Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
" . RBPEATBD MEASURBS DATA ANALYSIS WITH NONNORMAL OUTCOMBS by ALULA HADGU Department of Biostatistics University of North Carolina • Institute of Statistics Mimeo Series No. 2120T September 1993 REPEATED MEASURES DATA ANALYSIS WITH NONNORMAL OUTCOMES by Alula Badgu A Dissertation submitted to the faculty of the University of North Carolina at Chapel Bill in partial fulfillment of the requirements for the degree of Doctor of Public Bealth in the Department of Biostatistica. Chapel Bill, 1993 Approved by: ~. Advisor --~---Reader • ii ABSTRACT ALULA HADGU. Repeated Measures Data Analysis With Nonnormal Outcomes. (Under the direction of GARY G. KOCH). Longitudinal data present statistical problems of interest in public health and epidemiologic studies. In this dissertat.ion, we consider longit.udinal data in which out.come measurements are repeatedly taken on each subject., and there is primary interest in the dependence of the outcome variable on covariatee. This work is motivat.ed by a longitudinal study of women and their ectopic pregnancy outcomes in Lund, Sweden, and a randomized clinical trial of three mouthrinses in inhibiting the development of dental plaque. The common data structure of these studies is the presence of an intra-class or serial correlation within primary sampling units or subjects. Recently Liang and Zeger extended Generalized Linear Models by taking correlations within primary sampling units into account using quasi-likelihood. In this dissertation, we review and apply the Liang-Zeger methodology to the two data sets mentioned above. We further analyze the ectopic pregnancy data using Rosner's conditional modeling approach which is a generalization of the beta binomial models for correlated data. The specific questions we wish to address in the ectopic pregnancy study from Lund include: 1. Does the presence of pelvic inflammatory disease predispose a woman to a subsequent development of ectopic pregnancy? 2. Does the presence of mycoplasma at index laparoscopy increase a woman's risk of having ectopic pregnancy? 3. What are t.he other predictors of ectopic pregnancy? The objectives of the randomized clinical trial include: 1. 2. JJe the two experimental mouthrinses more effective than the standard mouthrinse? Is the effect. dependent on baseline plaque measurement? If so, in what way? iii Although the Liang-Zeger methodology gives regression estimates and variances that are robust against misspecification of the correlation structure, correct specification provides more efficient estimates. Rotnitzky and Jewell (Biometrika, 1991) provide a technique to verify if the "guessed" correlation matrix is reasonable. In this work, we modified the GEE software by writing additional IML codes 10 that one may use the technique of Rotnitzky and Jewell to verify if the "gueued" correlation matrix is reasonable. Furthermore, we propose to use the GEE approach u a way to obtain variance estimates of sensitivity and specificity estimates when the data consist of correlated binary outcomes. We give rust order approximations to the variances of sensitivity, specificity, predictive value positive, and predictive value negative • v TABLE OF CONTENTS Chapter • Page 1. INTRODUcrION ...•.......•••••••.•.••.•.•••.•.••••••..••.••••••••.•.••••••••..••••••••••••.•••.•.••...••••••...•••••.••• 1 2. LITERATURE REVIEW 3. 2.1 Linear Models for Repeated Measurement Data 2.2 Approach for NODDormal Data 5 REPEATED MEASUREMENT ANALYSIS USING GENERALIZED ESTlMATING EQUATIONS••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••.••••••••••••••13 3.1 Generalized Linear Models and Quasi-likelihood 3.2 Quasi Likelihood 3.3 Extension of GLM to Longitudinal Data 3.4 Logistic regression 4. THE EPIDEMIOLOGY OF PELVIC INFLAMMATORY DISEASE AND ECTOPIC PREGNANCY 5. 4.1 Epidemiology and Etiology of Ectopic Pregnancy 4.2 Epidemiology and Etiology of PID 3g ANALYSIS OF THE ECTOPIC PREGNANCY DATA USING GEE AND CONDITIONAL MODELS 5.1 Description of the Lund Study 5.2 Descriptive Analysis 5.3 Model Fitting 5.3.1 Analysis Using Standard Logistic Regression Model 5.3.2 Analysis Using GEE Techniques 5.3.2.1 Selection Between Correlation Structuna 4:1 vi 5.3.3 Analysis Using Rosner's Beta-Binomial Mod..:d 5.3.3.1 Extension to the General Case 5.3.3.2 Comparison with Ordinary Logiatic Regreaaion 5.3.3.3 Application of Roener's Model to the Ectopic Pregnancy Data 5.3.3.4 Comparison with Ordinary Logiatic Regression and Gee Model 5.3.3.5 Comparison of the Intercepts of the Models 5.3.4 Alternative Approach 6. ANALYSIS OF A MOUTH-WASH CLINICAL TRlAL 6.1 92 Descriptive Analysis 6.1.1 Mouthrinse Dispensation 6.1.2 Examination 6.1.3 Descriptive Analysis 6.2 Model Fitting 6.2.1 Analysis Using GEE Techniques 6.2.2 Selection Among the Correlation Structures 6.2.3 The Effects of Specifying Different Link Functions 7. SENSITMTY AND SPECIFICITY FOR CORRELATED OUTCOMES 7.1 Introduction 7.2 Sensitivity and Specificity in Clustered Data 7.3 Sensitivity and Specificity of Test Stripe 7.4 Sample Survey Approach 8. SUMMARY AND CONCLUSIONS 8.1 Summary 8.2 Future Directions BIBLIOGRAPHY 112 121 124 vii LIST OF TABLES Table 3.1 Comparison of the Discriminant and Maximum Likelihood !:;timates of the Logistic Regression Coefficienta Table 4.1 38 Selected c:ase-c:ontrolstudies to evaluate the relation between the use of an intrauterine device (IUD) and the occurrence of ectopic pregnancy 42 Table 4.2 Selected cue-c:ontrolstudies oral contraception (OC) use and the occurrence of ectopic pregnancy Table 4.3 3 Probability estimate of acute pelvic inflammatory disease for various combinations of symptoms based on a logistic regression analysis with no interaction Table 5.1 _ Number of Pregnancies Contributed by Each Woman During the 1960-1984 Time Period Table 5.2 46 52 Univariate Analysis of Ectopic Pregnancy that Occurred During 1960-1984 Time Period at the Department of Obstetrics and Gynecology, University Hospital, Lund, Sweden Table 5.3 53 Code Sheet for the Variables in the Lund Ectopic Pregnancy Data.............................................................................................................. 61 Table 5.4 Univariate Logistic Regression Models for Ectopic Pregnancy Table 5.5 Estimated Coefficients and Standard Errors for the _ 63 Multivariate Model Containing Variables Identified in the Univariate Analysis Tables 5.6 Estimated Coefficients Standard Errors for the Multivariate Model Containing Vanables Significant in Table 5.5. _ 66 viii Table 5.7 Log-likelihood, Likelihood Ratio Test Statistic (G), Degree of Freedom, and P-Value for Possible Interactions of Interest to be Added to the Main Effect Model Table 5.8 _ Estimated Coefficients and Standard Enors for the Multivariate Model Containing Main Effects and SignifiC&Dt Interaction Table 5.9 68 __.69 Estimated Odds ratios and 95% C.I. for variables not involving interaction terma............................................................................................ 69 Table 5.10 Expreesion for the Logits and Logit Differences in Tenm of the Estimated Parameters for the Possible Combinations of MYCO and SCORE _ Table 5.11 Values of the Estimated Logit Differences and Odds Ratioe Table 5.12 GEE Analysis of Ectopic Pregnancy with Logit Link and 71 Binomial Variance Table 5.13 Relative Difference Between Robuat and Naive 74 ~Statistics for the Independent and Exchangeable Correlation Structure Table 5.14 77 Results of the Ectopic Pregnancy Data U.ing Roener'. BetaBinomial Modeling Approach Table 5.16 Table 6.2 _ 86 Comparison of the Intercepts ofthe Standard Logistic Regression and Itosner'. Model Table 6.1 --84 Estimated Logistic Regression Coefficients from Alternative Model Predicting Ectopic Pregnancy in 1497 Women Table 5.17 76 The Eigenvalues of Matrix Q for the Independent and Exchangeable Correlation Structures Table 5.15 70 _.89 Summary Statistics of Demographic Characteristics by Treatment Group 94 Descriptive Statistics for Plaque Index Data. 95 ix Table 6.3 GEE Analysis of the Dental Clioieal Trial with Log Link and 103 G&II1JI1& Varianee Table 6.4 GEE Analysis of the Dental Clinieal Trial with Gamma Variance and Log Link Table 6.5 107 Estimated Regression Coefficients and Robust-Z Statistics for Gamma Error and Identity-Link Table 7.1 l06 Comparison of Robust and Naive Z Statistics· (Gamma Error Log Link) Table 6.6 _ 109 GEE Study of Sensitivity and Specificity of the Test Strip at 8 Minutes 118 x LIST OF FIGURES • • Figure 5.1 Ectopic Pregnancy Rates by Severity and Pregnancy Order Figure 6.1 Mean Plaque Index by Type of Mouthrinae Figure 6.2 Fitted Gamma Distribution for Dental Plaque - Month 3 ---98 Figure 6.3 Fitted Gamma Distribution for Dental Plaque - Month 6 _99 96 Figure 6.4 Gamma Quantile - Quantile Plot for Dental Plaque Index - Month 3 100 Figure 6.5 Gamma Quantile - Quantile Plot for Dental Plaque Index - Month 6 1Q1 Figure 6.6 Observed and Predicted Values for Log Link 1l0 Figure 6.7 Observed and Predicted Values for Identity Link • _.57 __ 1l1 CHAPTER 1 INTRODUCTION Longitudinal data present statistical problems of interest in public health and epidemiologic studies. In this dissertation, we consider longitudinal data in which outcome measurements are repeatedly taken on each subject and the primary question is the dependence of the outcome variable on covariates. The common data structure of repeated measurement designs is the presence of intradass correlation within subjects (dusters) and this intraclass correlation must be taken into account when one analyzes the data. This work is motivated by a longitudinal study of women and their pregnancy outcomes in Lund, Sweden. The objective of this work is to develop a statistical model and an analysis approach which : (1) allows missing data and covariates that change over time (2) allows for testing of hypotheses about regression parameters (3) accommodates to response observatioDS that are not independent. Epidemiologic studies of pregnancy outcome may focus on one of several adverse outcomes, such as stillbirth, low birth weight, ectopic pregnancy, malformations, and developmental abnormalities. In this dissertation we will study only one of these pregnancy adverse outcomes, namely ectopic pregnancy. Some of the covariates of interest are flXed characteristics of the woman and will remain constant for all of her pregnancies. variables will be referred to as time-independent variables. These The distinction between time- independent and time-dependent variables is important in models that employ repeated measurements. The logistic regression model is a commonly utilized statistical model in public health and epidemiologic studies. However, because of the correlation among pregnancies contributed by 2 the same woman, the assumption that the observations must be independent, is not satisfied. Researchers use different approaches in analyzing pregnancy data. Some researchers focus their analyses on only one pregnancy, usually the most recent pregnancy. The problem with this approach is that, though the assumption of independence is satisfied, there is a .. substantial loss of information and loss of statistical power. Other researchers have choeen to ignore the dependent nature of the data and proceed using logistic regression techniques. For repeated measures data, however, the independence assumption is usually violated. The dataset is made up of independent clusters, and within clusters the individual observations are usu&lly correlated. This correlation can not be ignored without damaging the inference on the regression coefficients. Recently an extension of generalized linear models that takes the correlations among measurements on the same individual into account has been developed by Lian,; and l-eger (1986). The technique of Liang and Zeger is semi-parametric since the estimating equations are derived without fully specifying the joint distribution of a subject's observations. Instead, • only the likelihood for the marginal distributions and a "working" covariance matrix for the observations are specified. Liang and Zeger show that the regression estimates obtained by this technique have optimal statistical properties such as consistency and asymptotic normality even if the correlation matrix has been misspecified. In this study we use the technique of Liang and Zeger, also known as the generalized estimating equations (GEE) method, to develop a statistical model for the Swedish ectopic pregnancy data. Our research interest is to study the marginal dependence of the response variable, namely the presence or absence of ectopic pregnancy on covariates. The specific questions we address include: (i) How does the prevalence of ectopic pregnancy change with age? (ii) How does the prevalence of ectopic pregnancy depend upon a woman's history of pelvic inflammatory disease? 3 (iii) How does the degree of PID severity and age affect the prevalence of ectopic pregnancy? (iv) How does the prevalence of ectopic pregnancy depend upon history of adnexal surgeries? The results of a atatistic:al analysis using the standard logistic regression model, which assumes independence of all the observations, are compared to the results obtained using the GEE approach. AB mentioned before, one of the advantages of the GEE methodology is that parameter estimates are robust against misapecification of the correlation structure. specification results in the mOlt efficient estimates. However, correct Rotnitzky and Jewell (1991) provide a technique to a.eea if the "guessed" correlation matrix is unreasonable. Thus, in this work we modified the GEE software by writing computer programs by which one may check if the "working" correlation matrix is reasonable. We further modify the GEE programs to provide adjusted Wald test statistics that take the correlation structure of the data into consideration as suggested by Rao and Scott (1984, 1987). These were originally introduced 88 extensions or adjustments to the usual Pearson chi-squared tests for data arising from complex survey sampling schemes. In this dissertation the GEE backbone program was obtained from Dr. Rezaul Karim of Johns Hopkins University. We then modified this program to obtain both robust and naive Wald statistics and their associated }>-values for any composite hypothesis regarding the regression coefficients. The scientific questions which can be addressed by repeated measurement studies are generally the marginal dependence of response on covariates or individual changes over time as in the case of transitional models. Marginal models describe the occasion-specific distribution of the response variable, whereas transitional models describe the distribution of individual changes over time. The distinction between marginal and transitional models will be discussed in the next chapters. General approaches for the analysis of repeated measures data are available for both continuous and categoric:al response variables. However mOlt of the methods are for continuous response variables based on parametric models assuming a multivariate normal 4 error structure (Ware, J.H. 1981; Louis, T 1988; Timm, N.H. 1980) The first general approach to the analysis of repeated measures when the response is categorical was described by Koch et al. (1977) and is based upon the weighted least squares (WLS) methodology of Grizzle, Starmer and Koch (1969). Many authors have extended the weighted least squares methodology to a variety of response functions of repeated measures categorical data with missing data (Stanish, Gillings, and Koch, 1978; Woolson and Clarke, 1984; Landis, Miller, Davia and Koch 1988). However, these methoda baaed on normal theory and categorical methods baaed on the weighted least squares approach are not always appropriate for the analysis of repeated measures data. First, the parametric &Uumption of multivariate normality may not be satisfied. Secondly, in situations when the response variable is dichotomous or ordinal, the WLS technique allows only categorical predictors, and thus cannot be used when there are continuous independent variables. In addition, the total sample size within each sub- population of the multi-way croea-c:lassification of response and covariates must be moderately large. Thus, this may impose some limitations in practice. In Chapter 2, some of the procedures for repeated measurements in the literature will be reviewed. In Chapter 3, the extension of generalized linear models to repeated measures data (Zeger, 1986; Liang 1986) will be discussed. The Swedish pregnancy data, and the epidemiology, etiology, and diagnosis of both pelvic inflammatory disease and ectopic pregnancy data is discussed in Chapter 4. In Chapter 5, the pregnancy data are analyzed using ordinary logistic regression, GEE methodology and Rosner's conditional modeling approach. In Chapter 6, we use the GEE techniques to study the effectiveness of three mouthrinses in inhibiting the development of dental plaque. In Chapter 7, we use the GEE approach to obtain estimates of sensitivity and specificity when the data consist of correlated binary outcomes. First order Taylor series approximations to the variances of these estimates are provided. Data from a dental clinical study are used to motivate and illustrate the method. In Chapter 8 we give summary, conclusions and suggestions for future research. CHAPTER 2 LITERATURE REVIEW A number of statistical papers, involving both continuous and categorical response variables, have been written on repeated measures analysis. A majority of the methods were developed for continuous response variables based on parametric models assuming a univariate or multivariate normal error structure. These parametric techniques are reviewed by Louis (1988), Ware (1965), and Timm (1980). One advantage of these classical techniques is the availability of existing software, such as PROC GLM and PROC LINMOD in the SAS system. In this chapter we will review (1) the common methods based on parametric assumptions and (2) the techniques for categorical response variables. 2.1 Linear Models for Repeated Measures There are two approaches to repeated measures analysis. The first approach is the population average model (Zeger, Liang, and Albert, 1988) where dependence of a response variable on the independent variables is constant and observed individual difference is due to sampling variability. In this approach, the regression coefficients have interpretations for the population and not for any individual (Zeger, Liang, Albert). This approach is also referred to in the statistical literature as marginal modeling. The second approach to modeling repeated measures is referred to as subject specific modeling, where dependence of a response on independent variables is subject-specific. This situation arises when obeerved individual differences are heterogeneous and this observed subject heterogeneity can be explicitly modeled. Example of subject-specific modela include the random effects model '(Liang and Zeger, 1990) and the mixed-eff'ects logistic model (Stiratelli, Laird and Ware, 1980; and Anderson and Aitkin, 1985) where one assumes natural heterogeneity among subjects in 6 regression coefficients. The ra.ndom-effects model is given by: where Yi is a vector o( responaes Xi is a known matrix o( covariate. P is a vector o( regression parameters bi - tid N(O, D) t 2 i - tid N(O, 17 ) independent o( b i example: Y·=PO 1 + PIX.1 + b01. + b 11·X.1 + t·1 where Yi = height o( child i Xi =age o( child i + P1Xi ) (Po is considered to be the population average growth function a.nd (boi + bliX i ) is considered as deviation specific (or child i. A. stated by Zeger, Liang, and Albert (1988), if the primary research interest is marginal dependence o( response on covariates then the population average model (marginal model) should be uaed. H the primary interest is to study the individual changes over time then subject specific models should be uaed. The approach to repeated measures data for continuous response is based on the standard general linear multivariate model (GLMM), which is a generalization o( the general linear univariate model (GLUM) which includes two or more dependent variables, all o( which share 7 the same, identical design matrix. The general linear multivariate model is given by (2.1) E[y]=XP denotes a K x t matrix of observations or measurements on t dependent variables or responae variables, Yl' Y2,·..'Yt· The jth column of Y is Yj and is a K x 1 vector which satisfies GLUM (Yj; XPj , C1'ij I K ) X denotes an K x q (K > where q) design matrix whose values are fixed constants which are assumed to be known without appreciable error. is a q x t matrix of unknown primary parameters. One of the objectives of repeated measures analysis is to estimate these primary parameters. An equivalent representation of model (2.1) is: E(vec[Y])= E = (It = ~ X) vec(P) ( 2.2) The constant variance-covariance matrix for each set of dependent variables is given by E, a txt matrix. ( 2.3) The correlation among the dependent variables is given by R =[R.~ = Diag (Ef 1/ 2 E Diag (Ef 1/ 2 ( 2.4) 8 When X is of less than full rank then fJ is non-estimable and not well-defined. P= when X is of full rank However, (X'Xr 1X'Y is an unbiased estimator of fJ. Moreover for each dependent variable separately, it provides a uniformly minimum variance estimator of fJ. Regardless of the rank of X, t t the estimator of E is given by = (Y - Xp)' (Y - XP) / (K - rank (X» =(Y'Y - P'X'Y)/(K - rank (X» t ( 2.5) is an unbiased estimator of E Bypothesis Testing: When testing hypotheses of interest we assume that the rows of Y are independently distributed as t-variate normal random vectors. The hypothesis of interest has the following form: . Bo : 9=CfJU=O HI: 9 = C fJ U ¢ where C is a X (2.6) 0 q contrast matrix consisting of any set of linearly independent row vectors and U is t x m and is of full rank m. C and U are chosen appropriately depending on the hypothesis of interest. If C(X'X)- C' is nonsingular then the above hypothesis is said to be testable. There are many test statistics for the hypothesis above and all of' these test statistics are functions of the following matrices: B = E 8' [C(X'X)- C'r 1 9 =(K - rank (X» U· t (2.7) U (2.8) Some of these multivariate test statistics include (1) Wilks' Lambda - lEI / IB + EI (2.9) 9 = Tr (H(H+Er 1] (2) Pillai's trace (3) Hotelling -Lawley Trace (4) Roy's Largest Root (2.10) = Tr [HE- 1] (2.11) = ..\1 = largest eigenvalue of HE"'1 (2.12) The general linear multivariate model has been extended by attaching a post stratification matrix T to equation (2.1). This approach assumes that polynomial regression of the observations on time adequately deae:ribes the dependence of the observations on time (Potthoff, 1964). This is the 10 called polynomial regression method usually used in growth curve problems to compare how the response changes over time among groups. The model is given by E[Y] = (2.13) XPT where Y, X and fJ are as described above and T is a matrix of polynomials in time. By appending a post matrix T as in equation (2.13), one imposes a time structure on the mean responses. An example of this type of modeling approach which incorporates covariates is given by Grizzle and Allen (1969). Kleinbaum (1969, 1973) has extended this model to allow for missing values. If from the estimate of the variance-covariance matrix, t, we learn that the 'covariance terms are approximately equal (compound symmetry) then it may be prudent to perform a univariate analysis, such as for a randomized block design with subjects as blocks so that we gain power and precision. Thus a vector of observations is represented as one index, such as an average of response. Then it equality of the average response over time between groupe is the question of interest, a univariate ANOVA can be used. This univariate approach then tests for main effects of groupe (between subjects effect), and main effects of time (within subject effect). The F statistic for the fixed effects of group is distributed as F as long as measurements are independent of group membership. The necessary and sufficient condition 10 for the F statistics to be exactly as F under the null hypothesis is that the correlation matrix of observation vector has the following form: 1 P P pip p p 1 p p p p p p p p 1 This is called "uniform" correlation by Cole and Grizzle (1966), and also referred to as intraclass correlation or exchangeable correlation. If this is not the case, F statistics for time efl'ecta are not distributed as F (Kirk, 1982). Adjustment factors to correct this problem have been suggested by Greenhouse 1959), Huynh and Feldt (1970), and Schwertman (1978) . 2.2 Approach I2t Nonpormal 12Aa The first general approach to the analysis of repeated measures when the response variable is categorical was given by Koch et aI. (1977) and is based on the weighted least squares methodology of Grizzle, Starmer and Koch (1969). Other authors have extended this technique to a variety of response functioDB for complete and incomplete repeated measures categorical data (Stanish, Woollon, Landis). In this technique the hypotheses of interest to be addressed in a repeated measures design include: (1) Differences among. aubpopulatiolll regarding the average distribution of the responses to say d conditions (e.g., treatments ). (2) Difference among the d conditiolll or treatments regarding the average distribution 11 across the s--subpopulations. Test statistics for the above hypotheses as well as parameter estimation are then undertaken within the general framework of weighted least squares by specifying the appropriate function F which represents estimators for the parameters under analysis and the corresponding operator matrices (e.g. operator matrices A, design matrix X, and contrast matrix C). A detailed explanation of the construction of these operator matrices is given in Landis et al (1988) . The weighted least squares methodology requires that the sample size within each subpopulation be moderately large and can not be used when there are continuous independent variables. Furthermore, it doea not accommodate time-dependent covariates, thus, in some practical situations this may impose a severe limitation. Wei and Stram (1988) describe another approach to analyzing repeated measures data with time-dependent and/or time-independent covariates. In this approach subjects are observed repeatedly at a common set of observation times. At each time point, the data are first analyzed using the technique of McCullagh and Neider (1982). The parameters of these models are assumed to be occasion-specific and are estimated by maximizing the occasionspecific quasi-likelihood equations. Let vec({3) = (P1' P2, u, P t )' denote the pt x 1 parameter vector, where P1, P2, ... , P t are the parameter estimates at time point 1, time point 2,... time point t. The authors show that the regression coefficients over different time points are asymptotically jointly 1 2 normal for large samples; i.e. K / (vec(P) - vec({3» is asymptotically normal with mean 0 -1 and covariance matrix V {3=I:~(X'X) of dimension pt x pt. Wei and Stram then provide a consistent estimator of V{3. The regression estimates from the time-specific univariate regressions can be combined to give a global picture of the effects of individual covariate:s on the response variable over the entire time period. For example, consider the t estimated regression coefficients for the kth covariate, say, age. 12 The sequential multiple test procedures of Marcus, Peritz and Gabriel (1976) can be used to test if there are significant changea in the t coefficients of age over time. If there are no significant changes then one may proceed to provide a global estimate for the covariate of interest. Wei and Stram also provide a lack of fit test for testing the adequacy of the quuilikelihood model used at each time point. The Wei and Stram method (1988) is cloeely related to the GEE method. If the GEE regression coefficients are allowed to vary in time and if the independence working correlation matrix is lUed, the two methods give identie&1 results (Zeger, 1988). However, for repeated measures, the GEE method is usually specified to give an estimate of a single regression parameter vector which is assumed to be constant over time. The Wei and Stram method gives separate regression estimates at each time point which then may be combined to give a global estimate if desired. Stram, Wei, and Ware (1988) describe an approach for analyzing repeated measures of an ordinal categorie&1 response variable with possibly missing observations and time-dependent covariates. In this approach, as in the case of Wei and Stram approach mentioned above, subjects are observed repeatedly at a common set of observation times. At each time point the data are first analyzed using the proportional odds regression model from a class of models for ordinal data proposed by McCullagh (1980). The first class of non-parametric methods for repeated measures was described by Koch, (1968). Other more recent nonparametric procedures include the Wei and Lachin method (1984) and the Wei and Johnson method (1985). Apart from marginal models for binary repeated measurements, no method was available for nODDormal outcomes with possibly missing observations and time dependent covariates until Liang and Zeger (1986). Liang and Zeger extended Neider and Wedderburn'. (1972) generalized linear modeb and Wedderburn'. quasi-likelihood method for repeated measures data so that it covers a wide class of outcome distributions. This will be discussed in the next chapter. . .. CHAPTER 3 REPEATED MEASUREMENT ANALYSIS USING GENERALIZED ESTIMATING EQUATIONS The first part of this chapter is a description of generalized linear models and quasilikelihood. The second part will discuss the extension of generalized linear models and quasilikelihood to repeated meuures data. Thus mOlt of this chapter is baaed on the concepts and expositions given by NeIder and Wedderburn (1972), Wedderburn (1974), McCullagh and Neider (1983), Zeger and Liang (1986); Liang and Zeger (1986), and Liang and Zeger (1991). • Generalized Linear Models 3.1 Generalized Linear Models (G LM) are simply an extension of classical linear models, and were first introduced by NeIder and Wedderburn (1972). In classical linear models set up recall that (3.1) where Y denotes a K x 1 vector of observations assumed to have mean IJ and constant variance and often a normal distribution. Similarly f is a K x 1 vector of error terms assumed to have mean 0 and a normal distribution. The fact that f has zero mean and a normal distribution excludes consideration of response variables that have Poisson, binomial, positive exponential, etc. distributions. In order to generalize the model to cover a wide range of dependent variables, it is important to separate clearly the deterministic linear regression component (systematic part) from the random component. reformulating equation (3.1) into three components. This can be done by 14 1. The systematic linear component part, the linear predictor '1 = X/3 where X is a Kxp matrix of known covariates and /3 is a px1 vector of unknown regression parameters. 2. The random component part Y comes from an exponential family with E(Y) = JJ • 3. The link component part that relates or links the mean of the response variable to the systematic linear component X /3. '1 = g (JJ) where g (JJ) is the link function which may be any monotonic differentiable function. The link function, g, changes from one generalized linear model to another as will be shown later. Thus generalized linear models allow the random component to come from an exponential family such as Poisson, Gamma, Binomial, etc.. It is known that if the likelihood has the exponential family form, maximum likelihood estimates of regression parameters can be found using the method of iteratively reweighted least squares (Bradley, 1973). The equation maximizing a likelihood function and the equation minimizing a weighted sum of squares are equivalent if the likelihood function arises from the exponential family of distributions. The link function may be any monotonic differentiable function. Likelihood Function ~ Generalized LiDYI Models The probability function for a scalar oheervation y from an exponential family is fy (y; 9, ;) =exp {[y9 - b(9)] / a(;) + C(y, ;)} (3.2) for some suitably choeen functions a( ), b( ), and c( ),. ; is a dispersion parameter assumed .. 15 known. If tP is unknown then fy(Yi 9, tP) may not be an exponential family (NeIder &l. Wedderburn 1972). The log likelihood for the datum y in equation (3.2) is given by y9-b(9) L = a(tP) + (3.3) C(y,4» Differentiating (3.3) with respect to 9, we get 8L y-b'(9) (3.4) 89 = a(tP) Further differentiating (3.4) with respect to 9, we get (3.5) From the well-known relationship (Kendall and Stuart) E (~) = 0 => E(y) = ~ = b'(9) (3.6) => E (y_~]2 = Yar(y) = b"(S) a(4)) (3.7) where b'(.) and b"(.) denote the first and second derivatives of b( ) with respect to 8, respectively. Thus note Yar(y) hu two components: (1) b" (9) - variance function dependent upon the canonical parameter 9 (and hence on the mean) a( 4» which is independent of the canonical parameter and hence independent of the (2) mean. The function a( 4» is usually of the form 16 a(<p) = <p/w where w is a known weight function. Example 1: For the Poisson distribution fy (YiJJ) = JJY e- u / y! = exp {[Yln(JJ) - JJ] + [-In(y!)]} = exp {(y8 ~e) + [-In(y!)]}, = exp ([[eye) - b(8)]/a(<p)] + C (y, <PH where D=InJJ b(D) =eO C(y, <p) = [-In (y!)] a(<p) = 1 Example 2. Normal distribution fy (Yi 0, <p) =~ 1211'6- exp[~(Y-JJ)2] 26 = exp {[yO - b(D)] / a(<p) + C(y, <PH = JJ b(O) = JJ2/2 where 0 a(<p) = 6 2 = <p C(y, <p) = - 1/2 (y2/6 2 + In (211'6 2)] From (3.7) Var(y) = E [y - JJ]2 = b"(8) a(<p) =lx6 2 =6 2 Example 3. Binomial Distribution ny - binomial (n, JJ) y =0, l/n, ..., 1 17 £(y, /J) = (oCn y ) /Joy (l_/J)o-oy = exp {ny log (/J/(1-/J)) = b(9) 9 log (1- /J) y log (/J/(1-/J» - log (1 - /Jr 1 l/n exp { = exp {( y9 - b(9» where +0 / a (C;) + log (oC oy )} + log nCny} + C(y, C;)} = log 1~ -/J =log (1 - /Jr 1 = log (1 + e9) ~ b' (0) = e 9 ~ = /J a(c;) = l/n C(y, ¢) = log nCny Var (y) = b" (9) a(¢) = e n(l Note that when a(¢) /J (1-/J) / n. = ¢/o When ¢ > 9 + eO = /J (l-/J) )2 then y is said to have an extra-binomial variation, Var (y) 1 then we have the case of overdispersion; if ¢ < 1 it is =¢ ~eferred to as underdispersion. For examples of overdiapened data see Crowder (1979), Williams (1982), and Cox and Snell (1989). Smith and Heitjan (1993) present a score test for assessing whether overdispersion is present in generalized linear models. the score function for P is given by 18 -h S(P) = K -h = i=1 L: = = t i=1 t t {I 8P + lJ9. t y.I - b'(9.) I } a(tIl) (1j) a9 i (ap) i=1 = ~) y.9 - b(9.) I a(tIl) 8 i=1 = log f(Yi; 9i' t C(Yi' til)} (chain rule lJL ~ lJL) = (89)t ~ 89 Yi - E[Yi] } { a(tIl) t i=1 ( 3.8) Thus the score function depends on the density function, f(Yi)' only through J.li Canonical = E(Yi) and LiW As defined by McCullagh and NeIder (1972) the link function is the function that relates the linear predictor " to the expected value J.l of the datum y. As shown before there are special links for the various distributions of interest. These links occur when Normal Poiaaon (I = f] and are eaUed canonical links. " = J.l 19 " = In (1-(p)p) Binomial Gamma Inverse Normal The generalized linear models are fitted by the method of maximum likelihood, using an iterative weighted least squares algorithm. the parameter Thi.e yields maximum likekihood estimates ~ of I!, and their corresponding standard errors. Hence estimates ij for the linear predictor and Pi for the fitted values can be obtained. Details are given in Neider and Wedderburn (1972). Just as important as the parameter estimates, are assessments of goodness of fit for the model fitted to the data. Given K observations one can fit models that contain up to K parameters. One extreme model is the null model which gives one common value for all the p'a. At the other extreme is the maximal model with K parameters, one parameter per observation. The maximal model assigns all variation between the Yi's to the systematic component of the model. It is said to be a completely specified model with the error term equal to zero. Even though the maximal model does not do much in terms of summarizing the data, it does provide a baseline against which models with fewer parameter:s can be compared and contrasted. Thi.e comparison is simply the difference between the maximum log likelihood achievable under the fujI model and that achieveable under the reduced model. Let ley, rP, y) = maximum likelihood achievable under the full model, l(p, and let tP, y) = maximum likelihood achievable under the reduced model, 8 and 9 be the estimated canonical parameters under the reduced and the full model respectively. Then the scaled deviance is: 5(y, M = - 2 log { I(p, tP, y)} ley, ,pi y) 20 = 2( l(y, tP, y) - 1 (p, tP; y)] K = 2 ~ (Yi(i - i) - b(i) + b(i)] / a(tP) i=1 =2 t wi(Yi(i - i) - b(i) + b(i)] / t/J i=1 (3.9) where D(Yl p» is referred to as the deviance of the reduced model relative to the full model; tP is the scale parameter. For the full model l(y, tPi y) = =f In (211'(7'2) For the reduced model l(p , tP; y) =-(2(7'2 r l t (since y = p). (Yi - Pi)2 - i=1 Therefore, the deviance D(Yi' p) = t In (211'(7'2). .t (Yi - p)2 which corresponda to the 1=1 residual sum of squares under the reduced model. Similarly, if the error probability distribution is the Poisson distribution then which corresponda to the log-likelihood ratio statistic often uaed as a test statistic in the analysis of contingency tables. 21 For the g&IIlma distribution K D(y, jl) =2L [-In (Yi / f£i) + (Yi - f£i) / f£i] i=1 For Normal distributions the scale deviance haa a X2 11 II distribution with = number of parameters in the full model minus the number of linearly independent parameters in the reduced model. For non-normal distribution&, the scaled deviance is distributed aa X211 onlyasymptotieally, thus little is known about the distribution of non-normal deviance for small sample sizes (Adena and Wilson 1982). Now suppoee we wish to compare two models, ml and m2, where m2 has 112 - III more terms than ml. Suppose S (ml, y) S (m2, y) = the scaled deviance of modell, = the scaled deviance of model 2; then S(ml, m2) = S(ml, y) - S(m2, y) has a x2112 _ III distribution if the two models have Normal errors. For other distributions S(ml, m2) haa an approximate x2112- 1I 1 distribution. Another measure of goodness of fit is the generalized Pearson X2 statistic, which is given by where V(jl) is the estimated variance function for the distribution of interest. 3.2 Quasi-likelihood In this eeetion we review quasi-likelihood met.hods and point. out lOme of the important 22 concepts and properties of quasi-likelihood techniques. This section is based upon the concept and expositions given by Wedderburn (1974), McCullagh (1988), and Liang and Zeger (1991). Unlike the likelihood approach of Section 3.1, in quasi-likelihood, analogues to equation 3.2 are missing, that is we are unsure about the random mechanism from which the data come from. Instead of assuming that y comes from an exponential family, we only specify the relationship between the mean and variance. This means that only a second moment assumption is required in quasi-likelihood analysis - thus avoiding the complete specification or the underlying distribution. This is a particularly useful approach if one does not have enough data to estimate a distribution form but sufficient enough to estimate the relationship of the first two moments (Pregibon. 1984). Hence the assumptions for quasi-likelihood analysis are: (1) g (1'.) 1 = x.p 1 (2) Var (Yi) = tP V(J'i) where V is a known function Definition: Suppose we have independent observations Y i (i = 1.....K) with E(Y.) =1" and 1 1 where V is some known function. Furthermore. suppose for each observation J'i some known function of a set of parameters Pl' P2..... 13 p • .= E(Yi) is Then for each observation Wedderburn defines the quasi-likelihood function. or more correctly. the log quasi-likelihood. for the ith observation. p. Q(Yii J'i) = j (3.10) Yi or equivalently the quasi-likelihood function Q. can be defined as a system of partial differential equations: 23 (3.11) The quaai-likelihood is the sum of components of (3.10) over the K observations. The statistical properties of quasi-likelihood functions are similar to those of ordinary likelihood. Details are given by Wedderbum (1974) and McCullagh (1983). Some of the major results are summarized below. (1) The quasi-score function is simply the first derivative of the log quasi-likelihood with respect to fJ and is given by (3.12) V(Pi) is referred aa the "working" variance of Yi. (2) .j(S(fJ) variance is asymptotically distributed aa multivariate normal with mean = 0 and =D where ~ () po t i l 8 Pi D = L..J (#) y- (Pi) Var(Yi) y- (Pi) 8! /K i=1 (3) Among linear estimators of fJ , quaai-likelihood estimates have minimum variance (Firth 1987). (4) ..fK (iJ - fJ) is also asymptotically distributed as multivariate normal with mean 0 and variance-covariance matrix I;1 D E- 1 where All the above results are baaed on the first-order Taylor Series expansion of the equations 24 determining p, S(P) = 0 in (P - p). Next is a table of quasi-likelihood, quasi-score functions, and variance functions for some of the familiar distributiona (Liang and Zeger, 1990). Distribution YUU Q(u;y) Normal _(Y-JJ)2/2 Poisson 1 Y In JJ-JJ Gamma JJ JJ2 -Y/JJ - InJJ log (l-JJ) Ylog~ -JJ + binomial JJ(l-JJ) As mentioned before, the only assumption we have in quasi-likelihood analysis is that we have to know the relationahip between the first two moments. Liang and Zeger extended this concept to longitudinal data where they view it as a multivariate version of the quasi- • likelihood function. (i.e. yt is now an n x 1 vector) . i 3.3 Extension of GLM to Repeated Measures In this section we summarize the major results of Liang and Zeger (1986) and. Zeger and Liang (1986). Recall that for quasi-likelihood to be applied to repeated measures data we require only that the variance of r be a known function of J! = E(r). Furthermore, the correlation structure must be taken into consideration when estimating the regression parameter P and its variance-covariance matrix. Let Yi t be the ni repeated measurements on individual i, i.e. ...·..·'Y·mi ) 1·-12 Yit = (Y·1'Y·2' 1 1 , , Let ~t = (~t'1 Xit ,....., 2 ....., K Xit ) be the vector of covariate8 corresponding to Yit' the tth p • 25 outcome in the ith individual where 1 = 1, 2,......, K and t = 1, 2,....., ni' Further let E(Yit) = lJit and R i (0) be ni x ni 'working' correlation matrix for each Yi' where 0 is an unknown vector of order s, by which R(o) can be fully specified. It is possible that the working correlation matrices can be different for different subjects. The varianee-covariance matrix of Yi is given by : V. 1 =A. 1/ 2 R..(o)A. 1/2 I fJ 1 1 (3.13) 1 where Ai = diag {var(Yit.)} is an ni x ni diagonal matrix which is a function of the mean, say g«lJit) as the tth diagonal element. Again R i (0) is an ni x ni 'working' correlation matrix that describes a postulated correlation structure of Yi' after having adjusted for the effects of the covariates. The quasi- • likelihood analogue of the score function is given by k S(P) = L: i=1 ot.1 Vi.1 (y.1 - IJ.) = 0 1 (3.14) where Note that this equation would be identical to equation 3.8 (regular score function) if y came from an exponential family and the obeervationa for each subject were independent. Ui (P, 0) = 0i Vi - 1 (Yi -lJi) is equivalent to the estimating function suggested by Wedderburn (1974) except that the Vi's here are functions of 0 and p. The only difference is the introduction of an additional parameter a to account for the aerial correlation of the data. 0 is consistently estimated by & which is a function of Furthermore ,p is consistently estimated by estimate of p, say PG consistent estimators of p and the scale parameter ,po ~ which itself is a function of p. Therefore an is obtained by iterating between aolutioDB to equation (3.14) and 0 and ,p as described by Liang and Zeger (1986). That is, PG is the 26 solution of t Ui i=1 {P, Q (P, ~(P») = 0 (3.15) where Ui(P, a) = Di t Vi -1 (Yi - Pi) and K is the number ofsubjects or clusters. Equation 3.15 is called a generalized estimating equation (GEE). Liang and Zeger show that as K goes to infinity, PG is a consistent estimator of P and that .JK (P G - P) is asymptotically multivariate normal with varianc:e-covariance matrix VG given by: (3.16) K t -1 where II = "" LJ D I. V.I D·I and i=1 A consistent estimator of VG' say estimators, PG' Q Io=t i=1 VG' D. t V.- 1 Cov(y.) V.- 1 D.. I I I I I is obtained by evaluating Di and Vi at the and ~, and replacing Cov(Yi) by (Yi - i'i) (Yi - i'i)'. Thus estimation of the parameters {3 G' a, and 4J is performed as follows: 1. Assume Ri (a) is the identity matrix and 4J = 1. Then estimate P using Fisher's scoring method. . 2. For any gIven /3- calcae ul t resl'd UA< .1. { ' } / (_ii)I/2 'Yit= Yit - Pit vi (3.17) where i'it depends on the current value of Pand iliii is the jJth element of the inverse of 11' 3. Estimate the nuisance parameter tP by ~- 1 - Lk Lni i=l t=l t 2 it / (N -p) (3.18) • 27 k L: N= where ni and p is the number of covariates. i=1 4. Estimate the 'working' correlation matrix K Ruv = L: 1iu1iv / (N - p) (3.19) i=1 The two Iteps, i.e. (3.15) and (3.17) - (3.19), are iterated until convergence. As stated by Liang and Zeger, incorrect specification of the correlation structure does not violate the consistency of the estimates PG and VG. However as shown by Liang (1986) specifying a working correlation matrix dose to the actual one increases the efficiency of the parameter estimates. (1) Some of the most common choices of the working correlation are: Identity matrix: This specification limply ignores the serial correlation structure of the data and treats the repeated observationa as if they are independent. In this situation one may use PROC LOGISTIC or PROC CATMOD to obtain estimates of regression coefficients. The resulting estimates will be consistent but not efficient if the correlation is large. 1 0 0 0 0 o 1 0 0 0 o 0 0 0 0 1 28 (3) Exchangeable correlation: This correlation specification assumes that the repeated observations share a common random component i.e. Corr(yit' Yiv) = Q for all t~v. The random effects model assumes this correlation structure: 1 Q Q Q Q Q Q Q Q 1 Q Q Q 1 Q Q 1 Under the assumption of exchangeable correlation, t i: a= ¢ (3) tit tit' I{ t>t' i=1 t Q is estimated by: .5 ni (ni - 1) - p} (3.20) i=1 Autoregressive Correlation: under autoregressive correlation structure the correlations decrease as the time spaces between the observations increases, i.e. Corr (Yit' Yia) =, Q It-sl: 1 Q Q 1 Q Q 1 Q a 1 t-2 29 Under autoregressive correlation structure Q can be estimated by the slope from the regression of log E ("rit "ris) on It-sl. (4) Besides the 3 correlation structures given above, the Unspecified correlation: correlation structure CAD al80 be totally unspecified if ni = n for all i. Then the 1/2 n(n-1) elements of R are estimated by It =~ / K ~ £- i=1 A.- 1/ 2 S. S.T A: 1/ 2 . I I I 1 (3.21) AB pointed out before, the GEE approach provides a powerful way of analyzing repeated measures data that arise from a wide range of distributions. I t also allows for accommodation to different choices of correlation structure as well as varying number of measurement times for subjects. Resulting estimates are consistent, asymptotically unbiased, and asymptotically normally distributed. However, all these nice properties are based upon large sample size assumptions. Paik (1988) investigates the small sample properties of the GEE technique for gamma error and AR-l correlation structure based on average cover~ge probability, i.e., the fraction of simulations whose confidence interval based on the t-statistic included the true parameter. Paik's simulation results indicate that the confidence intervala for the GEE estimator do not provide desirable coverage probability of the true' parameter due to considerably biased point estimates. Paik's Table I is shown on the next page. 30 Table I Average coverage probabilities for GEE interval estimation baaed on t-statistics and average interval length (parentheses) for on~parameter models baaed on 500 simulation samples nominal level = .95 correlation structure: AR-l a: .9 (upper entry) .3 (lower entry) number of parameters: 1 Sample SiR Working Correlation Independent Exchangeable AR-1 .924 • (2.37) .950 (1.49) .973 • (1.19) .950 (1.06) .980 • (.825) .917. (1.65) .940 .908 • (1.69) .934 (1.28) . 943 (1.11) .961 (.881) .897 • (1.55) .943 (1.21) .941 (1.02) .973. (.800) .930. (1.17) .937 (.999) .953 (.784) .920 • (1.23) .947 (1.02) .913. (.800) .890 • (1.59) .910 • (1.56) ~ .938 (1.72) (1.26) mentioned by Paik, the average interval length for sample size 20 and a = .9 under AR-l is close to that for sample size 50 under independent correlation, indicating that the same efficiency can be achieved with 40% of samples by specifying the correct correlation structure. If the sample size is small, say less than 30, some authors suggest that further steps be • 31 taken to obtain estimates with reduced bias. Thus Paik proposes three kinds of jackknife techniques, namely, the standard jackknife, the weighted jackknife, and the linear jackknife, which seem to reduce the bias and provide correct coverage probabilities. Asymptotic relative efficiency when the working correlation is misspedfied was examined by Liang (1986) when repeated measurements are binary and the true correlations are .3 and .7. When t.rue correlation was .3 and the 'working' correlation was misspecified as independent, AR-l, or exchangeable, the estimates were still efficient with lowest asymptotic relative efficiency of 0.95. However, when t.he t.rue correlation was 0.7 the asymptotic relat.ive efficiencies ranged from 0.23 (true correlation=exchangeable and working correlation=!- dependence) to 0.99 (true correlation=exchangeable and working correlation=independence). Chi.Squared Tests fQr ~ Regression Parameters Most of this section is based upon the concepts given by Rotnitzky and Jewell (1990), Rao and Scott (1981), and Rao and Scott (1987). From previous sections we have that: Var (f3G) = 11- 1 10 11- 1 (3.22) where 11 and 10 are defmed in Section 3.3. Equation 3.22 above is sometimes called as the sandwich estimator. Let us now consider partitioning the regression coefficients such that: a_G-_[~] I-' CI where {3G is a pxl vector and 'Y is a qx1 vector of the first q components of {3 G [1 ~ q ~ p ] and 6 is the remaining (JHl) components of {3G' We are interested in testing the hypothesis that: Ho : 'Y = 'Yo versua Ha : 'Y ::f; 'Yo· Then from the asymptotic normality of PG , the generalized Wald test statistic is given by 32 Tw (3.23) P where t is a qxl vector of the first q component of G and V"T is the qxq sub-matrix of V{J which is given in Equation 3.22. If we usume that the "working" correlation is the true correlation structure then we may use a modified "working" Wald statistic, (3.24 ) where W"T is the qxq submatrix of W {JI where W Q = K (~D. V.- 1 D·r 1 ~ L...tll I Rotnitzky and Jewell (1990) then proceed to provide an important result in the form of a They prove that the modified "working" Wald statistic is simply a weighted Theorem. combination of q independent x21 random variables I Le., T * w -- where 2: c·J X·J2 cl~c2~ ( 3.25 ) I ... ~Cq~O are the eigenvalues of a matrix Q which itself is a function of Dj' Vi' and Ai defined previously. In particular (3.26) where 1 -, -1· Qo=1('" ~D.V. D. L...t I I I Q 1 = 1('"1"" .LJ (3.27) n.v.- 1 Cov (y.) V.- 1 f>. I I I I I (3.28) 33 ifq<p (3.29) ifq=p where D (1) and D (2) are the ni xq and nix(M) matrices of the first q and last M columns i i of Di respectively. The implication of the above result is that the coefficients of the linear combination (Equation 3.25) are identically one when the "working" conelation is gueeeed correctly. That is to say if the "working" correlation is guessed conectly then c1 ~1 and ~ 1 where c 1 = q-l L Cj and C2 C2 = q-l LCj-2. However, note that cl ~1 and c2~1 does not necessarily imply that the "working" correlation matrix is indeed the true but unknown conelation matrix. Since clustering can have substantial effects on the distribution of T w • similar to that of the standard Pearson chi-squared test statistic, some adjustment of T w· is necessary. Following Rao and Scott (1981) Rotnitzky and Jewell propose adjusting T w • as follows. (3.30 ) In this dissertation the GEE backbone program was obtained from Dr Karim, Johns Hopkins University. We then modified this program to obtain the generalized Wald test statistic, T w , and the modified "working" Wald statistic, T W ·' IML codes to obtain the matrices QO' Q1' and Q given above have been added to the program. This will help users of the GEE program to verify the adequacy of their guessed "working" correlation matrices. 3.4 Logistic Regression Logistic regression is used widely to study the relationship between a dependent dichotomous variable and eeveral independent variables which may be all continuous, all 34 categorical or a combination of the two. The logistic regression model is given by the mathematical function: w-(x) =Prob(Y=llx)= 1+expt-Q-~x) where where Q and ~ -00 < x < 00 (3.31) are unknown regression parameters. The function w-(x) is monotonic:ally increasing from 0 to 1 .. x increases from -00 to +00. Since w-(x) h.. a sigmoid shape like that of many epidemiologic events, the logistic function is often used by epidemiologists and statistici&nl to model risk or probability of disease development during lOme specified time interval .. a function of several independent variables (Kleinbaum, Kupper and Morganstern, 1982). The function is approximately linear for 0.2 < ",(x) < 0.8. There are two methods of estimating the logistic risk function - the discriminant function (DF) approach suggested by Truett, Cornfield, and Kannel (1967) and the iterative weighted least squares method which is equivalent to the maximum likelihood • method (ML) suggested by Walker and Duncan (1967) and Day and Kerridge (1967). The DF approach assumes normality of the independent variables while the ML approach requires no assumptions on the form of the distribution of the independent variables. Next we discuss the two estimation approaches. 1. The Discriminant Function Approach In the case of the discriminant method, the posterior probability of an event occurring based on the logistic model is developed directly from the assumption of normality of the independent variables and Bayes' rule. That is, Prob[y.=lIX.] = I 1 (3.32) 35 where the conditional distribution of X.1 given y.1 = 1 is assumed to be normal (PI' 0'2) and the conditional distribution of X. given y. = 0 is assumed to be normal (PO' 0'2); that is, the 1 1 conditional distributions have different means, but the same variance. By equating (3.31) and (3.32) it can be shown that Q = -In[(I-p)/p] - [1/(20'2)][(P1-PO)(P1 + PO)] fJ =(PI - pO)/0'2 and p is the unconditional probability that Y i = 1. Sample estimates of Q and fJ are obtained by substituting the maximum likelihood estimates of PI' PO' P and 0' 2 Specifically where where So 2 and S2 1 are sample variances of Xi given Yi = 0 and Yi = I, respectively and KO and K 1 are the sample sizes. 2. The Maximum Likelihood Approach The ML approach (Walker and Duncan, 1967 and Halperin, Blackwelder, and Verier, 1971) requires no assumption about the distribution of the Xi's 36 The likelihood function for an event occurring is The logarithm of the likelihood is Taking the derivative of the logarithm of the likelihood function with respect to Q and f3 we obtain the maximum likelihood equations 810 L = ~ y. _ J... x. p. = 0 ~ 2..., vQ i=1 81n L_ '"'a'ir op t • 1=1 I i=1 I I k y.x. - "x.P.. I I k..J 1 1 These equations are solved iteratively using the Newton-Raphson technique, which is based on the first order Taylor series expansion of T(6) = 81n L 8(Q,P) • 6 r +1 such that = 6• r + [T'(6• r )r 1 T(I)• where 6 =(Q,P)' The asymptotic variance-covarianee matrix is the Frechet-Cramer-Rao lower bound which is - {E[T'(6)]r 1 (Brooks et al 1988). 37 Halperin, Blackwelder and Verter (1971) did a comparative study of the DF and ML approaches to logistic regression using both theoretical and empirical development techniques. They showed, theoretically, that for the case of one independent variable, if P = 0, the DF approach gives correct estimates of a and p. If, however, P :;:. 0, the estimates of a and P may differ markedly from the true population values of a and P even when the logistic regression model holda. Empirically, these authors concluded that, the DF approach works reasonably well, even for nonnormal data such as thoee from the Framingham Study. However, they found that the ML method usually gave a slightly better fit to the model, as evaluated from the observed and expected number of cases per decile of risk. The overall conclusion of these authors was that the ML approach is preferable for most situations commonly encountered in epidemiology. Brooks, Clark, Hadgu and Jones (1988) performed a Monte Carlo simulation study to compare the robustness to the assumption of normality of the discriminant function and maximum likelihood methods of estimating the logistic regression coefficients for the simplified case of one independent variable. These authors generated data from four distributions - the normal, exponential, Bernouilli, and Poisson. The overall conclusion of this study is that when the data are normally distributed, both the DF method and the ML method provide good estimates of a and /3. For nonnormal distributions and moderate to large sample size, the ML estimates are preferred; however, for very nonnormal distributions and small sample sizes and highly significant beta, alternate methods to both procedures should be sought. A:I1 example of such a case is attached in the next page which is adapted from Brooks et al. Table 3.1 Comparison of the Discriminant and Maximum Likelihood Estimates of the Logistic Regression Coefficients Poisson po Uncondi tional Probability Sample Size 0.05 500 0.2.5 Model Parameters 50 100 500 0.50 50 100 .500 =0.5 Maximum Likelihood Estimates Disl..Timinant Estimates ,.. Rei-bias 6 ,..a S.E.~ -2.94 -2.740 -2.512 0.0 -0.5 2.0 -2.983 -2.012 -2.709 1I.01lS 0.007 11.007 0.001 -0.406 -0.903 0.010 0.000 0.004 -1.099 -0.902 -0.666 0.0 -0.5 -2.0 -1.148 -0.981 -0.800 11.014 0.014 0.013 OJlO4 -0.434 -1.151 0.016 0.015 0.012 -1.099 -0.902 -0.666 0.0 -0.5 -2.0 -1.146 -0.940 -0.788 0.010 0.003 0.008 0.013 -0.453 -1.133 0.011 0.010 0.008 ---- 0.000 0.197 0.432 0.0 -05 -2.0 -1.100 -0.931 -0.784 0.004 0.004 0.004 -0.004 -0.439 -1.109 0.005 0.004 0.004 ----- 0.000 0.197 0.432 0.0 -0.5 -2.0 -0.006 0.210 0.466 0.012 0.011 0.013 0.017 -0.506 -1.663 0.014 0.015 0.021 ---0.011 -0.168 0.000 0.197 0.432 0.0 -0.5 -2.0 0.00.5 0.205 0.450 0008 0.008 0.009 0.007 -0.493 -1..581 0.009 ----- O.oII 0.012 0.015 -0.209 0.000 0.197 0.432 0.0 -0..5 -2.0 -0.005 0.196 0.433 0.004 0.003 0.004 0.010 -0.492 -1..533 0.004 0.004 0.005 -0.015 -0.233 a Roundml wu DeTfonned after ccrnOUIatIOOS. 6 S.E.(~ or8' -----0.189 -0.549 -----0.131 -0.424 -0.093 -0.433 -0.122 -0.446 ---- Rei-bias ...a S.E.~ it -2.963 -2.761 -2.528 0.008 0.008 0.007 -0.086 -0.621 -2.204 0.021 0.023 0.033 ---0.242 0.102 -1.123 -0.93.5 -0.673 0.014 0.014 0.013 -0.116 -0.676 -2.313 0.036 0.040 0.05.5 0.352 0.156 -1.132 -0.904 -0.663 0.009 0.009 0.008 -0.036 -0.637 -2.301 0.024 0.027 0.040 0.273 0.150 -1.097 -0.906 -0.664 0.004 0.004 0.004 -0.020 -0.527 -2.046 0.010 0.010 0.015 0.254 0.023 -0.007 0.211 0.447 0.012 0.011 0.011 0.038 -0..5.57 -2.430 0.030 0.039 0.056 ---0.110 0.213 0.005 0.20.5 0.444 0.008 0.008 0.007 0.014 -0..516 -2.1.53 0.019 0.026 0.031 0.037 0.076 -0.00.5 0.196 0.432 0.004 0.003 0.003 0.019 -0.510 -2.02.5 0.008 0.010 0.013 --0.020 0.013 S.E.<f) ort ---- ---- ---- ---- Extracted from Broots. Clar .. HJJdgu nd Jones w 00 CHAPTER 4 THE EPIDEMIOLOGY OF PELVIC INFLAMMATORY DISEASE AND ECTOPIC PREGNANCY The fIrst part of this chapter summarizes the descriptive epidemiology of ectopic: pregnancy and di.scuues proposed risk factors for ectopic pregnancy. The second part diacusses the epidemiology of pelvic inflammatory disease including the etiology and diagnosis of PID . It also addresses the biological plausibility of PID as a risk factor for the development of ectopic pregnancy. 4.1 Ectopic: Pregnancy Ectopic: pregnanc:y is a major health problem world wide. Ectopic: pregnancy occurs when a fertilized egg is implanted outside the uterine corpus. The most common implantation sites are the fallopian tubes, with less frequent occurrence in the ovaries, cervix and other abdominal areas (Pritchard et ai, 1985). Ectopic pregnancy increases have been documented in many parts of the world (Beral, 1975; Glebatis et al, 1983; Meirik 1981; Hockin et al, 1984; Westrom et al 1981). In the United States alone the annual incidence of ectopic pregnancy increased from 4.5 per 1000 pregnancies in 1970 to 14.3 per 1000 pregnancies in 1983 (CDC; MMWR, 1986), increase. & three fold In Lund, Sweden the rate of ectopic: pregnancy per 1000 diagnosed pregnancies increased from 5.8 during the early 1960's to 11.1 in the late 1970's for women of reproductive age. Ectopic: pregnancy has been referred to as an unqualified disaster in human reproduction 40 (DeCherney 1983). It presents a life-threatening event, accounting for about 10% of all maternal mortality (Schnider, Berger, Cattell, 1977). Although the annual maternal mortality rate from ectopic pregnancy in the US has declined over the years, it remains as the leading cauae of maternal deaths during the first trimester or pregnancy (Atrash et al, 1986; Frau et al 1986; Budnick and Parker, 1982). The recunence rate of ectopic pregnancy is 10 to 20 percent (BroDJOn, 1977) and 20 to 40 percent of women with ectopic pregnancy will not be able to conceive again (Mueller et al, 1987, Sherman et al, 1987). Furthermore the consequences or ectopic pregnancy are alao associated with great emotional stress, can have a major effect on a woman's sexual as well as reproductive health by threatening her fertility. In the developing nations of Africa, Asia, and South America, where health care services are limited or unavailable for the majority of the population, the morbidity and mortality of ectopic pregnancy are even more serious. Thus the prevention of ectopic pregnancy by identifying its poeeible risk factors is a significant and useful public health effort. Etiology A. Pelvic Inflammatory Disease A number of ease-control studies have shown an association between pelvic infection and ectopic pregnancy, with adjusted odds ratios ranging between 2.0 to 7.5 (Bren et al, 1974; Levin et al, 1982; WHO, 1985). A cohort study of pelvic inflammatory disease undertaken in Lund, Sweden provided t.he strongest evidence of an association between ectopic pregnancy and pelvic inflammatory disease. The hypothesis that pelvic inflammatory disease is associated with ectopic pregnancy is based upon the fact that PlO-causing organisms such as Neisseria gonorrhea, Chlamydia trachomatis, uroplasmas and mycoplasmas ascend to the fallopian tubes causing tubal infection and eventually tubal occlusion thus preventing the implantation of a fertilized ovum in the uterus. ~ 41 B. Intrauterine Device A number of case-control studies have shown an association between IUD's and ectopic pregnancy, with adjusted odds ratios ranging between 0.5 to 32.2 (Table 4.1). The hypothesis that IUD's are associated with ectopic pregnancy is based upon studies that have shown that a device in utero affects tubal motility and ovum transportation (Mastroianni, Rosseau, 1965; Bengtsson, Mowad, 1966) C. Oral Contraceptives The uaociation between ectopic pregnancy and oral contraceptives has been studied for both prior to and at time of conception. A. .hown in the Table 4.2 except the Women'. Health Study and the WHO study with non-pregnant matched controls there does not seem to exist an association between ectopic pregnancy and oral contraceptives as evidenced by the 95% confidence intervals for the odds rati06. 4.2 Pelvic Inflammatory Disease Pelvic inflammatory disease is a serious and common health problem for women of reproductive age. Every year an estimated one million women in the United States alone experience an episode of acute pelvic inflammatory disease; resulting in 1.7 million private office visits (Blount et al, 1983) and 260,000 hospitalizations (Washington A, et ai, 1984). Pelvic inflammatory disease has long been known as one of the major causes of female infertility (Bedbug, Epeyz, 1958; Westrom, 1975) and increases a woman's risk' for chronic abdominal pain, ectopic pregnancy and recurrent pelvic inflammatory disease (Westron, 1975; Westrom, 1980). According to Westrom, 11% of women with one episode of pelvic inflammatory disease develop tubal occlusion, and 23% women with two episodes of pm develop tubal occlusion. The tubal occlusion rate increases to 54% for women with three or more episodes of PID. Table 4.1 Selected case-control studies to evaluate the relation between thc use of an intrautcrine device (IUD) and the occurrence of ectopic pregnancy Study location and time author No. of cases No. and type of controls Odds ration (95% Confidence interval) IUD use at conception Past IUD use Controlled variables Finland, 1972-1974 Rantakyala 71 ROO legal abonion 6.4 (1.1-13.0) Not available None Australia, 1971-1977 Pagano 287 3,200 obstctric 32.2 (21.8-47.7) Not available None Finland, 1976 Savolainen ct aI 136 196 legal abonion 155 delivery 16.3 (9.6-27.8) 1.0 (0.3-3.0) None; excluded current users when estimated risk for past use Boston, 1976-1978 Levin et al 85 498 delivery Not available 1.4 (0.7-2.5) 13 variables; all multigravida; excluded current IUD users Women's Health Study, 1976-1978 Ory 615 3,453 nonpregnant hospital 0.8 (0.6-1.0) 1.4 (1.1.-1.7) Age, race, parity, previous ectopic pregnancy pelvic inflammatory disease Seattle, 1975-1980 O1ow et aI 155 456 populationbased delivery Not available 1.7 (1.1-2.6) Race. gravidity. smoking status, reference year, condom us douching; excluded current IUD users World Health Organizations, 1978-1980 1,108 1; 108 pregnant 1.108 nonpregnant matched triplets 6.4 (3.3.-12.2) 0.5 (0.3-{).7) 0.7 (0.4-1.2 0.7 (0.1-3.9) Age, parity, marital status, hospital of treatment WHO Adopted from ChOw et aI ~ ~ N Table 4.2 Selected case-eontrol studies oral contraception (OC) use and the occurrence of ectopic pregnancy Odds ration (95% Confidence interval) OC use at conception Past OC use Study location and time (author) No. or cases No. and type or controls Finland, 1972-1974 Rantakyala, et al 61 780 legal abortion 3.1 (0.9-10.4) Not available None; excluded current users of intrauterine devices Finland, 1976 Savolainen, et al 81 182 legal abortion 155 delivery 2.1 (0.6,7.1) 0.5 (0.2-1.5) None; excluded current users of intrauterine devices Boston, 1976-1978 Levin, et al 85 498 Delivery Not available 0.6 (0.3-1.2) 13 variables; all multigravida 475 2,923 nonpregnant hospital 0.1 (0.1-0.2) Not available Age, race, parity, previous ectopic pregnancy, pelivic inflammatory disease; excluded past IUD users 1,108 1,108 pregnant 1,108 nonpregnant mathced triplets 2.2 (1.0-5.0) 0.1 (0.1-02) Not available Not available Age, parity, marital status, hospital of treatment Women's Health Organizations, 1978-1980 (48) WHO World Health Organizations, 1978-1989 (48) WHO Controlled variables Adopted from Chow et al J::V-J Economically, PIO and its consequences also represent a substantial economic burden to society. The direct and indirect costs including PIO-related ectopic pregnancy and infertility costs are staggering - an estimated 3.5 billion dollars per year (Peterson, et al, 1991). Both Neisseria gonorrhea and Chlamydia trachomatis, which are prevalent sexually transmitted diseases, have long been recognized as agents causing pelvic inflammatory disease. Other microbial isolates such as mycoplasmas and uroplasmas have also been implicated as agents causing pelvic inflammatory diaeues. Thus the ascent of these microorganisms from the lower genital tract to the upper genital tract is, by definition, a cause of pelvic inflammatory diaeale. Oespite PIO's enormous medical and public health implications the clinical diagnosis of PIO by clinical measures alone is difficult and has a high degree of inaccuracy. In a Swedish study in 1969 acute pelvic inflammatory disease was correctly diagnosed clinically in only 65% of 814 consecutive cases admitted to the University Hospital, Lund, Sweden (Jacobson I, Westrom L, 1969). A poor correlation between the clinical diagnosis of pelvic inflammatory disease and the "correct" diagnosis (laparoscopically confirmed diagnosis) was also reported in two other studies. (Farr, Findlay, 1928; Chaparo, Snehanshu, et al 1978) . Hadgu et al (1986) provide a mathematical model that provides a more accurate diagnostic algorithm for classifying women as having PIO or not. According to this model the best predictors of PIO include: (1) purulent vaginal discharge (2) erythrocyte sedimentation rate >, 15 mm/hr (3) positive gonorrhea result (4) adnexal swelling on bimanual examination and (5) temperature> 38 0 C (6) age less than 25 years (7) single marital status The logistic probability estimates for vanous combinations of the diagnostic variables 45 (Table 4.3) indicate that if a woman is positive for all seven of the predictor variables, then her estimated probability of having PID is 97%. If a woman is positive for all the variables except for purulent vaginal diKharge, then her probability of PID is estimated to be 86%. If all the predictor variables are negative, then the probability of acute PID drope to 7% (Badgu , Westrom, Brooks, 1986). Furthermore Badgu et al developed hybrid models which combine these simple clinical parameters and laparoscopy ( the Gold Standard). For the hybrid model the sensitivity, specificity, and overall classification rates were 100%, 67.2%, and 89.2% • ~ pointed out before, the hypothesis that PID is associated with ectopic pregnancy is based on the theory that PID may prevent or retard migration of the fertilized ovum to the uterus, thus predisposing a woman to ectopic gestation. 46 Table 4.3 Probability estimates of acute pelvic inflammatory disease for various combinations of symptoms based on a logistic regression analysis with no interaction Variables Value Value of prediction variable Intercept -2.5191 1 1 1 1 1 1 1 1 Discharge 1.5333 0 0 0 1 0 1 0 1 Erythrocyte 1.3458 0 0 0 0 0 0 1 1 Gonorrhea culture results 0.9091 0 1 1 0 1 1 1 1 Adnexal 0.6646 0 0 0 0 1 1 1 1 Fever 0.6602 0 0 0 1 1 1 1 1 Age 0.3392 0 0 1 1 1 1 1 1 Marital status 0.3302 0 0 1 1 1 1 1 1 .07 .17 .29 .60 .61 .87 .86 .97 sedimentation rate Probability (acute pelvic inflammatory disease) 0= No I = Yes Extracted from Hadgu et al CHAPTERS ANALYSIS OF THE ECTOPIC PREGNANCY DATA Section 5.1 Description of The Lund Study In this Section we provide longitudinal study consisted of & & simple description of the Lund study. The Lund cohort of 2501 women, under 36 years of age, who underwent a laparoscopic examination at the Department of Obstetrics and Gynecology, University Hospital, Lund, Sweden, Crom January 1, 1960 to December 31, 1984. The catchment area of the Department includes the City of Lund with surrounding rural districts in Southern Sweden. It is the only Department for in-patient treatment of gynecological disorders in women residing in the area. It also serves female non-residents studying at the University of Lund. In 1960 the female population 10-35 years of age was close to 40,000 women, increasing to just over 50,000 in 1985. All women in this study were inpatients with signs and symptoms of acute pelvic inflammatory disease who underwent index (initial) laparoscopy within 24 hours of admission. The laparoscopic technique used was pr~nted in earlier reports (Westrom L, 1975) and remained constant for the period of the study. The intra-pelvic findings of the laparoscopy were graded as follows: Mild £In:. Fallopian tubes reddened and swollen, but freely movable and with normal morphology of the fimbriated ends. Presence of an infectious purulent or seropurulent exudate in the pelvic cavity or leaking Crom the fimbriated ends of the tubes. Moderately severe PID: Tubes not freely movable. Fimbriated ends abnormal or not clearly visible. Infectious exudate an! fibrin deposits on serosal surfaces. 48 ~ fm:. Pelvic peritonitis and/or abecess formation. Laparaseopic visibility significantly impaired by inflammatory masses. Each woman had to satisfy the minimum criteria for clinical diagnosis of PID before she underwent laparoecopy to confirm the clinical diagnosis. The minimum criteria for clinical diagnosis of PID were: From 196~1969: Lower quadrant bilateral abdominal pain or pelvic pain of shorter duration than three weeks plus two or more of the following: 1) abnormal vaginal discharge 2) rectal temperature> 380 C 3) vomiting 4) menstrual irregularities 5) urethritis symptoms 6) proctitis symptoms . 7) marked tenderness of the pelvic organs on bimanual examination 8) adnexal maaa 9) erythrocyte sedimentation rate >15 mm per hour From 1970 • 1984: All three of : 1) low bilateral abdominal or pelvic pain of shorter duration than three weeks 2) purulent vaginal discharge diagnoeed by direct microscopy of a wet mount of the vaginal contents or painful intermenstrual bleeding 3) increased motion tenderness of the uterus and adnexa at a bimanual pelvic examination. There were 3,250 women who fulfilled the clinical diagnosis of PID. However 749 women were excluded from the follow-up study for the following reasons. a) Women subjected to index laparoec:opy after their 36th birthday. These women were excluded because earlier studies (Westrom; 1975) had revealed that pregnancy seeking behavior was rare after the age of 35. b) Women in whom the index laparoeeopy revealed signifieant fertility reducing (large fibroida, endometriosis, genital malformations, etc) factors 49 c) Women who suffered from malignant diseases, chronic diseases of the heart, lung/chest, kidneys, or who had chronic metabolic disorders. d) Women who had consulted for infertility before the index laparoecopy. e) Women who were known drug addicts or suffered from chronic psychotic diseases such as schizophrenia. f) Women who refused laparoscopy or in whom a laparoscopy was contraindicated for other medical reasons. g) Women given a diagnosis of puerperal PID, ie., disease episode within six weeks of delivery. Treatment All patients with PID were treated in hospital with bed rest and antibiotics for at least ten days. During the 25 year period of this study, a number of different antibiotic treatments were used. These include streptomycin/penicillin during the 1960-65 time period; chloramphenicol/penicillin during 1966-71; and ampicillin-doxycycline during the 1971 to 1984 time period. Follow-Up Out of the 2501 women, 1004 were further excluded from the study for the following reasons: a) 514 women who did not expose themselves for the chance of pregnancy, i.e., women who avoided pregnancy voluntarily. b) 221 women who tried to conceive but were unsuccessful due to proven tubal or factor infertility. c) 168 women were lost to follow-up. d) 19 women had history of ectopic pregnancy at initial pregnancy (at baseline). e) 82 had incomplete examination. non-tubal 50 Of the 1497 women who became pregnant after initiallaparoscopy, 744 were pregnant once, 382 twice, 242 three times, 88 four times, and 41 more than four times. Thus a total of 2806 pregnancies are included in this analysis. Of the 2806 pregnancies 1497 were firstorder, 753 were eecond-order, 371 were third-order, and 185 were Courth-order or higher order pregnancies. The major research questions in this analysis include: 1. Does the presence oC pelvic inflammatory disease predispoee a woman to have ectopic pregnancy? 2. Bow has ectopic pregnancy changed since 1960? 3. What other factors predict high risk for ectopic pregnancy? 5.2 Descriptive Analysis Table 5.1 shows the number of pregnancies contributed by each woman during the 19601984 time period. Meet women contributed only one or two pregnancies in this analysis; the mean number of pregnancies contributed by each woman is 1.9 (Table 5.1). One could analyze a single pregnancy from each woman, thus using only 1497 (53%) of the 2806 pregnancies, and resulting in a loes of information and loes of statistical power. The magnitude of the loes in statistical power would be even greater if there were only a small correlation among the pregnancies from the same woman. Furthermore if we limit the study to the outcome of the first pregnancy only we can not distinguish changes in ectopic pregnancy rate over pregnancy order. Biologically, the damage to the fallopian tubes resulting from pelvic inflammatory disease may be irreversible, and if it is serious enough to cause an initial ectopic pregnancy it may also be serious enough to cause subsequent ectopic pregnancies. Thus it is important to study the impact of pelvic inOammatory disease on ectopic pregnancy for each pregnancy order separately as well as all pregnancies together. 51 Simple descriptive statistics are presented in Table 5.2. The risk of ectopic pregnancy increases with increasing age at pregnancy, prior adnexa.l operations and time-period of index lapar08COpy. More importantly the risk of ectopic pregnancy appears to increase markedly with prior PID scores and number of PID episodes. The PIO-score index was constructed by accounting for the number of previous PIO episodes and their severity. A PIO-score of 0 indicates no previous history of PIO; a score of 1 indicates a history of mild PID, and a score of 2 indicates a history of moderately severe to severe PID. Previous ectopic pregnancy wu alao a..ociated with current ectopic pregnancy status. Of the microorganisms isolated at index laparoec:opy only the presence of Mycoplasma at index laparoscopy was significantly associated with ectopic pregnancy. None of the birth controls used at index laparoecopy (IUD, pill, barrier methods) was associated with ectopic pregnancy. However, the information on birth control methods is severely inadequate since it wu obtained only at initial laparoecopy and not during the follow-up time. Note that strictly speaking standard-c:hi squared tests are not appropriate for these data since pregnancies within a woman (cluster) are correlated. 52 Table 5.1. Number of Pregnancies Contributed by Each Woman During the 1960-1984 Time Period. Number of Pregnancies Conoibuted Number of Women Total number of pregnancies 1 744 (50%) 744 (26%) 2 382 (25%) 764 (27%) 3 242 (16%) 726 (26%) 4 88 (6%) 352 (13%) 5 29 (2%) 145 (5%) 6 9 (1%) 54 (2%) 7 3 (0%) 21 (1 %) Total 1497 2806 Average number of pregnancies conoibuted by each women ~= 1.9 1497 53 Table 5.2 Univariate Anal ysis of Ectopic Pregnancy that Occurred During 1960-1984 Time Period at theDepanment of Obstetrics and Gynecology. University Hospital. Lund. Sweden. * Ec«mics * Pregnancies Variable •• Percent P-Value .000 Mother's age at Pregnancy S24 41/1107 3.7 25-30 101/1132 8.92 31 + 51/567 8.99 S 19 7011259 5.56 20-24 8811039 8.47 25+ 35/508 4.7 1960-1967 ron92 7.58 1968-1976 10211267 8.05 1977-1984 31n46 4.16 84/1215 6.91 97/1377 7.04 8/118 6.78 0 104/1494 6.96 1 5'2n54 6.90 2 271372 7.26 3+ 10/186 5.38 0 158/2640 5.98 1 261132 19.70 2+ 9/34 26.47 Age at index laparoscopy 0.023 Index Laparoscopy Time Period .003 Marital Status Married or Cohabiting Single - Other .98 Prior Pregnancies .86 Prior adnexal operations .000 54 Variable P-Value # EctQpics # Pregnancies Percent Prior PIO episodes 0 18n89 2.28 1 148/1789 8.27 2 21/189 11.11 6139 15.38 3+ .000 Previous Still Births 0 188/2769 1+ 5/37 6.79 0.108 13.5 Previous Ectopic Pregnancy 0 168/2639 6.24 1 25/113 22.12 2+ .000 28.57 Prior PIO Score 0 18n89 2.28 1 91/1457 6.25 2 84/560 15.0 .000 Microorganisms Isolated at Index Laparoscopy Gonorrhea (GC) Yes 51n05 7.23 No 142/2101 6.76 Yes 13/190 6.84 No 180/2616 6.88 Yes 44/444 9.91 No 149/2362 6.31 0.66 ECOLI .98 Mycoplasma (MYCO) 0.006 55 Variable P-Value # Ectopics # Pregnancies Percent Strep Yes 42/499 8.42 No 151/2707 6.55 .134 Birth Controls Used at Index laparoscopy IUD (ever) Yes 47n18 6.55 No 14612088 6.99 Yes 51/800 6.38 No 14212006 7.08 Yes 15/246 6.10 No 17812500 6.95 .684 Pill .508 Barrier ** = Chi-squared test .613 56 Figure 5.1 shows the longitudinal ectopic pregnancy rates by PID-score (PIO severity) and pregnancy order. Ectopic pregnancy was related to PID-severity cross-sectionally. Women with no PIO (score = 0) showed lower ectopic pregnancy rates compared to women with PIO (score 1 • acore2) at all pregnancy orden. increased with the severity of tubal damage. The ectopic pregnancy proportion Moreover, the damage did not decline longitudinally; rather, it seemed to stay constant acrou pregnancy order, thus indicating that the risk of ectopic prepancy from PID does not seem to decrease with increasing pregnancy order. Though the ectopic pregnancy proportion for severe PIO decreases at 3rd order pregnancy, a chi-squared teat indicates no statistically significantly diff'erentce (p=0.337) from the proportion of first order or second order pregnancy. However, for women with NO-PIO a chi-squared test indicated a significant increase (p=0.04) in the ectopic pregnancy proportion from fint-order pregnancy to third order pregnancy. Thus though the ectopic pregnancy proportion for women with no PID is less than those of women with mild or severe PIO c.rosssectionally, proportions for the no-PIO groups tend to increase longitudinally, unlike those of PIO women. Figure 5.1 Ectopic Pregnancy Rates by Severity and Pregnancy Order E c t 20 o P I c 15 I:. P r e g n 10 .- a n c y R a t e s 5 .- o I I I I P1 P2 P3+ I Pregnancy Order No PIO ---*- Mild PIO ~ Severe PIO P3+-Pregnancy Order 3 or more P1-Pregnancy Order 1 P2-Pregnancy Order 2 Bars are standard errors -S- Vl '-l 58 5.3 Model fitting 5.3.1 Analysis Using Standard Logistic Regression Model (Ignoring Dependence) In this aection we analyze the ectopic pregnancy data using ordinary logistic regression techniques assuming the pregnancies within a cluster (women) are statistically independent. First our model-building strategy involves seeking the most parsimonious model that explains the data but also includes the scientifically relevant variables (such as PIO and g~nital mycoplasmas) whether they are statistically significant or not. There!ore our model-building strategy and particularly our variable selection proc:esa we take the following into account: 1. A univariate logistic regression analysis of each variable will be undertaken. 2. Any variable whose univariate test has a p-value of < 0.25 will be considered as a candidate for the multivariate logistic regression. The cut-off' point of 0.25 is based upon the work of Bendel and Afifi (1977) for linear regression and Mickey and Greenland (1989) for logistic regression. The problem with this approach is that an explanatory variable mAy be weakly associated with a dependent variable in a univariate analysis; however, it may become an important variable in the presence of other covariates. 3. We will then fit a multivariate logistic regression model and the importance of each variable included in the model will be verified. This will be done by the examination of the Wald statistic for each variable. Variables that do not contribute to the model are then eliminated and a new model is fitted. The new model will be compared to the old model through the likelihood ratio test. 4. Once we have obtained the model that we feel conta.ins the essential variables, we will then consider the need for including interaction terms among the variables. Finally we also use stepwise logistic: regression methods to see if the variables selected using the above variable selection procedure cillIer from those selected by a stepwise logistic regression technique. In order to make the logistic regression analysis easier, an abbreviated code sheet is given 59 in Table 5.3. which describes the independent variables and provides the code names for the variable that will be used. The results of fitting the univariate logistic regression models to these data are given in Table 5.4. In Table 5.4 we show the following: 1. the estimated regression coefficient(s) for the univariate logistic regression model containing only this variable. 2. the estimated standard error of the regression coefficient. 3. the estimated odds ratio, which is obtained by exponentiating the estimated regression coefficient. 4. the 95% CI for the odds ratio. 5. the value of the log-likelihood for the model. 6. the likelihood ratio test statistic, G, for the hypothesis that the slope . coefficient is zero. This is obtained by: G- -2 1 {likelihood without the variable(constant only)} og likelihood with the variable - Under the null hypothesis this quantity will follow the chi-squared distribution with 1 degree of freedom for the independent variables with two categories and 2 degrees of freedom for the independent variables with 3 categories. From Table 5.4 we see that with the exception of NPREV (# previous pregnancies) and ISTATUS (marital status at index laparoscopy), there is evidence that each of the variables has some association with the outcome variable, namely ectopic pregnancy. This observation is based upon an inspection of the confidence interval estimates, which for most variables either do not contain 1 or just barely so. Therefore, based on the univariate results, we begin the multivariate model with all the variables except NPREV and 1STATUS. On the basis of the results shown in Table 5.5 it appears that all of the variables except AGE and PREVSTlL show considerable importance 60 in the multivariate logistic regression model. This suggests that we fit a new model which does not contain AGE and PREVSTIL. The results of this reduced model are displayed in Table 5.6. The likelihood ratio test for the difference between the models in Table 5.5 and 5.6 yields a value of G = 1256.9 - 1256.7 = 0.2. Comparing this to the chi-squared distribution with 2 degrees of freedom yields a non-significant ~value, demonstrating that the variables AGE and PREVSTIL add little to the model once the other variables are retained. Also note that the inclusion or exclusion of AGE and PREVSTIL to the model does not alter the values of the other regression coefficients by much. An alternative approach to variable selection is to use the stepwise method in which variables are selected either for inclusion or exclusion from the model in a sequential fashion based on a pre-specified inclusion/exclusion criterion. We used the PROC LOGISTIC procedure with the STEPWISE option and obtained variables identical to the ones shown in Table 5.6. Table 5.3 Code Sheet for the Variables in the Lund Ectopic Pregnancy Data Variable Abbreviation Identification Code ID Ectopic Pregnancy ECf 1 = ectopic pregnancy o = not ectopic pregnancy Age at Index Laparoscopy 0= Age ~ 19 1 = Age ~ 20 Age at Pregnancy 0= Age ~ 24 1 = Age ~ 25 Marital Status at Index Laparoscopy AGE AGEGP ISTATUS o = Married or Cohabiting I = Single 2 = Other Year initial laparoscopy was perfonned LAPID 0= 1977-1984 1 = 1960-1976 Genital Mycoplasma at Index laparoscopy MYCO 0= No 1 = Yes Prior Pregnancies NPREV o = no prior rregnancy 1 = 1 prior pregnancy 2 = 2 or more prior pregnancies Prior adnexal surgeries o = no previous adnexal surgery 1 = 1 or more adnexal surgery NSG 61 62 Variable Abbreviation PID Score (Laparoscopy) SCORE SCORE1=1 (Mild PID) SCOREI SCORE1=O (Otherwise) SCORE2 =1 (Severe PID) SCORE2 =0 (Otherwise) SCORE2 Table 5.4 Variable Univariate Logistic Regression Models for Ectopic Pregnancy Parameter Standard Odds Ratio Estimate Error Estimate 95% c.1. 1) SCOREI 1.0485 .2619 2.85 (1.71,4.77) SCORE2 2.0227 .2662 7.56 (4.49,12.74) 2) AGE .1624 .1090 1.1 H 3) AGEGP .9379 .1H04 4) ISTATUSI .0518 ISTATUS2 -2 log L G p-value 1326.045 79.670 .0001 (0.95,1.46) 1402.941 2.723 .0989 2.55 (1.79,3.64) 1374.52 31.137 .0001 .1526 1.05 (0.78,1.42) 1405.55. .117 .943 0.0107 .3829 1.01 (0.48,2.14) .6773 .2009 1.97 (1.33,2.92) 1392.70 12.964 .0003 6) NPREVI -.0100 .1760 .99 (0.79,1.40) 1405.594 .07 .9655 NPREV2 -0.0522 .1982 .95 (0.64,1.40) 7) NSG 1.4344 .2072 4.20 (2.80,,6.30) 1367.15 38.474 .0001 8) MYCO .4909 .1800 1.63 ( 1.15,2.32) 1398.777 6.887 .0087 9) PREVSTILL .7632 .4868 2.14 (0.83,5.57) 1403.608 2.057 .1515 5) LAPID 0\ W 64 Next we proceed to assess interactions among the variables in the model. A total of 10 . possible two-way interactions were formed from the model in Table 5.6. However the interaction between PIn score (SCORE) and age at pregnancy (AGEGP) could not be assessed because BOme of the cells contained zero observations. At this point we are interested in assessing the contribution of each interaction term to the previously developed model Table 5.6. model are presented in Table 5.7. The result of adding each interaction to this Upon inspection of Table 5.7 we see that there is one interaction term that is statistically significant. This is the SCORE x MYCO interaction. The addition of the SCORExMYCO interaction term appears to provide a significant improvement over the main effects only model (p=0.0497). Therefore the inclusion of this interaction into the model offen the possibility to better describe the effect of PIO and genital mycoplasma on ectopic pregnancy. The estimated coefficients and their standard errors are given in Table 5.8. We now turn to the interpretation of the fitted model in Table 5.8. For covariates not involving interaction terms the estimated odds ratios are obtained by simply exponentiating the estimated regression coefficients. For example, the odds of developing ectopic pregnancy increases by 2.4 fold (=exp(O.8801» for women whose age at pregnancy was 25 or more. Estimated odds ratios and 95% C.I. for the variables not involving interaction terms are given in table 5.9. A modification of the above approach is necessary to estimate the odds ratios for variables involving interactions. Specifically we would like to: (1) Estimate the odds ratio of PIO (SCORE) separately among those women with genital mycoplasma and those without mycoplasma. (2) Estimate the odds ratio of mycoplasma separately among those women with PIO and those women with NO PIO (SCORE1 = 0 and SCORE2 = 0). (3) A fifth odds ratio of interest may be the odds ratio for both PIO and mycoplasma (SCORE=l and MYCO = 1) relative to women with no genital mycoplasma and No-PID 65 (MYCO =0 and SCORE=O ) controlling for other predictors. Each of these odds ratios is obtained by exponentiating an appropriately calculated logit difference. We will first present the logit and the differences in general terms, and then use the values of the estimated regression coefficients in Table 5.9 to get numeric values. First let p·X· represent the contribution by the terms containing AGEGP, LAPIO and NSG, terms that do not involve interactions. Then the values of the estimated logit differences are given in Table 5.10 For example, the logit difference d 1 (Table 5.10) represents the effect of SCOREI (mild pm) among women without genital mycoplasma (MYCO=O). We now turn to the interpretations of the estimated odds ratios in Table 5.11. From Table 5.11 we see that PIO (whether it is mild or severe) by itself is an important risk factor when the woman does not have genital mycoplasma. The estimated adjusted odds ratios for mild and severe PIO among women with no genital mycoplasma are 4.10 and 10.19 respectively. However, among women with genital mycoplasma the odds ratios for mild or severe PIO are not as impressive. The estimated odds ratios for mild and severe PIO among women with genital mycoplasmas at index laparoscopy are 1.08 and 2.60 respectively. These differences in the odds ratios for PIO among those with mycoplasma infection compared to those without mycoplasma infection results from the negative interaction between PIO and mycoplasma. This indicates that a positive finding for PIO (or mycoplasma) is important if the finding for mycoplasma (or PIO) is not positive. Similarly we see that genital mycoplasma is also a significant risk factor among women who do not have PIO. The estimated odds-ratio for mycoplasma among women with no PIO is 5.10. The estimated odds-ratios for mycoplasma among women with mild PIO and severe- PIO are 1.34 and 1.30 respectively. When taken together, the odds ratio for both mild PIO and mycoplasma (SCOREI=1 and MYCO=I) relative to women with no PIO and no mycoplasma (SCORE1=0 and MYCO=O) is 5.51. The odds ratio for both severe PIO and mycoplasma (SCORE2=1 and MYCO = 1) relative to women with no pm and no mycoplasma (SCORE 2 = 0 and MYCO=I) is 13.27. Table 5.5 Estimated Coefficients and Standard Errors for the Multivariate Model Containing Variables Identified in the Univariate Analysis Variable Parameter Estimate Standard Error Wald Oli-Square Pr> Oli-Square INTERCPT -4.9152 0.3303 221.5086 0.0001 SCOREI 1.0159 0.2639 14.8184 0.0001 SCORE2 1.9127 0.2707 49.9265 0.0001 AGEGP 0.8813 0.2053 18.4194 0.0001 AGE -.0217 0.1151 0.0356 0.8504 LAPID 0.5462 0.2078 6.4056 0.0086 NSG 0.9608 0.2210 18.8962 0.0001 MYCO 0.4465 0.1898 5.5316 0.0186 PREVSTlLL 0.1703 0.5258 0.1048 0.7461 -2log L = 1260.394 66 Table 5.6 Estimated Coefficients and Standard Errors for the Multivariate Model Containing Variables Significant in Table 5.5. Parameter Estimate Standard Error Wald Oli-Square Pr> Variable INTERCPT 4.9266 0.3267 227.3643 0.0001 SCORE1 1.0169 0.2638 14.8656 0.0001 SCORE2 1.9174 0.2699 50.4633 0.0001 AGEGP 0.8677 0.1855 21.8819 0.0001 LAPID 0.5506 0.2073 7.0540 0.0079 NSG 0.9664 0.2200 19.2888 0.0001 MYCO 0.4505 0.18902 5.6816 0.0171 -2log L = 1260.540 Chi-Square 67 68 Table 5.7 Log-likelihood, Likelihood Ratio Test Statistic (G), Degree of Freedom. and P-Value for Possible Interactions of Interest to be Added to the Main Effect Model. Interaction -2 log-likelihood G d.f p-value Main Effects Only 1260.5 SCORE x LAPIn 1260.1 0.4 2 .8183 SCORE x NSG 1259.5 1.0 2 .6065 SCORE x MYCO 1254.5 6.0 2 .0497 AGEGP x LAPID 1260.5 0.0 1 1.0000 AGEGP x NSG 1260.4 0.1 1 0.7518 AGEGP x MYCO 1260.5 0.0 1 1.0000 LAPID x NSG 1260.4 0.1 1 0.7518 LAPID x MYCO 1257.2 3.3 1 0.0693 NSG 1259.2 1.3 1 .2542 x MYCO • . Table 5.8 Estimated Co-efficient and Slandard Errors for the Multivariate Model Containing Main Effects and Significant Interactions. Variable Parameter Estimate Standard Error Wald Chi-Square Pr> Chi-Square INTERCPT -5.3073 0.4111 178.4401 0.0001 SCORE1 1.4112 0.3647 20.4032 0.0001 SCORE2 2.3215 0.3709 47.0801 0.0001 AGEGP 0.8801 0.1862 24.0466 0.0001 LAPID 0.5652 0.2086 7.7541 0.0054 NSG 0.9375 0.2207 16.7725 0.0001 MYCO 1.6290 0.4992 9.5599 0.0020 SCORE1 x MYCO -1.3336 0.5708 4.9463 0.0261 SCORE2 x MYCO -1.3653 0.5828 4.9977 0.0254 Log-likelihood = 1254.5 Table 5.9 Estimated Odds Ratios and 95% C.I for variable not involving interation terms: Variable Estimated Odds Ratio 95% C.I. AGEGP 2.41 (1.67.3.47) LAPID 1.76 (1.17.3.47) NSG 2.55 (1.66.3.94) 69 70 Table ~lO Expression for Ole Logita and Logit Differences in Termll of the Estimated Parameters for the P08lIible Combinations of MYCO and SCORE. Score (pm) MYCO o 1 o Logit Difference d 1 =iJ 1 SCORE 1 • + iJ*X* + iJ 3(SCORE x MYCO)+ iJ*X* iJ 3(SCORExMYCO) Logit DiB'. iJ 3(SCORExMYCO) 71 Table 5.11 Values of the Estimated Legit Differences and Odds Ratios Effect Among Legit Difference Odds Ratio SCOREI = 1 MYCO=O 1.4112 4.10 SCOREI = 1 MYCO = 1 (1.4112-1.3336) 1.08 MYCO = 1 SCOREI = 0 1.6290 5.10 MYCO = 1 SCOREI = 1 1.6290-1.3336 1.34 (1.4112 + 1.6290 - 1.3336) 5.51 SCOREI + MYCO SCORE2 = 1 MYCO =0 2.3215 10.19 SCORE2 = 1 MYCO= 1 (2.3215-1.3653) 2.60 MYCO = 1 SCORE2 = 0 1.6290 5.10 MYCO = 1 SCORE2 = 1 (1.6290-1.3653) 1.30 (2.3215 + 1.6290 - 1.3653) 13.27 SCORE2 + MYCO 72 Analysis !.Wllg Q.EE Techniques 5.3.2 In this section we use the GEE method explained in Chapter 3 to fit the ectopic pregnancy data, with the logit link function h(p) = log-IP and the variance function V(p) = -p p(l-p). The GEE method was used to model the marginal dependence of E(Yit) on the covariate:s. Table 5.12 shows estimated regression coefficients and Z-statistics for a binomial error and a logit link model. Three different correlation structures , namely independent, exchangeable, and I-dependent correlation structures, are fitted to investigate the effect of specifying different correlation structures. One-dependent correlation structure assumes that correlations one occasion apart are the same (but non-zero) and correlations two or more occasions apart are zero. From the magnitude of the regression coefficients of the model with one-dependent correlation structure, it can be seen that pelvic inflammatory disease is the strongest predictor of ectopic pregnancy, followed by history of genital mycoplasma at index laparoscopy and age at pregnancy. The presence of severe pelvic inflammatory disease increases the odds of having ectopic pregnancy by 10.93 ( = exp(2.3918)) fold for women without genital mycoplasma. This odds increases to 14.38 (=exp(2.3918+1.654o-1.3803) for women with genital mycoplasma. Similarly, the odds of developing ectopic pregnancy after mycoplasma are 5.23 times higher (=exp(1.654) for women without PID. The estimate of the 'working' correlation, Q, is 0.089. Similarly, the estimate of the scale parameter is 0.9532 for the I-dependent correlation structure. One can note that for dichotomous outcomes the true value of 4J is one. Thus a way in which the results of Table 5.12 could be adjusted to a scale parameter of one is through division of regression estimates and their standard errors by the square root of the estimated scale parameter, e.g., for I-dependence division by the square root of 0.9532 would be applied. Since such results would be similar to those already shown no such adjustment has been made. Furthermore, we find the computational procedure to be reassuring by providing a scale parameter estimate which is so close to the known value of one. 73 For the time-independent variables note that the coefficients and the z-statistics for the three correlations are similar. However, for the time-dependent variable, NSG, the coefficients and their associated robust z-statistics are different under the three correlation structures. As pointed out by Zeger and Liang (1986) " the effect of assuming a non- independent correlation structure is to use weighted linear combinations of both Y's and X's for each subject in GEE. It is therefore sensible that covariates which vary with time may experience larger changes in their coefficients than do time-independent covariates for finite sample." 74 Table 5.12 GEE Analysis of Ectopic Pregnancy with Legit Link and Binomial Variance Correlation Structure Estimate Intercept independent exchangeable I-dependence -5.3073 -5.3108 -5.3410 SCORE1 independent exchangeable 1- dependence SCORE2 S.E. - Robust z- Robust 0.382 0.401 0.397 0.421 0.429 0.425 -12.60 -12.39 -12.56 1.4112 1.4284 1.4326 0.336 0.355 0.349 0.363 0.370 0.365 3.89 3.86 3.92 independent exchangeable l-dependence 2.3215 2.4263 2.3918 0.342 0.360 0.355 0.370 0.376 0.373 6.27 6.45 6.42 LAPID independent exchangeable I-depcndence 0.5651 0.6085 0.6024 0.204 0.214 0.211 0.217 0.222 0.219 2.61 2.74 2.74 AGEGP independent exchangeable I-dependence 0.8801 0.8818 0.8910 0.182 0.179 0.183 0.189 0.184 0.189 4.65 4.79 4.71 NSG independent exchangeable I-dependence 0.9375 0.3495 0.6097 0.215 0.247 0.234 0.222 0.291 0.252 4.23 1.20 2.42 MYCO independent exchangeable I-dependence 1.6270 1.6295 1.6540 0.483 0.523 0.506 0.527 0.544 0.540 3.09 2.99 3.06 SCORElx MYCO independent exchangeable I-dependence -1.3336 -1.3375 -1.3575 0.554 0.596 0.578 0.588 0.610 0.604 -2.27 -2.19 -2.25 SCORE2 x MYCO independent exchangeable I-dependence -1.3655 -1.3899 -1.3805 0.566 0.605 0.588 0.615 0.634 0.629 -2.22 -2.19 -2.19 Covariate For independent correlation structure scale parameter = 0.9618 For exchangeable correlation structure scale paramenter = 0.9441 working correlation = 0.1043 For l-depcndence correlation structure scale pararnenter = 0.9532 working correlation = 0.089 S.E Naive 75 5.3.2.1 Selection Between ~ IIu:tt Correlation Structures Comparison of the naive z-statistics and robust z-statistics can be used to evaluate whether a specified correlation structure is incorrect. Note that naive standard errors are calculated assuming that the 'working' correlation structure is the true one. Recall that from equation 3.22 we have that When the correlation structure is assumed to be correctly specified then iO = iI-I. One can therefore expect relatively large differences between the robust and naive z-statistics if the working correlation is far from the true correlation. However, if they are similar it does not necessarily imply that the assumed correlation structure is similar to the true correlation structure. Table 5.13 shows the absolute relative difference between the robust and naive z-statistics for the two correlation structures. For the independent correlation structure, naive z-statistics show larger values than robust z-statistics. This suggests that the variance is underestimated since correlated data are treated as independent and counted as separate information. For almost all covariates the independent correlation structure showed relatively higher relative difference in absolute terms. Both the exchangeable and one-dependent correlation structures appear to have similar relative changes. Perhaps another way of choosing among 'correlation structures is to use the technique of Rotnitzky and Jewell given in Section 3.3. Table 5.14 gives the eigenvalues of the matrix Q defined in Section 3.3. From Table 5.14 we see that (~1' ~2) values for the three different correlation structures are similar with the I-dependent correlation structure having values slightly closer to one, but not by much. Rec.all that i l is the average of the eigenvalues of matrix Q and ~2 is the average of the squared eigenvalues of matrix Q given in equation 3.26. Table 5.13 Absolute Relative Difference Between Robust and Naive Statistics for the Independent and Exchangeable Correlation Structure. Exchangeable Independent Covariate Z~Robust Z-Naive Relative Change Z-Robust Z-Naive INTERCEPT -12.60 -13.88 10.1% -12.39 -13.25 SCOREI 3.89 4.21 8.2% 3.86 SCORE2 6.27 6.80 8.4% LAPID 2.61 2.77 AGEGP 4.65 NSG MYCO I-Dependence Relative Change Relative Change Z-Robust Z-Naive 6.9% -12.56 -13.46 7.2% 4.02 4.1% 3.92 4.10 4.6% 6.45 6.75 4.6% 6.42 6.74 5.0% 6.1% 2.74 2.85 4.0% 2.75 2.86 4.0% 4.84 4.1% 4.79 4.92 2.7% 4.71 4.88 3.6% 4.23 4.35 2.8% 1.20 1.41 17.5% 2.42 2.61 7.9% 3.09 3.37 9.1% 2.99 3.12 4.3% 3.06 3.27 6.9% INTI -2.27 -2.41 6.2% -2.19 -2.24 2.28% -2.25 -2.35 4.4% INTI -2.22 -2.41 8.6% -2.19 -2.30 5.0% -2.19 -2.35 7.3% INTI = SCOREI X MYCO INTI = SCORE2 X MYCO ..... 0\ Table 5.14 The Eigenvalues of Malrix Q for the Independent and Exchanageable Correlation Structures. Eigenvalues Independent Exchangeablc I-Dependent c. 1.4268 1.5030 1.3689 C2 1.2835 1.3357 1.3319 cJ 1.2412 1.2015 1.2208 e. 1.1809 1.1133 1.1587 C5 1.0689 1.0459 1.0410 e6 ].0332 0.9979 1.0039 C7 0.9832 0.9910 0.9780 c8 0.8997 0.8995 0.9107 c9 0.8380 0.8181 0.8247 c. 1.1062 1.1117 1.0932 c2 1.2563 1.2740 1.2264 ...... ...... 78 5.3.3 Analysis Using Rosner's Beta-Binomial Model In this section we provide a summary of Rosner's (1984,1989) regression technique for correlated binary data. Rosner's model was derived as an extension of the beta-binomial model of Williams (1975). For simplicity, we first restrict our attention to the case where each primary sampling unit has exactly two subunits. Let Y1 and Y2 be the binary outcomes associated with the two units in the ith pair. Associated with each member of the pair is a pdimensional vector of covariatea xi' Let us now consider a primary sampling unit with x =Xo = (0 0 •.• 0). Roener then the beta binomial distribution to model the probability of the four disease states U8e8 (++, +-, -+, -) for such a PSU or cluster. Specifically, he assumes that this cluster has probability Pi of having any particular sub-unit affected, where Pi follows a beta distribution with parameters a and b over the class of all individuals with covariates xO' It then follows that the probability of disease states for such a PSU is given by 1 P(++ I xO) = of p.2 p.a-1 (1- p.)b-1 1 1 1 r(a+b) dp. r(a)r(b) 1 (a+l)a = (a+b+1)(a+b) Similarly P(+ - I xa) = P(-+) = (a+b+l)ab (a+b) (b+1)b P(- I xO) = (a+1>+1) (a+b) Then the probability distribution of disease states for a PSU with arbitrary covariates x. is 1 given through two conditional logistic regression models: 79 In where crl = In ( b ~ 1) , and cr2 = In (b+1)b (a+1)a Rosner then shows that, using (5.1), the joint probability distribution of Y 1 and y 2 It is clear that when cr2 18 = 2cr1 ' Rosner's model reduces to the standard, uncorrelated logistic model for two independent observations. Rosner shows the odds ratio for association between y 1 and Y2' controlling for x, is given by: = (a+1)(b+1) ab Parameter estimation in (5.2) is then undertaken by the method of maximum likelihood. Since there does not exist a closed form of (5.2), an iterative procedure based on the NewtonRaphson method may be employed. The interpretation of a regression parameter of Rooner's model, {3x, is as follows: f3x is the increase in the log-odds of response for subunit holding all other covariates and the response of the fellow subunit fixed. a particular 80 5.3.3.1 Extension to ~ General ~ Let us now consider the general case, where the i th primary unit has ni subunits, i = 1, .., K. Let Yij = 1 if the jth subunit of the primary unit is affected, and 0 otherwise, j = 1,...., ni' Le t ! i !i (1) (0)' ,.... !i 4 (ni) (1)' ,...., Xi (ni)' ' h (0) . . f' . d d . d were Xi IS a matnx 0 tlme-lD epen ent covarlates an .: f' d - d . are matrlcles 0 tlUle- epen ent covanates. Rosner (1984, 1989) generalize the beta-binomial model to accommodate both timedependent and time-independent covariates using a particular polychotomous logistic regression model. Under this model the likelihood for the i th primary unit is given by (0) ni &si bo.-s. P(y-'1 1 exp 1 [.E J=1 YlJ" (.B!i + 1'!i 0) )] I !1') = --------....::...-.;;..w-:---------- E Zi (0) ni a,rz.. IJ b(n._Ez..) 1 exp IJ [E j=1 Zij (.B:i where a and b are the parameters of a beta distribution, + ~ i .B and (5.3) 0) )] • l' are regression coefficients corresponding to time-independent and time-dependent variables, Yij=1 if the jth subunit in = · . IS . auecte - er d an d Oot h erwlse. . the 1.i hpnmary uDlt bm ~~ (b + j), m ~ 1, aO F urt h ermore, am m-l = j=O 11' ( a+J, .) = b O = 1, and the summation in the denominator. is over all possible permutations z· = (z.l' z'2' .., z· ) of 0'5 and l's. And s· denotes the number of 1 1 1 lDi 1 subunits of the the i th primary unit which are affected. If no covariates are present, then according to Rosner (1989) the denominator in (5.3) is (a+b)n.1 and Prob(y.) reduces to _1 lsi' bo.-s. / (a+b)n.. which is a beta model with parameters a and b. Moreover for dusters 111 of size 1, the model reduces to ordinary logistic regression and the coefficients of fJ and l' have a marginal interpretation. The full likelihood is then given by L = II P(Yi I Xi) i and the Newton-Raphson technique 81 is used to obtain maximum likelihood estimates and perform significance tests concerning a, b, p, and "'(. The measure of dependence in this model is the pairwise odds ratio between subunits, (a + 1) (b + 1)/ ab, where a and b are the parameters of a beta-binomial distribution. Rosner dermes the pairwise odds ratio to be the odds of disease for one cluster member given that another cluster member is also affected, divided by the comparable odds given that the other cluster member is not affected. 5.3.3.2 Comparison with Ordinary Logistic Regression AD. important distinction between Roener's model (5.3) and the ordinary logistic regression model is that Rosner's model explicitly includes the outcome of previous visits as additional predictor covariates, whereas ordinary logistic regression treats the visits within a cluster as separate and independent observations. In a subsequent paper Rosner and Milton (1988) have presented an approach for obtaining appropriate ordinary logistic regression estimates and standard errors from Rosner's regression estimates: below: l: . 1 = n·1 ( a+b+n.) 1 (a+b+l)l: n·1 • 1 l: n.[a.+b+1+(n.-1)p] . 1 1 1 = Var(pp)OL Var(pp)R 1 (a+b+l) = = 1 a+b+l E ni Ei n. (a+1>+n. )[x. (0)]2 l l Ip 1 a+b+l ~{[a+1>+1 +(ni-1)Pe]4= [xie(j)]2} 1 J These results are given 82 for balanced data (ni = n), we have = Var(re)OL = a+b+l+{n-l)Pe Var(re)R a+1>+1 where {{3 P)OL and {{3 P)R are the time-independent regression coefficients from the ordinary logistic regression model and Rosner's model. Similarly respectively {re)OL and ("Ye)R are the time-dependent regression coefficients Crom the ordinary logistic regression and Rosner'. model respectively. The above formulae are obtained by assuming all the covariates are independent and that the regression of the eth time-dependent covariate, #- k) Xe 0) on Xe (k) (j is approximately linear (Rosner and Milton, 1988). Glynn and Rosner (1993) in a simulation study investigated the performance of Rosner and Milton's conversion formula for both time-dependent and time-independent covariates. These authors conclude that the conversion formula performed well in terms of coverage probability, consistency, and bias. 5.3.3.3 Application of Rosner's Model to the Ectopic Pregnancy Data We now present the application of Rosner's beta-binomial model to the ectopic. pregnancy data described in Section 5.1. Estimates of these regression coefficients that condition on the status of previous pregnancies were obtained by using the polychotomous logistic regression program peHLE of Rosner which was written in FORTRAN IV. Table 5.15 summarizes results of the beta-binomial modelling approach with the use of the Lund ectopic pregnancy data set. From the magnitude of the regression coefficients of the model, it can be seen that pelvic inflammatory disease is the strongest predictor of ectopic pregnancy I followed by history of genital mycoplasma and age at pregnancy. The conditional pairwise odds ratio of ectopic pregnancy between two pregnancies is , 83 (a + 1) (b + 1)/ (ab) = 2.2. The conditional pairwise odds ratio of 2.2 means that on the average the odds of a subsequent ectopic pregnancy increases by 2.2 fold if the prior pregnancy was ectopic after having adjusted for the effects of the covariates given in Table 5.15. The formal interpretation of the regression parameters for Rosner's beta-binomial model is as follows: /l = odds of developing ectopic pregnancy at pregnancy order t for an exposed (e.g. PID) versus an unexposed (e.g. no-PID) woman at pregnancy order t conditional on: (a) values of all other covariates being the same at pregnancy order t. (b) outcome status at pregnancy 1,2,..., t-I the same (or # ectopic pregnancies among the previous t-I visits must be the same) for exposed and unexposed women. Thus for binary longitudinal data of the Lund ectopic pregnancy type we have: logit (Pr (ectopic pregnancy at the tth pregnancy order I ~ , outcome status at visit 1, ..., t.- 1)] + f3 x (0) - - + r x (t) -- where a and b are the two parameters from the beta-binomial distribution and S_t is the number of ectopic pregnancies out of the previous t.-I pregnancies. Thus the interpretation of regression coefficients is not invariant with respect to cluster size unlike the coefficients of the marginal model, where the interpretation of fJ is the same regardless of cluster size. 84 Table 5.15 Variable Results of the Ectopic Pregnancy Data Using Rosner's Beta-Binomial Modeling Approach P-value Parameter Standard Estimates Error 2 sided PARISE a 0.89963 0.40438 2.22468 0.2610 b 18.25537 8.41525 2.16932 0.03006 LAPID 0.47757 0.19926 2.39676 0.01654 MYCO 1.45227 0.40534 3.58279 0.00034 SCOREI 1.28516 0.30068 4.27417 0.00002 SCORE2 2.08186 0.30703 6.78053 0.00000 AGEGP 0.82971 0.18338 4.52463 0.00001 NSG 0.79874 0.21230 3.76239 0.00017 SCIMYCO -1.18991 0.46765 -2.54445 0.01094 SC2MYCO -1.21646 0.46238 -2.63087 0.00852 85 5.3.3.4 Comparison with Ordinary Logistic Regression Ami ~ Models Table 5.16 summarizes the results from ordinary logistic regression, generalized estimation equations, ud Rosner's beta-binomial as applied to the ectopic pregnucy data set. Both levels of statistical signifiC&llce ud estimated odds ratios appeared to vary according to the estimation procedure. Fitting the studard logistic regression model to all available pregnucies without consideration of the correlation between the outcomes of the pregnancies within a woman produced unbiased estimates of effects but underestimated SE's ud overstated statistical signifiC&llce of covariates slightly. Standard errors, obtained from the GEE approach of Liug ud Zeger, were up to 14% higher thu the unadjusted SE's obtained from the studard logistic regression model. The estimated regression coefficients and their SE's that were obtained from Rosner's model were uniformly smaller thu the corresponding estimates that were obtained from the estimating equations approach with the exception of the NSG variable. However, the levels of statistical significance were in generally good agreement between these two models. The fact that estimated logistic regression coefficents and their associated SE's from Rosners model were closer to the null thu the estimates from the GEE approach is attributable to the underlying difference in the models with regard to the inclusion of outcome for previous pregnucy as u additional predictor variable. The GEE model which is a marginal model, excludes the status of previous pregnancies as predictors, whereas Rosner's model (which is a conditional model) includes the status of previous pregnucies as covanates. Thus, if the goal of the study is to estimate marginal effects of a covariate (e.g. effect of PID on ectopic pregnucy irrespective of the status of previous pregnancy outcomes), then the results of the estimating equation approach are more appropriate. On the other hud if the focus is, say, individual patient counseling of a specific womu, then it is useful to Table 5.16 86 -- Estimated Logistic Regression Coefficient from Alternative Model Predicting Ectopic Pregnancy in 1497 Women Variable Standard Logistic Regression GEE Rosner's Model SCORE1 (SE) 1.4112 (0.3647) 1.4326 (0.365) 1.2852 (0.30058) SCORE2 (SE) 2.3918 (0.373) 2.4263 (0.3764) 2.08186 (0.30703) AGEGP (SE) 0.8910 (0.189) 0.8818 (0.1841) 0.82971 (0.18338) LAPIDI (SE) 0.6024 (0.219) 0.6085 (0.2223) 0.47757 (0.19926) NSGI (SE) 0.6097 (0.252) 0.3495 (0.2910) 0.79874 (0.21230) MYCO (SE) 1.6540 (0.540) 1.6295 (0.5443) 1.45227 (0.40534) SCIMY (SE) -1.3575 (0.604) -1.3375 (0.6097) -1.18991 (0.46238) SC2MY (SE) -1.3803 (0.629) -1.3900 (0.6341) -121646 (0.46238) 0.089 0.0496 2.2 Intraclass Correlation Pairv.ise odds ratio 87 condition on the outcome of previous pregnancies, since conditional models such 88 Rosner'. model provide estimates of risk in the population of women that have the same history specific woman under consideration for counselling. 88 the Furthermore, Rosner's approach estimates the joint distribution of outcomes and thus enables one to assess goodness of fit of specific models since the model specifies a likelihood (Equation 5-3) assuming the bet&binomial model is correct. 5.3.3.5 Comparison Qf Th Intercepts Qf The Models Let Pi denote the random variable for the probability that any of the ni pregnancies of woman i will result in ectopic pregnancy at baseline. This probability has subscript i, indicating that it is specific to woman i. Insufficient information precludes estimating Pi for each woman, but it is possible to estimate the distribution of the Pi across women. Some women have higher probabilities of ectopic pregnancy and their Pi is higher (closer to 1), whereas other women have lower probabilities of aborting and their Pi is smaller (closer to • 0). Assuming Pi has a beta distribution we have, Var(Pi) = P(1-P) (a+t+1) = P(1-P) r Thus the expected value for a woman's probability of ectopic pregnancy is P, and the variance among these probabilities can be high or low depending on the value of r. Here the parameter r plays the role 88 the intraclass correlation coefficient. A comparison of the intercept terms for the standard logistic regression and Rosner'. model show that they are the same (Table 5.17). For the standard logistic model there is only one parameter, BO' which is estimated to be -3.0078. This corresponds to a baseline probability of ectopic pregnancy of 0.0471. For Rosner's model, the beta distribution 88 parameters are estimated as a = 0.89963 and b = 18.25537. The baseline probability is thus estimated as 0.0522. The intraclass correlation coefficient is r = a+~+1 0.0496. The hypothesis that the pregnancies from the same woman are statistically independent of each other (B O : r = 0) is tested by comparing the difference in the -2 log (L) between the two models to a chi-squared distribution with one degree of freedom (Rosner, personal communication). In this data set, this hypothesis is rejected (log likelihood difference =9.8, p < 0.001), thus we conclude that pregnancies from the same woman are positively correlated , though this correlation is small. The low correlation among pregnancies from the same woman coupled with the low average number of pregnancies contributed by each woman (1.9) are perhaps the reasons why the correlation does not have a large effect on parameter estimation and testing. In this work we have not investigated the behavior of the log likelihood ratio test statistic when the true parameter is near or on the boundary of the hypothesis region. Feder (1968) has studied the behavior of -2 log A near the boundary of the hypothesis region; additional work is probably needed. • Table 5.17 Comparison of the Intercepts of the Standard Logistic Regression and Rosner's Model Standard Logistic Model --- .---- ._._-----------~---._--- Parameter estimate -- Rosner's Model _. - - ---- 80=- 3.cX>78 a= 0.89963 b= 18.25537 Filled probability of ectopic Pregnancy when all predictors = 0 _I_ = 0.0471 I +exp(3.0078) P= il = 0.0522 a+b r=_I_ = 0.0496 a+b+l 210g L 1254.5 1244.7 co \0 90 5.3.4 Alternative Approach An alternative strategy for binary dependent outcomes is Bonney's (1987) regressive Bonney'. approach expresses the likelihood of a set of binary logistic regression approach. dependent outcomes as a product of conditional probabilities each of which is assumed to be logistic. Thus, the dependent problem is transformed to independent univariate logistic regression as follows: Pr (Y I ~) =Pr (yl' Y2'· .. , Yn I ~ ) = Pr = (Yll~) Pr (Y21 Yl'~)··· Pr (Ynl Y 1, Y2,···, Yn-l'~) Q + "Yl Z·1 1 + "Y2 Z·2 1 + ... + "Y'1- 1 Z.. 1,1- 1 + {3 X.1 • where, Z.. = 2Y. - 1 1J J o if j <i ifj~i Then, the likelihood for the n dependent observations is: n Pr (Y I ~) = rr exp(O.Y.) 1 1 Where the product is over observed Y's. Thus, in the case of the Lund ectopic pregnancy data both Rosner's and Bonney's model structures have some similarities in the sense that both include the status of previous pregnancies as predictors, though the likelihood functions are different. As mentioned before, 91 the likelihood for Rosner's model is derived as an extension of the beta-binomial model, whereas, the likelihood for Bonney's model is simply the product of conditional probabilities each of which is assumed to be logistic. Secondly, Rosner's model assumes that Yl' Y2"'" y.~ 1 have equal and additive effect on Y., whereas Bonney's model can accommodate both 1 equal and unequal effect of YI' Y2,m, Yi- 1 on Vi' Thirdly, Rosner's model describes only one pattern of dependence, namely the exchangeable correlation structure of the Y's given the X's. On the other hand, Bonney's model can accommodate more specialized patterns of dependence including Markov dependence of order 1. Finally, Rosner's model provides direct estimates of pairwise odds--ratio whereas with Bonney's model estimates of pairwise oddsratios are complicated to obtain. In the case of clusters of size 2 or 3 formulas are given in Bonney (1987 formulas 2.8 and 2.9). CHAPTER 6 ANALYSIS OF MOUTHRINSES IN INHIBITING THE DEVELOPMENT OF SUPRAGINGIVAL DENTAL PLAQUE 6.1 Descriptive Analysis In this clinical trial subjects were generally healthy adult male and female volunteers, age 18-55, with pre-existing plaque and gingivitis but without periodontal disease. Prior to entry subjects were screened for a minimum of 20 sound natural teeth, a minimum plaque mean of 2.0 and a minimum gingivitis mean of 2.0. Subjects with gross oral pathology or on antibiotic, antibacterial or anti-inflammatory therapy were excluded from the study. Subjects were then randomized to two new mouthrinses (A & B) or a standard mouthrinse with double blinding. 6.1.1 ~fouthrinse Dispensation Subjects began their assigned regimen of rinsing vigorously with 15 ml of the assigned rinse for 30 seeonds, twice daily, morning and afternoon, for six months. All rinsings were under supervision of the representative of the Principal Investigator on week days. Subjects were also given 16 ounce bottles and graduated plastic cups and were instructed to follow their assignment at home over weekends and to maintain a diary. During the study, subjects followed their usual oral hygiene and dietary habits but were instructed to refrain from using commercial mouthrin.ses and to advise the investigators of the use of antibiotic or antiinflammatory drugs. 6.1.2 Examinations Subjecta were instructed to refrain from all oral hygiene and use of experimental products 93 on examination days until after the plaque collections and/or examinations, at baseline and at 3 and 6 months following the initiation of rinses. Separate case report forms were used at each examination interval. Gingivitis was scored at baseline, 3 months and 6 months by the Modified Gingival Index. Plaque was also scored at baseline, 3 months and 6 months by the Turesky modification of the Quigley-Hein Index (1970). A summary of the demographic characteristic of each treatment group is given in Table 6.1. Randomization appears to have produced treatment groups which are reasonably comparable with respect to the distribution of demographic characteristics such as smoking status, age, baseline plaque index and baseline gingivitis index. However, the sex distribution was different in that experimental monthrinse A had significantly more females than the standard mouthrinses or B. §.J.....a. Descriptive Analysis In the next sections we will study the effectiveness of the 3 mouthrinses in inhibiting the development of dental~. Table 6.2 provides summary statistics of the Modified Quigley-Hein Plaque index (PINDEX) at 3 months and 6 months. From Table 6.2 we see that for all subjects, the mean plaque index decreased over time and the variance decreased as mean decreased. However the coefficient of variation was constant. The sample correlation among the observations within individual was relatively high (correlation 0.6275). = From the attached plot of time since study enrollment versus mean plaque index (Figure 6.1), the experimental mouthrinses A and B appear to be more effective in inhibiting the development of dental plaque. There was a 17.8% reduction in mean plaque from the 3 month to 6 month period for experimental mouthrinse B and only a 4.1% reduction for experimental mouthrinse A. Table 6.1 Summary Statistics of Demographic O1aracteristics by Treatment Group Treatment Group Covariate Standard (n= 39) A (n=34) B (n=36) p. value Sex Female Male 15 (38.5) 24 (61.5) 24 (70.6) 10 (29.4) 19 (S2.8) 17 (47.2) 0.023 Smoking Status Nonsmoker Smoker 33 (84.6) 6 (15.4) 28 (82.4) 6 (17.6) 31 (86.1) S (13.9) 0.910 Age Median Mean Range 27 27.5 (23. 38) 26 29.1 (22.48) 2S 27.4 (23.47) 0.3843 Baseline Plaque Index Median Mean Range 2.5 2.56 (2.07.3.3) 2.51 2.57 (2.04, 3.57) 2.43 2.48 (2.02.3.13) 0.S364 Baseline Gingivitis Index Median Mean Range 2.15 2.22 (2.02. 2.81) 2.16 2.19 (1.98. 2.88) 2.13 2.16 (2.01. 2.54) 0.3797 \0 ~ .. ... .. • Table 6.2 Descriptive Statistics for Plaque Index Date 3-months 6-months Mean 1.464 1.346 Correlation 1 0.6275 0.6275 1 Variance 0.4856 0.2925 0.2925 0.4473 Coefficient of Variation 0.476 0.497 \0 VI 96 Q) , .....S-4t: I • I rn ..= ='0 • I ~ ::s • f I I 't-t 0 II Q) Q.. ...• .••. ~ ~ • L ~ c ..0 >< Q) "t:S .. .I t:r tIS ...... .. % .....t: Q) ::s ~ t: (\1 Q,) ::s. . .. .... I CO Q) '" S-4 ::sQO ..... ~ JC. ...I- .• ..•. .... .... ..... .. ..•. •... ... •... •... ..•. ..... ..... ..... ... . 0 ~ 0 0 97 Therefore mouthrinse B appears to be both effective (reduces plaque quickly) and is longlasting (continues to reduce plaque over the entire time of the clinical trial) whereas mouthrinse A seems to be effective but its long-lasting effect, ie., its ability to continue to reduce plaque after 3 months is not impressive. Figure 6.2 and 6.3 show the histogram of the plaque measurements at 3 and 6 months respectively. The histograms were skewed to the right. The Shapiro-WUk test of normality at the two time points showed w = 0.969 (p = .101) and w = .964 (p =0.043), which suggest that these data do not represent samples from normal distributions. To account for the skewness in the distribution of plaque measurements at 3 and 6 months, a gamma distribution was evaluated. Moreover when a gamma distribution was superimposed on the histogram it appeared to provide an adequate fit to the data. This observation is further confirmed by the ~? goodness of fit test (Snedecor-Cochran, 1980) which yielded a p-value of 0.4723 for the 3-month measurement and 0.1689 for the 6-month measurement. The null hypothesis tested is that plaque observations are randomly drawn from a gamma distribution. Maximum likelihood parameter estimates were obtained using PROC CAPABILITY in the SAS-QC System (1989). Gamma quantile-quantile plots are shown in Figures 6.4 and 6.5. In general the points on the plot fallon or near the line with slope equal to the scale parameter, although, three or four observations appear to be outliers. 98 CD ::s~ • 0 I ~ ~U! ~ CD • i 0 • 0 ..... , ..•• ..••• II ..... ::s ..c ..... ,I I L 0 ~ I I U .. •· 0 I I• x i • ..· ... ! J.. .. - ~! III I rn~ ·.... 1 Oz L I IS I 8 8IS I ~ .. •· .' 0 "1j CD ~ ~ ..... - · •· rz.. 0 . co •· • CD J.4 I •• •• .. • L.LU.C., ..• • 0 •• i •• .•• -•I .• ~ ~ . t-' ~ 0 0 I ~ ::sbO ..... •-• •• ...• .. •• J.. N . .. ..... .....IS IS tl.. . • L :I U Figure 6.3: Fitted Gamma Distribution for Dental Plaque MONTH-' 11-' _·-__ 0····. - .----. , ,- , ·°1 / r~ J e r II c e n t 10 I o I 1 -0.1 1-.£ --.-_-1_-,.-- , 0.1 0.' ..--' 1.0 . I .• ' ... I -1 1.1 I.' 3.0 'INDEX Curve: • ' ••• I1... t.-o 1".p.-3.3 Ic.l.-0 .• U \D \0 100 >< CU .. • + ~ .....s:: CU ;. + IS ..... a.. ..... IS • ., s:: CU ..•. Q .... .....,0 -••. .. . 0 ..... a.. • •• ~ ...... "1 ~ .....CU $ -. •• •ic s:: ... lSi :Sz a • • I •• I .....cu .... ., s:: •• • •u • ,; • w • •• ... ~ ~ ~ L ~ I . -• c ~ IS :s I a .. IS e e •• IS CJ .. . co ~ ....cu :staa .... rz.. . . . . . .. •... •. •. •. •. L ... •. za.,x •. • •. • • 101 .. "u • 't' d + ..... u :J 0' IS ~ C4 • ~ IS ~ s::u ..•. Q s.. ....0 -•. It ~ • •• ...oSa It 0 ~ C4 ~ ..... ~i a U ~ '. ...•• .... ~! •ic • • I •• I u .... .... ~ s::IS ~ •• L. ~ ~ ... c •I •• a IS e e • IS ~ . lO CD U s.. .... ..• ~ :J :J t:.a r&. • ...•• •u • 0 ...• •. It •. It •. • •. • •... L"ZDIIIX •... •. • •. • • 102 6.2 Model Fittmg 6.2.1 Analysis Using GEE Techniques Since characteristics of the plaque data appear to be distributed as gamma, a generalized linear model is fit with gamma error and log link. Log link has the physical interpretation that the covariates have multiplicative effects. Thus we used the GEE method to model the marginal dependence of E(Yit) on the independent variables. Table 6.3 shows estimated regression coefficients and z-statistics for a gamma error and log link model. A. in the case of the pregnancy data exchangeable and independent correlation structures are fitted to investigate the effect of specifying different correlation structures. From the magnitude of the regression coefficients in Table 6.3 and their associated robust confidence intervals it is clear that the variables AGE, SEX, and SMOKE are not important. The Wald chi-squared value for testing the hypothesis that the coefficients for AGE, SEX, and SMOKE are jointly equal to zero was 2.2405 with 3 degrees of freedom (p = 0.5240). Thus we fitted a parsimonious model without these variables. The results of this model are displayed in Table 6.4. Table 6.4 also shows the results of a grossly misspedfied exchangeable correlation of 0.9, rather than its estimated value of 0.4299. Our purpose here is to show the effect of grossly misspecifying the correlation on the regression estimates and their robust z-statistics. 103 Table 6.3 GEE Analysis of the Dental Clinical Trial with Log Link and Gamma Variance Correlation Structure Estimate S.E. (Naive) Intercept Independent Exchangeable -1.3078 -1.3114 .2718 .3220 .2921 .2915 -4.47 -4.50 Rinse A Independent Exchangeable -0.3547 -0.3582 0.0729 .0863 0.0896 .0892 -3.96 -4.01 Rinse B Independent Exchangeable -0.3672 -0.3677 .0699 .0829 .0700 .0696 -5.25 -5.28 Baseline Independent Exchangeable 0.6265 0.6282 0.869 .1029 0.0890 .0886 7.04 7.09 Age Independent Exchangeable 0.0074 0.0075 .0057 .0067 .0083 .0083 .89 .90 Sex Independent Exchangeable 0.0546 0.0541 .0648 .0769 .0741 .0738 .74 .73 Independent Exchangeable 0.0129 0.0153 .0817 .0968 .0980 .0973 .13 Independent Exchangeable -0.0812 -0.0823 .0571 .0436 .0439 0.0414 -1.85 -1.86 Covariate Smoke Time For Independent Correlation Structure: Scale parameter 0.1740 Mean square error 0.2916 c 1 1.36. c~ 2.3322 = = = = For Exchangable Correlation Structure: Scale parameter .1745 Mean square error 0.2916 Working correlation 0.4148 c 1 1.0355. c~ 1.2748 = = = = = S.E. (Robust) Z-Robust .16 104 For exchangeable correlation structure, the coefficient estimated for experimental mouthrin.se B was -0.3641 (95% C.I. (-0.5023,-0.2259». This means the plaque measurement for mouthrinse B subjects are expected to be as the plaque measurements of the standard mouthrinse multiplied by 0.6948 (=exp(-.3641». Similarly the coefficient estimate for experimental mouthrinse A was -.3244 (95% C.I.(-0.5043,-0.1445) , i.e. the plaque measurements for mouthrinse A subjects were .7230 times the plaque measurement for standard mouthrinse subjects. Furthermore, as time elapsed, plaque level is that of previous time multiplied by 0.9178 (=exp(-0.0858). The regression estimates and their robust z- statistics did not change by much between the three correlation structures. This is consistent with the theory of Liang and Zeger that z-statistics are robust to the mis-specification of the correlation structure provided we have the correct link function. 6.2.2 Selection Among the Correlation Structures Comparison of robust-z and naive-z statistics helps to suggest which correlation structures might be incorrect. Recall that naive z-statistics are calculated from the variance estimate of the regression coefficients by assuming that the 'working' correlation sturcture is indeed the true one. Thus we can expect relatively large differences between these two statistics if the working correlation is far from the true correlation structure. Table 6.5 shows the robust z-statistic and naive z-statistic and their relative difference for the three correlation structures. The grossly misspecified correlation structure showed more marked difference between robust-z and naive-z. The least relative change was observed for exchangeable correlation structure (a = 0.4299). For independent correlation structure, naive z-statistics show larger values (in absolute terms) than robust-z for the time-independent variables. This indicates that variance is underestimated for time-independent variables since correlated data are treated as independent and counted as separate information. Note that ignoring the correlation leads to incorrectly interpreting the data as weak evidence for the effect of time (z-naive = -1.47). 105 Using the technique of Rotnitzky and Jewell (Section 3.3) the (~1' ~2) values for the exchangeable, independent, and a grossly misspecified exchangeable correlation structures are (0.9824, 1.1755), (1.2379, 2.0479) and (1.7423, 7.2325) respectively. These results tend to indicate the independent and grossly misspecified exchangeable correlations are inappropriate, and 80 with it our choice of covariate adjusted correlation is the exchangeable correlation structure = 0.4299. 6.2.3 The Effect of Different Specification of Links For gamma error three links are commonly used. These are: 1. Log link: '1 = log (1') 2. Reciprocal link: '1 =b 3. Identity link: = '7 I' Log link has the interpretation that the systematic part of the model is multiplicative on the original scale, and hence additive on the log-scale. Identity link has the interpretation that the systematic part of the model has an additive effect. As pointed out in Section 3.1, reciprocal link is the canonical link for gamma distribution in which sufficiency of resulting statistics and convergence are guaranteed (McCullagh &. Neider). The reciprocal link has the interpretation as the rate of a process or rate of change which may not be appropdate in this case. Table 6.4 Covariate Intercept GEE Analysis or the Dental Clinical Trial with Gamma Variance and Log Link Correlation Structure Independent Exchangeable «=0.9 Rinse A Independent Exchangeable «=0.9 Rinse B Independent Exchangeable «=0.9 Baseline Independent Exchangeable «=0.9 Time Independent Exchangeable «=0.9 Estimates S.E. (Naive) S.E. (Robust) Z-Naive Z-Robust -0.9890 -0.9935 -0.9982 .2315 .2753 .3159 .2581 .2573 .2565 -4.27 -3.61 -3.16 -3.83 -3.86 -3.89 -0.3203 -0.3244 -0.3287 .0709 .0844 .09683 .0922 0.0918 0.0916 -4.52 -3.84 -3.39 -3.48 -3.53 -3.59 -0.3631 -0.3641 -0.3651 .0699 .0834 .0959 .0701 0.0705 .0706 -5.19 -4.37 -3.81 -5.12 -5.17 -5.21 0.6132 0.6156 0.6181 .0878 0.1049 0.1204 .0939 0.0937 .0934 6.99 5.87 5.13 6.53 6.57 6.62 -0.0850 -0.0858 -0.0869 .0577 0.0438 0.0184 0.0438 0.0440 0.0443 -1.47 -1.96 -4.71 -1.94 -1.95 -1.96 For Independent Correlation Structure Scale parameter = .1778 c1 = 1.2379, Cz = 2.0479 For Exchangeable Correlation Matrix Scale parameter = .1783 c. = 0.9824 Cz = 1.1755 « = .4299 For grossly misspecified Correlation Scale parameter = .1788 c. = 1.7423 Cz =7.2325 « =.9 ...... ~ Comparison of Robust and Naive Z Statistics (Gamma Error Log Link) Table 6.5 Exchangeable Independent Covariate Z-Robust Z-Naive Intercept -3.83 -4.27 Rinse A -3.48 Rinse B Baseline Time Mean Relative O1argc Z-Robust Z-Naivc I J% -3.86 -3.61 -4.52 30% -3.53 -5.12 -5.19 1% 6.53 6.99 -1.94 -1.47 Grossly MisspeciOed Relative O1argc Relative O1argc Z-Robust Z-Naive 6% -3.89 -3.16 19% -3.84 9% -3.59 -3.39 6% -5.17 -4.37 15% -5.21 -3.81 27% 7% 6.57 5.87 11% 6.62 5.13 22% 24% -1.95 -1.96 1% -1.96 -4.71 140 14.6 8.4% 42.8 .... o ....., 108 Table 6.6 shows results of a gamma error and identity link model. For the exchangeable . correlation structure the predicted mean plaque measurements for the standard mouthrinse, experimental mouthrinses A and B are 1.787 (= 0.3674 0.4490 + .8408 X 2.562), 1.343 ( =-0.3674 - + 0.8408 x 2.568), and 1.235 ( =-0.3674 - 0.4817 + 0.8408 X 2.479) respectively. The corresponding predicted mean plaque measurements for the gamma error, log link and exchangeable correlation structure are 1.793 (e(-0.9935 .9935 - 0.3244 + + .6156 x 2.567», 1.301 (=exp ( - .6156 x 2.568», and 1.183 (exp(-9935 - 0.3641 - .6156 x 2.479». The observed mean plaque measurements were 1.763, 1.288, and 1.144; thus the log-link model seems to predict slightly better than the identity link. Figures 6-6 and 6-7 show observed and predicted mean plaque values for each mouth-rinse group by time points. Comparison of the predicted mean values and observed mean values shows that the two link functions provide a good fit, with the log-link providing a slightly better fit at 6 months. Furthermore, the fluctuations in the robust-z values across the different correlation structures were more marked for the identity link than the log link but not by much. Inconsistency with respect to robust-z values across the different correlation structures contradicts the property of robustness against the misspecification of the correlation structure. This could suggest that the log-link may be the more appropriate link since consistency of the robust-z with respect to the misspecification of the correlation structure can be preserved only when the model is correct. " • Table 6.6 - Estimated Regression Coefficients and Robust-Z Statistics for Gamma Error and Identity Link Coefficient Estimates Z-Statistics Independent Exchargcablc Misspccilied Independent Exchargeable Misspecified Intercept -0.3691 -0.3674 -0.3319 -0.87 -0.87 -0.70 Rinse A -0.4385 -0.4490 -0.5061 -3.39 -3.51 -3.95 Rinse B -0.4735 -0.4817 -0.5464 -4.60 -4.70 -4.64 Baseline 0.8385 0.8408 0.8449 5.18 5.19 4.63 Time -0.1215 -0.1234 -0.1264 -1.99 -2.01 -2.04 a 0 0.4383 0.9 .... ~ 110 ~ • .....d ~ taO 0 ~ • f-4 0 """" rn Q) .....~IS > '0 Q) ... •.• • • c c ~ .....(J '0 Q) f-4 • . i z 0.. '0 d IS '0 • Q) :> f-4 Q) rn ..0 0 .. •• CO • CO Q) f-4 ::sbO x. ..... i- ••. ••. ..•. ••. ..•. ••. •... •... .... •... •... ..•. ... ~ ..... •... ..... •.. • ~ 111 J4 d • ....to.::I ..., .... ..., ~ d • CD ..... 't' .... 0 'toot In CD w Ii w c ....='IS • > c • 't' ...,CD ....'t' til CD .... ~ 't' d • IS 't' CD i:CD .. In ,.Q 0 z-..• •• CO CD .... ....='taO ~ .I % CJ ..x.. •. •. ... •. ... •. ..•. ..•. ... •... •.. ..•. ..... •... ..... •... ...i · • • • • • • ~ • CHAPTER 7 SENSITMTY AND SPECIFICITY FOR CORRELATED OUTCOMES In this chapter a general estimating equations approach is uaed to obtain estimates of sensitivity, specificity, predictive value positive (PVP), and predictive value negative (PVN) when the data cODSist of conelated binary outcomes. First order approximations to the variances of estimated sensitivity, specificity, PVP, and PVN are given. Data from a dental study are used to motivate and illustrate the methods. 7.1 Introduction The use of diagnostic testa for the detection and evaluation of various diseases has grown markedly in recent years. A survey of published material in this field was presented by Miller et a1. (1977). The performance of a new diagnostic test for predicting the presence or absence of a medical condition is customarily evaluated by estimating its sensitivity and specificity with respect to a traditionally used and accepted test which is regarded as a "gold standard" in making the diagnosis. In this context, sensitivity is the probability that the new test indicates that the condition is present when the gold standard does, also. And specificity is the probability that the new test indicates that the condition is absent when the gold standard indicates it is absent, also. The positive and negative predictive values of a screening test are the probabilities of the presence or absence of the disease given positive and negative results of screening test, respectively. Fleiss (1981) discusses the problem of estimating sensitivity and specificity when each observation may be regarded &8 being independent of all other observatioDS; Lachenbrucb (1988) discusses the situation where a new test is given to 113 each subject several times; and Hujoel, Moulton and Loesche (1990) discuss the use of the specialized correlated binary models of Bahadur (1961) and Kupper and Haseman (1978) for obtaining standard errors of sensitivity and specificity estimates when observations are correlated. . In this Chapter we use the generalized estimating equations (GEE) approach developed by Liang and Zeger described in Chapter 3 for estimating sensitivity, specificity, PVP and PVN and their robust standard errors from clustered observations. Research presented in this paper was motivated by a dental study. The purpoee of this study was to assess the efficacy of a new test for predicting progression to gingivitis. For each subject in the study, data was collected at each of 5 different tooth sites in the mouth. At each selected tooth site a small flat rectangular strip (measuring approximately 0.0625 inch x 0.25 inch) was inserted into the crevice between the gum and the tooth. These strips were specially designed for the collection and evaluation of elastase concentration in gingiv81 crevicular fluid of individual teeth. Since elevated concentrations of elastase are believed to be precursors of gingivitis (i.e. loa of attachment between the tooth and the gingival tissue), indications of high elastase • concentration from the strips provide a new test for predicting progression to gingivitis of a person. Progression to gingivitis was measured by the change in attachment loss between an initial baseline measurement and a second measurement 6 months later by X-ray. A change in loss of attachment of greater than 0.6 mm was taken to indicate progression and represents the" gold standard". In estimating the measures of test validity and their standard errors in the dental study, it is reasonable to assume that the observations for progression or not to gingivitis within a mouth are correlated; due to the man~er in which an individual cares for his/her teeth or genetic factors, one would expect that if some teeth have progressed to gingivitis other teeth within the same mouth are likely to progress also. Therefore, statistical methods that recognize and account for the correlation of observations within an individual are appropriate. 114 Zeger (1988) gives an overview of different statistical methods for analyzing correlated binary data. Donald and Donner (1990) present results of a simulation study comparing estimators of the common odds ratio when the data are correlated. In the next sections we provide first order approximations to the variances of sensitivity, specificity, PVP and PVN. 7.2 Sensitivity, Specificity, PVP and PVN in Clustered Data. This Section is concerned with estimating sensitivity, specificity, PVP and PVN by treating both the "gold standard" and diagnostic test results as bivariate random variables. In terms of the dental example the gold standard corresponcU to the determination of whether the selected site in a subject's mouth progressed to gingivitis and the new test corresponds to whether the test strip indicated an elevated elastase concentration in the gingival crevicular fluid of corresponding teeth. Therefore, sensitivity and specificity may bP. modeled directly. Let T ij = 1 if the "new test" is positive in the ith subject and jth tooth o otherwise G ij = 1 if the "gold standard" is positive in the ith subject and jth tooth o otherwise A GEE model with Bernoulli variance function and a logit link is given by: Prob [T..=l I G..=G] lJ = lJ exp(P O + PI G) 1 + exp (PO + PI G) Thus sensitivity of the new test is given by: Prob (T.. = I I G.. = 1] lJ and specificity lJ ~11 115 = 1 + ~ (PO) = 11'00 • Prob (Too = 0 I Goo = 0) IJ I .. IJ Here we further assume that, 1) Pr (Til = 1 I Gil = G) = Prob (Ti2 = 1 I G i2 = G)= .•. =Prob (TiS = 1 IG i5 =G) for G=O,l for all values of gl' g2' •.. , g5 and for all i=l, 2, •.•, n j j=1,2, •..,5. Now let ~ 11 = _ (exp(P O+P 1) exp(P O+P 1» - [1 + exp (PO + P )]2 ' [1 + exp (PO + P )]2 1 1 (811'11 811'11» 8P O '8P 1 811'00 ~OO = ( 8P O 8"'00 = 8P ) ( - exp (PO) / [exp (1 1 + pO)]2, 0) d an . Thus the first order Taylor series approximations to the variances of estimates of sensitivity ... and specificity that account for clustering are - where VR . . • ~11 V R ~11 T is the GEE sandwich variance estimator of PG given in Section 3.3. Similarly PVP and PVN can be estimated indirectly using Bayes' rule. First let ¢ denote the prevalence of gingivitis by the "gold standard" test ; also we assume E(G ij )=¢ for all i, j. Then, Prob (Too = 1 I Goo = 1) Prob (Goo = 1) PVP= Prob [G ij = 1 I Tij = 1] = IJ IJ IJ Prob (Too = 1) lJ 116 11 Similarly, PVN= Prob [Goo = 0 I Too = 0] = lJ lJ = Prob (Too = 0 I Goo = 0) Prob (Goo = 0) lJ P b{T _ 0) 1J ro ij- 11'00 (1- t/l) 11'00 ( 1- 9'» + 1I'01t/l = MOO' where, 11'11 = Prob(T..=IIGoo=I)= 1 exp (PO + PI) + exp (P 0 + P1) ?rOO = Prob(T..=OIG..=O)= I + exp1 (PO I + exp (PO lJ IJ lJ IJ ?rIO = Prob(Too=IIG..=O)= IJ IJ exp(P O) ) ) ., 117 = 0o(1-tP)lI'OO(1 - 11'00) + (l-tP) = 1I'00[ {1-tP) 11'00 (1 -11'00) 2 °0 I 01 = 11'11 tP + 11'10 (l-tP) 00 = 11'00 (l-tP) + tP 11'01(1-"'01)] where and + 11'01 tP First order approximations to the variances of estimates of PVP and PVN that account for clustering are: Var (M n ) Var (MOO) • • • - T Al VR Al - • • and - T A O VR A O where again VR is the GEE variance estimator of PG given in Section 3.3. Section 7.2 Sensitivity and Specificity Study of the Test Strips To assess the sensitivity and specificity of the test strips, 30 subjects were recruited for a prospective study. For each subject five tooth sites were selected and a test strip was applied to each selection site. After 8 minutes the test strips were removed from the sites and "read" by the dentist. Six months later patients returned to the clinic where it was determined whether the selected sites had progressed to periodontitis. When held under ultraviolet light, strips fluoresce along the length of the strip. The extent to which each strip fluoresces along its length depends upon the elastase concentration in the crevicular fluid. Standardised 118 locations along the length of each strip enabled the dentist to determine the integer score for the strip that could range between 0 indicating weak elastase concentration and 4 indicating strong concentration. Part of the aim of the study was to pick a strip score for use as a standard in practice by dentists to predict patients' chances of progressing to periodontitis within 6 months with acceptable sensitivity and specificity. Table 7.1 lists the estimated sensitivity and specificity of the strip device accounting for clustering of observations within subjects' mouths of the prospective study. Diagnostic performance of the strip was estimated at each of four cut points. Table 7.1 GEE study of sensitivity and specificity of the test strip at 8 minutes. Specificity Sensitivjty 9!! I!2i!ll L ~ Ql estimate !L ~ Q1i It ~ ~ estimate !L ~ ~ 1+ 0.89 0.96 1.00 0.12 0.19 0.26 2+ 0.61 0.77 0.93 0.58 0.67 0.76 3+ 0.16 0.40 0.64 0.72 0.80 0.88 4+ 0.04 0.22 0.39 0.91 0.95 0.99 Table 7.1 shows that as strip cutpoint increases, sensitivity decreases from 0.96 to 0.21 and that specificity increases from 0.19 to 0.95. The strip cut point that provides moderate diagnostic performance corresponds to a cutpoint of 2+. At this level the estimates of sensitivity and specificity are 0.77 and 0.68, respectively. Cluster adjusted 95% confidence 119 intervals are moderately broad. Lower confidence bounds of these intervals suggest a less than impressive diagnostic performance. A dentist at the Centers for Disease Control suggested that current knowledge in oral radiology indicates that measurement error in ascertaining extent of bone 1088 is large enough to prohibit accurate measurement in bone .. height within 1-2 mm even after standardization of radiographic methods. Further, the method permits one to observe the proximal sites of teeth, only. In this regard, the CDC dentist suggests that the results we have obtained portray a somewhat "optimistic" performance of the sensitivity and specificity of the new test. Indeed, upon close scrutiny of the radiologic data we found some bone height measurements taken at 6 months to be greater than the baseline measurements indicating measurement error. The estimated "working correlation" matrix is 1.00 0.10 0.10 0.10 0.10 0.10 1.00 0.10 0.10 0.10 0.10 0.10 1.00 0.10 0.10 0.10 0.10 0.10 1.00 0.10 0.10 0.10 0.10 0.10 1.00 indicating that the correlation between clustered observations within a mouth may be quite small. In this case we may regard observations within subjects as being nearly independent and our effective sample size as being more like 30 x 5 = 150 independent observations than 5 correlated observations taken from 30 subjects independently. A. a last and interesting point, the CDC dentist suggests that the low observed correlation may be more a function of untoward variation resulting from poor definition of periodontal disease in its appropriate quantification and warns that our result pertaining to the test may not be reproducible using more appropriate data. 120 7.4 SAMPLE SURVEY APPROACH An alternative strategy for obtaining estimates of sensitivity and specificity is to use a sample survey approach. In this approach sensitivity and specificity are defined in terms of • ratio estimates as follows: Let, Zl=l if both the 'new test' and the 'gold standard' indicate disease presence o otherwbe. Z2 = 1 if the"gold standard" indicates disease presence o otherwbe. Z3= 1 if both the "new test" and the "gold standard" indicate no disease o otherwise. Z4= 1 if the" gold standard" indicates no disease o otherwise. . Thus sensitivity of the new test is given by .. Prob (T=l I G=I) Prob(T=I, G=l) = Prob(G=I) and specificity I G-O) _ Prob(T=O, G=O) P ro b (T-O Prob(G=O) = E Z3 EZ4 Thus, this approach enables the lack of independence of multiple sites from the same subject to be taken into account. Ratio estimates of sensitivity and specificity and their associated standard errors were obtained with the PC-CARP software for survey data analysis (Fuller WA, Kennedey W, Schnen D et ai, 1986). At the optimal cutpoint of 2+, the estimates of sensitivity and specificity were 0.80 [95% C.I. (0.63,0.97)] and 0.67 [95% C.I. (0.58,0.76)] respectively. These estimates and their 8S808iated 95% confidence intervals are similar to the results obtained in the previous Section. CHAPTER 8 SUMMARY AND CONCLUSIONS 8.1 Summary In this dissertation, longitudinal data analysis in which outcome measurements are repeatedly taken is investigated. By incorporating the correlation structure among measurements within individuals, NeIder's generalized linear models approach is extended for the repeated measurements data using quasi-likelihood methods by Liang and Zeger. The Liang and Zeger method is appropriate when one has repeated non-normal data and the primary interest of research is marginal dependence of response on covariates. Hence this modeling approach is useful in analyzing repeated measurements data in the following sense: 1. '" A wide class of error structures can be specified. It can be applied to various types of response variables that have normal, Poisson, binomial, positive exponential, gamma, etc, distributions. 2. Any "working" correlation structure can be specified, and yet regression coefficient estimates are still consistent even when the correlation is misspecified. 3. It accommodates discrete and/or continuous covariates which may be time-dependent or time-independent. 4. It accommodates imbalances in the number of measurements across subjects. 5. Regression coefficient estimates from generalized estimating equations are consistent and asymptotically unbiased and normally distributed. In Chapter 2, some of the procedures for analyzing repeated measures data were reviewed• • These included the classical methods of analysis for continuous response variables and those based on parametric models assuming a multivariate normal error structure. The methods 122 for analyzing repeated measures data when the responae variable is categorical include the weighted least squares approach (Koch et al, 1977), the Wei and Stram (1988) aproaeh with time-dependent and/or time independent covariates, and the Stram, Wei, and Ware (1988) approach for analyzing repeated measures data of an ordinal categorical responae variable. In Chapter 3, we reviewed generalized linear models, quasi-likelihood, and Liang and Zeger's extension of generalized linear models and quasi-likelihood to repeated measures data. In Chapter 4, we provided a descriptive epidemiology o( ectopic pregnancy and discussed the proposed risk factors for ectopic pregnancy including pelvic int1ammatory disease. The hypothesis that pelvic inflammatory diaeaae is a risk factor (or ectopic pregnancy is baaed upon the conjecture that PID-causing organisms such as Neisseria gonnorrhea, Chlamydia trachomatis, uroplasmas and mycoplasmas ascend (rom the lower genital tract to the fallopian tubes causing tubal infection and eventually tubal occlusion thus preventing the implantation of a fertilized egg in the uterus. In Chapter 5, the generalized estimated equations approach of Chapter 3 is implemented to the ectopic pregnancy data from Lund, Sweden. From the Lund ectopic pregnancy data, we learned that pelvic inflammatory disease is the strongest predictor of subsequent • development of ectopic pregnancy and that there is a dose-response relationship between PIO severity and ectopic pregnancy. We also learned that the presence of Mycoplasma from lower or upper genital tract sites at index laparoscopy is also a strong predictor. of ectopic pregnancy. Other correlates of ectopic pregnancy include age at pregnancy and history of gynecological surgery. We further analyzed the ectopic pregnancy data using Rosner's conditional modeling approach which is a generalization of the beta binomial models for correlated data. The GEE modeling approach was contrasted to Rosner's conditional model, and recommendations for choosing between the two was dependent on the objective of the study. • In Chapter 6, we used the GEE approach to study the efl"ectivenes8 of three mouthrinses in inhibiting the development of dental plaque. From this study we learned that both experimental mouthrinses A and B are more effective than the standard mouthrinse. e. 123 In Chapter 7 we used GEE techniques to obtain estimates of sensitivity and specificity when the data consist of correlated binary outcomes. First order Taylor series approximations to the variances of estimated sensitivity and specificity were given as a function of the robust variance estimator given in Chapter 3. 8.2 Future Directions In this section shortcomings of the GEE analysis approach and suggestions for future research are outlined • 1. The small-sample properties of the GEE methods have not been studied extensively. Though regression estimates from GEE are consistent, unbiased and normally distributed asymptotically, it is important to assess the performance of GEE's for small number of clusters. A simulation study of the GEE approach for analyzing correlated gamma data indicated that parameters were quite biased and had poor coverage probability in small • samples (Paik, 1988). • 2. The effect of the quantity of, and mechanism for, missing data have not been investigated in depth. Furthermore, in contrast with parametric normal theory methods and the weighted least squares approach, which require only that the missing data mechanism be missing at random, the GEE methods require the stronger missing completely at random assumption (Laird, 1988). That is, whether an observation is missing can not 'depend on previous outcomes nor covariate values. 3. Though the Rotnitzky and Jewell approach is an important and first step, it is not a necessary and sufficient condition for choosing the correlation matrix, R. Furthermore estimates of c 1 and c 2 by themselves without some kind of variance estimates may not provide very useful information. Thus, sampling properties of the estimates ofc1 and c2 are • needed to quantify how far these estimates are from one• 4. Further studies that look at bias and coverage probabilities of our Taylor series variance estimates for sensitivity and specificity are warranted. 124 BIBLIOGRAPHY Adena, M.D. and Wilson, S.R. (1982). Generalized linear models in epidemiological research, Instat Foundation, Sydney, Australia. Anderson, D.A. and Aikin, M. (1985). Variance components models with binary response: interviewer variability. J. Royal Statistical Society. Series B 47, 203-210. Atrash, H.K., J.M. and Hogue, C.J.R. (1986). Ectopic pregnancy in the United States, 1970-1983. CDC Surveillance Summaries. ~ 35, 29-37. Bahadur, R.R. (1961). A representation of the joint distribution of responses to n dichotomous items in Soloman, H. (ed), Studies in item analysis and prediction. Stanford University Press, Stanford. Bende11, R.B. and Afifi, A.A. (1977). Comparison of stopping rules in forward regression. Journal of the American Statistical Association, 72, 46-53. Bengtsson, L.P.H. and Moawad, A.H. (1966). Re: Lippes Loop and myometrial activity. (Letter). Lancet 1, 46. Bera1 V. (1975). An epidemiological study of recent trends in ectopic pregnancy. Br. J. Obstet. Gynaecol. 82, 775-82. Bonney, G.E. (1987). Logistic regression for dependent binary observations. Biometrics 43, 951-973. Budnick, L.D. and Parker, J. (1982). Ectopic pregnancy in New York City. Am. J. Public Health 72, 580-4. Bradley, E.L. (1973). The equivalence of maximum likelihood and weighted least squares estimates in the exponential family. Journal of the American Statistical Association 68, 199-200. Bronson, R.A. (1977). Tubal pregnancy and infertility. Fertil Steril, 28, 221-8. Brooks, C.A., Clark, R.R., Hadgu A and Jones A. (1988). The Robustness of the logistic risk function. Communications in Statistics-Simulation and Computation, 17, 1-24. • Crowder, M.J. (1978). Beta-binomial Anova for proportions. Appl. Statist 27, 34-7. • Centers for Disease Control (1986). Ectopic Pregnancy - United States. 125 ~ • 35, 289-91. Centers for Disease Control (1987). Ectopic Pregnancy - United States . 39, 401-4. ~ Chaparo, M. Snehanshu, G., Nasheol, A. and Poliak, A. (1978). Laparoscopy for the confirmation and prognostic evaluation of pelvic inflammatory disease. Int J Gynecol Qbstet 15, 307-9. Chow, W-H, Daling, J.R., Cates, W. and Greenberg, R.S. (1987). Epidemiology of ectopic pregnancy. Epidemio1osic Reviews 9, 70-94. Chow, W-H, Da1ing, J.R. and Weiss, N.S. et al. (1986). IUD use and subsequent tubal ectopic pre~a~cy. Am J Public Health 76, 536-9. Connolly, M and Liang, K-Y. (1988). Conditional logistic regression models for correlated binary data. Biometrika 75, 501-506. Cox, D.R. and Snell, E.J.(1989).Analysis of Binary Data, 2nd edn., New York: Chapman and Hall. Day, N.E. and Kerridge, D.F. (1974). A general maximum likelihood discriminant function. Biomet=~ 23, 3i3-323. • Donner, A. and Donner A. (1990). A si~u1ation study of the analysis of sets of 2 x 2 contingency tables under cluster sawpling; estimation of a common odds ratio. JASA 85, 410; 537-5~3. • Farr, C. and findlay, R. (1928). Salpingitis: a detailed analysis based on the study of five hundred ~(rty·five cases - January 1914 to December 1927, inclusive. Sure. Gynecol. Obstet. 49, 647-62. Feder, P.J. (1968). On the distribution of the log likelihood ratio test statistics when the true par~eter is near the boundaries of the hypothesis regi~ns. The Annals~athematical Statistics. 39, . No.6,2044-55. Firth, D. (1987). On the efficiency of quasi-likehood estimation. Biometrika 74, 233-245. Fleiss, J.L. (1981). Statistical Methods for Rates and Proportions, 2nd edn. Wiley-Interscience, New York. Frau, L.M., MacKay, H.T. and Hughers, J.M. et al. (1986). Epidemiologic aspects of ectopic pregnancy. In: Langer A., Iffy, L., eds. Ectopic Pregnancy. Littlefield, MA: PSG Publishing. • Fuller, W.A., Kennedy, W., Schnell, D., Sullivan, G., Park, H.J. (1986). PC Carp. Ames, Iowa: Statistical Laboratory, Iowa State University. Glebatis, D.M. and Janerich, D.T. (1983). Ectopic pregnancies in Upstate New York. JAMA 249, 1730-5. 126 Glynn, R.J. and Rosner, B. (in press). Comparison of alternative regression models for paired binary data. Statistics in Medicine. Greenhouse, S.W. and Geisser, S. (1959). On methods in the analysis of profile data. Biometrics, 1969,23,357-81. Grizzle, J.F., Starmer, C.F, and Koch, G.G. (1969). Analysis of Categorical data by linear models. Biometrics 25, 489-504. • Hadgu, A. Westrom, L., Brooks C.A., Reynolds G.H., Thompson, S.E. (1986). Predicting acute pelvic inflammatory disease: A multivariate Analysis. Am J Obstet Gynecol 155, 954-60. Halperin, M., Blackwelder, W.C. Verter, J. (1971). Estimation of the multivariate logistic risk function: A Comparison of the discriminant function and maximum likelihood approaches. Journal of Chronic Disease 29, 125-158. Hedburg, E. and Speyz, S. (1953). Acute Salpingitis: views on prognosis and treatment. Acta Obstet Gynecol Scand 37, 131-54. Hockin, J.C. and Dessamine, A.G. (1984). Trends in ectopic pregnancy in Canada. Can Med Assoc J 131, 737-40. Hugoel, P.P., Moulton, L.H. and Loesche, W.J. (1990). "Estimation of sensitivity and specificity of site-specific diagnostic tests. Journal of Periodontal Research 25, 1~3-196. HuYnh, H. and Feldt, L.S. (19 'v). Conditions under which mean square ratios in repeated measurments design have exact F-distributions. Journal of the American Statistical Association, 65(332), 1582-89. • Jacobson, L. and Westrom, L. (1969). Objectivized diagnosis of acute pelvic inflammatory disease: diagnostic and prognost~c value of routine laparoscopy. Am J Obstet Gyneco1 lOS, 1088-98. Kirk, E.R. (1982). Experimental Desgn, 2nd edn. ,Brooks/Cole Publishing Company. K1einbaum, D.G., Kupper, L.L., and Morgenstern J (1982). Epidemiologic Research, Belmont, California: Lifetime Learning Publications K1ienbaum, D.G. (1973). A generalization of the growth curve model which allows missing data. Journal of Multivariate Analysis 3, 117-124. Koch, G.G. (1970). The use of nonrarametric methods in the statistical analysis of a complex split ~lJt experi~ent. Biometrics 26, 105-128. Koch, G.G., Landis, J.R., Free~an, J. L., Freeman, D.H. and Lehnen, R.G. (1977). A general methodology for the analysis of experiments with repeated measurement of categorical data. Biometrics 33, 133-158. • 127 Kupper, L.L. and Haseman, K.J. (1978). The use of a correlated binomial model for the analysis of certain toxicological experiments. Biometrics 34, 69. Lachenbruch, P.A. (1988). Multiple reading procedures: The performance of diagnostic tests. Statistics in Medicine 7, 549-557 . • Landis, Jr.R. and Koch G.G. (1979). The analysis of Categorical data in longitudinal studies of behavorial development; in J.R. Nesselroade and P.B. Baltes (eds). Lon&itudinal MethodololY in the Study of Behavior and Development, Academic Press, New York, Chapter 9, pp.233-26l. Laird, N.M. (1988). Missing data in longitudinal Studies. Statistics in Medicine 7, 303-315. Landis, J.R., Miller, M.E., Davis, C.S. and Koch, G.G. (1988). Some general methods for the analysis of categorical data in longitudinal studies. Statistics in Medicine 7, 109-137. Levin, A.A., Schoenbaum, S.C. and Stubblefield., P.G. et al. (1982). Ectopic pregnancy and prior induced abortion. Am J Public Health 72, 253-6. • Liang, K-Y and Zeger, S.L. (1986). Longitudinal data analysis using generalized linear models. Biometrika 73, 13-22 . Liang, K-Y and Zeger, S.L. (1991). Regression Analysis for Dependent Responses, Johns Hopkins University, Department of Biostatistics. Louis, T. (1988). General methods for analyzing repeated measures, Statistics in Medicine 7, 29-45. Mastroianni, L. and Rosseau, ~.H. (1965). Influence of the intrauterine coil on ovum transport and sperm distribution in the monkey. Am. J. Obstet. Gynecol 93, 416-20. McCullagh P, and NeIder, J.A. (1989). Generalized linear models. Chapman and Hall, London. McCullagh,. P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society, Series B 42, 109-142. , & Meirik, o. (1981). Ectopic pregnancy during 1961-78 in Uppsala County, Sweden. Acta Obsted Gynecol Scand 60, 545-8. Mickey, J. and Greenland, S. A study of the impact of Confounderselection criteria on effect estimation. American Journal of Epidemiology 129, 125-137. Miller, M.C., Westphal, M.C., Reigert, J.R. and Barner, C. (1977). Medical Diagnostic Models: A Sibliography (Microfilm). Ann Arbor, Michigan: University Microfilms International. 128 Mueller, B.A., Daling, J.R. and Yeiss, N.S., et al. (1987). Tubal pregnancy and the risk of subsequent infertility. Obstet Gynecol 69, 722-6. NeIder, J.A. and Yedderburn, R.Y.M. (1972). Generalized linear models. J. Roy. Statist. Soc. Ser. A. 135, 370-384. Newhaus, J.M, Ka1bfleish, J.D. and Hauck, Y.Y. (1991). A Comparison of cluster-specifi~ and population averaged approaches for analyzing correlated binary data. Inte~ational Statistical Review 159, 25-35. • Ory, H.Y. (1981). The Yomen's H~altt Study. Ectopic pregnancy and intrauterine coutraceptive de~'i~es: new perspectives. Obstet. Gyneco1 57, 137-44. Pagano. R. (1981). Ectopic pregnancy: a seven-year survey. Med J Aust 2, 586-8. Paik, M.C. (1988) Repeated measurement analysis for non-normal data in small samples. COmmunication in Statistics-Simulation and Computation 17, 1155-1171. Potthoff, R.F. and Roy, S.N. (l9'>4). A generalized multivariate analysis of variance model useful especially for growth curve problems. Biometrika 51(314), 313-326. Pritchard, J.A., MacDonald, P.C. and Grant, N.F. (1985). Yilliams Obstetrics 7th ed. Norwalk, CT: Appleton-Century-Crofts. Rantakyla, P., Y10stalo, P, and Jarvinen, P.A., et al. (1977). Ectopic pregnancies and the use of intrauterine device and low dose progestogen contraception. Acta Obstet Gynecol Scand 56, 61-2. Rao, J.N.K. and Scott, A.J. (1987). On simple adjustments to chi-squared tests with sample survey data. Ann. Statist. IS, 385-97. Rao, J.N.K. and Scott, A.J. (1981). The analysis of categorical data from complex sample surveys: chi-squared tests for goodness of fit and independence in two-way tables. J. Am. Statist. Assoc. 76, 221-30. Rosner, B. (1984). Multivariate Methods in Ophthalmology with application to other paired-data situations. Biometrics 40, 1025-1035. Rosner, B. (1989). Multivariate methods for clustered binary data with more than one level of nesting. JASA 84, 373-80. Rosner, B. and Milton, R.C. (1988). Significance Testing for Correlated Binary Outcome Data. Biometrics 44, 505-512. Rotnitzky, A. and Jewell, N (1990). Hypothesis testing of regression parameters in semiparametric generalized linear models for cluster correlated data. Biometrika 77, 485-97. • 129 Sacks, B.P. and Layde, P.M., rrubin, G.L. and Rochat, R.W. (1982). Reproductive mortality in t~e ry~ited States. ~ 247, 2789-92. SAS Institute, Inc. (1985). Institute, Inc. • ~A~'/IY~ ase~'s Guide. Cary, NC: SAS SAS Institute, rnc. (1989). SAS/QC Reference Version 6, First Ed., SAS Institute, Inc. Savalainen, E. and Saksela, E. (1978). Ecoptic pregnancy: relationship to the preceding contraception. Ann Chir Gynaecol 67, 198-202. Schneider, J.~ Berger, C.J. due to ectopic pregnancy: A 557-566. a~d Catt~l, re'r~~w ~f C. (1977). Maternal mortality 102 deaths. Obstet Gynecol, 49, Schoen, J.A. and Nowak, R.J. (197 u ). Re?~at ectopic year clinical survey. Obstet G~p-col 45, 542-6. ~regna~cy: A 16- Schwertman, N.C. (1978). A note on th~ Geiosser-Greenhouse correction for incomplete data split-plot enal~7sis. Journal of the American Statistical Association, 7:,3S3-3So . • Sherman, D., Langer, R. and Sado~sky, G., et al. (1982). Improved fertility following ectopic pregnancy. Fertil Steril 37, 497-502. Smith, P.J. and Heitjan, D. (1~93). Testing and adjusting for overdispersion in generalize~ l~nee~ oo~els. Journal of the Royal Statistical Society. Seri~~_, 36,lAl-191. Stanish, W.M., Gillings, D.B. ~nd Koch G.G. (1978). An application of multivariate ratio methods fc~ the nnaly~i9 of a longitudinal clinical trial with Missing data. ~iometric~, 3b ., 305-317. Stiratelli, R. Laird N and Ware, J. (1984). Random effects Models for serial observati~ns with binary ~ecpanses. Bio~etrics 40, 961-71. 0= Stram, D. 0., Wei, L. J. and ~~Le. :. S. (1988). Analysis repeated ordered categorical outcones l":th ,~~~i~ly missing observations and time-dependent covariates. Jq~~~~~ ~~~e ~erican Statisttca1 Association, a3, 531-637 , Timm, N.H. (1980). Mu1tivaria~~.an&lysis of variance of repeated measurements. in P.R. Krishnaiah (ed), Handbook of Statistics. Volume I: Analysis of Variance, North-Holland, New York. pp 41-87. Tosteson, A.N.A. and Begg, C.B. (1988). A general regression methodology for ROC curve estimation. Med~cal-Decisi~n Making, 8, 204-215. Truett J, Cornrie1d J. and Kan.~el, W (~.967). A multivariate analysis of 130 the risk of coronary heart Diseases, 20, 511-524. dise~se in Framingham. Journal of Chronic Turesky, S., Gilmore, N.D. and ~lickman, J. (1970). Reduced plaque formation by the chloromethyl analogue of Victamine. C.J. Periodontol 41, 41. Wald, A. (1943). Tests of statistical hypotheses concerning several parameters when the number of ~bservations is large. Trans. Amer. Math. ~, 54, 426-82. • Walker, S.H. and Duncan, D.B. (1967). Estimation of the probability of an event as a function of several independent variables. Biometrika 54, 1 and 2, 167-179. Ware, J. H. (1988). Linear models Statistics in Medicine, 7, 29-45. f~r analyzing repeated measures. Washington, A.E., Johnson, R.E. and Sanders, L.L. (1987). Chlamydia trachomatis infections in the United States. What are they costing us? JAMA. 259, 2070-2. Wedderburn, R.W. (1974). Quasi-likelihood functions, generalized linear models, and the Gauss-Newton ~ethod. Biometrika, 39, 439-447. • Wei. L.J. and Johnson, W.E. (1985) Combi~ing dependent tests with incomplete repeated measurements. Biometrika, 72, 359-364. Wei, L.J. and Lachin, J.M. ~lS~4). Two-sample asymptotically distribution-free tests for incomplete cultivariate observations. Journal of the American Statist~~al Association, 79, 653-661. Wei, L. J and Stram, D. O. (19,38~. Analyzing repeated measurements with possibly missing observations ~) modell:ng marginal distributions. Statistics in Medicine, 7, l39-~.48. Westrom, L. (1975). Effect of acute pelvic inflammaatory disease on fertility. Am. J. Obstet. Gynecol. 121, 707. Westrom, L. (1980). Incidence, ?revalence, and trends of acute pelvic inflammatory disease and its consp.qu~nces in industrialized countries. Am. J. Obstet. Gynecol. 138, 880-92. Westrom, L. Bengtsson, L.P.h. ~r.d M~rdh, P.A. (1981). Incidence, trends, and risks of ectopic pregna~cy ~~ a population of women. Br. Med. J. 282, 15-18. . Williams, D.A. (1982). Extra-bi~~mial variation in logistic linear models. Appl. Statist. 31, 1~4-e. Woolson, R.F. and Clarke, ~.~. (1984). A~alysis of categorical incomplete longitudinal cata. ~~l of the Royal Statistical Society. t 131 Series A, 147, 87-99. • World Health Organization (1985). Task Force on Intrauterine Devices for Fertility Regulation. A mu1tinatio~a1 case-control study of ectopic pregnancy. C1in Reprod Fertil ~, 131-43. Zeger, S.L. and Liang, K.Y., TAngitudina1 data analysis for discrete and continuous outcomes. Biomet~i~s 42. 121-130. Zeger, S.L. Liang, K.Y. and Albert, (1988). Models for longitudinal Data: A generalized estimating equation Approach. Biometrics 44, 10491060. Zeger, S.L. (1988). Commentary. Statistics in Medicine 7(1/2), 161-168 . • • ,