Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
An Indicator of Nonresponse Bias Derived from Call-back Analysis Paul P. Biemer RTI International and UNC Outline Ignorable vs. non-ignorable nonresponse Bias in the nonresponse adjusted estimator Call-back model for estimating non-ignorable nonresponse Application for estimating drug use prevalence Future research Estimation for Population Proportions Consider a SRS of size n Want to estimate some proportion, Let Let yi denote the observed dichotomous variable ri response indicator i E (ri | i ) response propensity ˆi estimator of response propensity Nonresponse Adjusted Estimator Estimator of is ˆ adj n 1 1 ri yi ˆi n i 1 which is unbiased if nonresponse is ignorable w.r.t. ˆi i.e., if the error in ˆi is uncorrelated with yi Bias in the Adjusted Estimator B(ˆ adj ) E (ˆ adj ) 1 n ˆ i i E yi n i 1 ˆ i Cov( yi , ei ) ˆi i where ei error in the estimate of i ˆi Cov( yi , ei ) 0 if nonresponse is ignorable Call-back Model Analysis Goal is to estimate i when nonresponse is non-ignorable Uses yi and call-back patterns to predict i ; note, yi are only observed for respondents. For example, suppose yi 1 for a alcohol user (positive) 0 for non-alcohol user (negative) Using data on call outcomes at each call-back for users and nonusers, we can estimate response propensity as a function of yi Then Cov( yi , ei ) 0 Call-outcomes by LOE for Alcohol Call-back Data for Alcohol Negative 7000 Positive Interviewed positives Refused Number of Cases 6000 Other NR 5000 4000 Interviewed negatives 3000 2000 1000 0 0 5 10 15 LOE 20 25 Call-outcomes by LOE for Marijuana Call-back Data for Marijuana Number of Cases Negative 9000 Positive 8000 Refused 7000 Other NR Interviewed negatives 6000 5000 4000 Interviewed positives 3000 2000 1000 0 0 5 10 15 LOE 20 25 Call-outcomes by LOE for Cocaine Call-back Data for Cocaine Negative 10000 Positive 9000 Refused Interviewed negatives Number of Cases 8000 Other NR 7000 6000 5000 Interviewed positives 4000 3000 2000 1000 0 0 5 10 15 LOE 20 25 Call-back Notation 1 = interview 2 = non-interview 3 = noncontact Call pattern 31111 => noncontact followed by interview Once interviewed, stays interviewed (absorbing state) Once non-interviewed, stays non-interviewed (absorbing state) Call-Back Data for LOE=5 Pattern Definition Users Nonusers Total 11111 Interviewed at call 1 n(1,1|1) n(1,1|2) n(1,1|+) 31111 Interviewed at call 2 n(2,1|1) n(2,1|2) n(2,1|+) 33111 Interviewed at call 3 n(3,1|1) n(3,1|2) n(3,1|+) 33311 Interviewed at call 4 n(4,1|1) n(4,1|2) n(4,1|+) 33331 Interviewed at call 5 n(5,1|1) n(5,1|2) n(5,1|+) 22222 Non-interviewed at call 1 n(1,2|+) 32222 Non-interviewed at call 2 n(2,2|+) 33222 Non-interviewed at call 3 n(3,2|+) 33322 Non-interviewed at call 4 n(4,2|+) 33332 Non-interviewed at call 5 n(5,2|+) 33333 Never contacted n(5,3|+) Simple Call-back Model for NI-NR LOE-5 Log-Likelihood log ‹ ( ) 5 n(l ,1| g )log l 1 g 5 l ,1| g Likelihood of interview after l calls n(l ,2 | )log( g l ,2| g ) n(5,3 | )log( g 5,3| g ) l 1 g Likelihood of non-interview after l calls g Likelihood of no contact after 5 calls Obtain parameter estimates by maximum likelihood Simple LOE-5 Model Parameters 1 (true prevalence) 1,l|g , l 1, 2,3, 4,5, g 1, 2 (interview probabilities) 2,l|g , l 1, 2,3, 4,5, g 1, 2 (non-interview probabilities) 11 parameters and 10 degrees of freedom Over-parameterized; requires constraints These constraints reduces parameters to 7: 1,2|g 1,3|g 1,4|g , g 1, 2 2,2|g 2,3|g 2,4|g , g 1, 2 Application – Drug Use Survey Compared estimates of alcohol, marijuana and cocaine past year use prevalence for unadjusted current (traditional) adjustment call-back model adjustment Current adjustment incorporates 13 grouping variables and their interactions including a number of state specific components Call-back model incorporated call-back data (for up to 15 call-backs) and the drug use variable of interest Estimated Response Propensities for Simple LOE-15 Model Positive % Negative % Overall % Alcohol Marijuana 53.4 96.9 94.9 58.6 62.8 62.8 Cocaine 95.4 62.0 62.8 Prevalence Estimates for Simple LOE-15 Model Unadjusted % Alcohol Marijuana Cocaine Current % Call-back % Bias Due to NI-NR 65.86 16.98 65.15 10.58 77.92 11.00 -12.77 -0.42 3.63 2.35 2.39 -0.04 Future Work Test feasibility of incorporating call-back data in the nonresponse adjustment process Enter # call-backs into the current logistic regression model (does not adjust for NI-NR) Apply the simple call-back model to the drug use data after traditional adjustment to provide second adjustment factor for NI-NR Use the simple call-back model to assess NI-NR bias following traditional adjustment approach