Download Pill, Patch or Shot? Subjective Expectations and Birth Control Choice"

Pill, Patch or Shot?
Subjective Expectations and Birth Control Choice
Adeline Delavandey
Abstract
When choosing a contraception method, women base their decisions on their subjective expectations
about the realizations of method-related outcomes. Examples of outcomes include getting pregnant and
contracting a sexually transmitted disease (STD). I combine innovative data on probabilistic expectations
with observed contraceptive choices to estimate a random utility model of birth control choice. The
availability of expectations data is essential to identify preferences from beliefs. E¤ectiveness, protection
against STDs and partner’s disapproval are found to be the most important factors in the decision
process. The elicited expectations and inferred preferences parameters are used to simulate the impact
of various policies.
Keywords: Subjective expectations, contraception, random utility model.
JEL classi…cation: J13, C42, C35.
1
Introduction
Contraception has signi…cant social and economic impacts. At the family and society level, demographers
and economists have long emphasized the quality-quantity trade-o¤: smaller families invest more in their
children’s human capital.1 Furthermore, women’s control over their fertility may a¤ect their marriage and
education decisions, their labor force participation and, as a consequence, their socioeconomic independence.
Manuscript received May 2006; revised December 2006.
am indebted to Tim Conley, Chris Taber, Luis Vasconcelos and especially Chuck Manski for extremely valuable sug-
yI
gestions. I am grateful to Ran Abramitzky, Gadi Barlevy, Raquel Bernal, Iliyan Georgiev, Jacob Klerman, Raphael Lalive,
Pedro Mira, Francesca Molinari, Arthur Van Soest, Madeline Zavodny, two anonymous referees and seminar participants at
Cornell University, University of Michigan, Universidade Nova de Lisboa, Northwestern University, University of Pittsburgh,
University of Pennsylvania, University of Bristol, RAND, the 2004 PAA Economic Demography Workshop, the ESPE 2005
Annual Conference and the Fundacion Ramon Areces conference “What can we learn from individual expectations of economic
variables?” for helpful comments and discussions. Financial support from Northwestern University Graduate Research Grant
and a grant from the Searle Fund is gratefully acknowledged.
1 See for example Becker and Tomes (1976).
Despite the importance of fertility control, unintended (i.e. unwanted or mistimed) pregnancy is a common
phenomenon in both developing and developed countries. In the United States, almost half of the pregnancies
are unintended, and approximately half of these end in an abortion (Henshaw, 1998). The choice of birth
control methods in‡uences the extent to which women experience an unintended pregnancy. Yet, little is
known about women’s motives to choose a particular birth control method. In this paper, I use innovative
data on expectations and preferences to estimate a random utility model of contraceptive choice, without
the need to impose assumptions on women’s expectations.
Women decide to use a contraception method in order to in‡uence the occurrence of certain outcomes that
enter their utility functions; for example, getting pregnant, contracting a sexually transmitted disease (STD)
or experiencing side e¤ects. Each birth control method implies a particular probability for the occurrence of
these outcomes, which may not be known to decision-makers. Thus, when choosing a contraception method,
women base their decision on their subjective beliefs about this probability.
Despite the central role economic analysis allocates to agents’expectations in their decision-making, few
studies are based on empirically founded knowledge about subjective expectations. Researchers interested in
predicting behavior often make non-veri…able assumptions on expectations (e.g., that expectations are equal
to average population outcomes) and employ choice data to reveal preferences. However, a basic di¢ culty
is that observed choices may be consistent with many combinations of expectations and preferences. To
illustrate this, suppose that only two birth control methods exist and that the …rst method yields a smaller
objective probability of getting pregnant, but is associated with side e¤ects. Observing that a woman
chooses the second method is consistent with two possibilities: she either cares more about side e¤ects than
pregnancy, or she values e¤ectiveness more but believes that the second method is more e¤ective. Choice
data alone do not enable one to discriminate between these two competing explanations.
A possibility for overcoming the identi…cation problem is to use additional data on expectations. To
this end, I designed and conducted a survey in which I elicit directly probabilistic expectations regarding
contraception methods from young sexually active women.2 I compare the elicited subjective probabilities
concerning pregnancy, side e¤ects and protection against STDs with statistics from medical studies, and …nd
that median answers are consistent with those statistics for a wide range of questions and methods. This
suggests that respondents are able to meaningfully answer questions eliciting their expectations in probabilistic form. However, median answers do not re‡ect the wide range of answers provided by respondents. The
women I interviewed, mostly college students, exhibit substantial heterogeneity in beliefs, which emphasizes
the need to be cautious when making assumptions on expectations. This heterogeneity can stem from the
2 The
term “probabilistic expectations” refers to the fact that respondents were asked a probability rather than an expected
behavior (e.g., “what do you think is the percent chance that you would become pregnant” rather than “do you expect to
become pregnant?”).
2
fact that women have misperceptions about the risk they face,3 or that they face di¤erent risk (e.g., because
some might have more facility to conceive than others) and have private information about it.
I employ the subjective expectations data and the actual contraception method chosen by the respondents to estimate a structural model of birth control choice. This approach contrasts with existing studies
which focus on the association between women’s demographic characteristics and contraceptive choice (see,
e.g., Tanfer et al., 1992; Bankole et al., 1999; Grady et al., 2002) without studying choice behavior.4 The
heterogeneity in subjective beliefs enables me to identify respondents’preferences for each outcome considered. The underlying reason for the heterogeneity in beliefs is not relevant to the preferences’ estimation.
The method rests on the assumption that women choose their birth control methods based on their beliefs.
I …nd that pregnancy, contracting an STD and partner’s disapproval are the most important factors in the
choice of a contraception method in my sample. To evaluate the relevance of collecting expectations data,
I compare my results to estimates that would be obtained under regular assumptions on expectations (e.g.,
that women’s beliefs about pregnancy rates are equal to actual pregnancy rates in the population) and …nd
that preferences parameters for getting pregnant would be quite di¤erent.
Delving into the process of decision-making is crucial for policy design. For example, unintended pregnancy may be prevalent because (i) some women misevaluate their risk;5 and/or (ii) their preferences are
such that the disutility associated with contraception use does not compensate for the disutility that accompanies an unintended pregnancy. The methodology developed in this paper is useful to assess whether
there exist widespread misconceptions about contraceptive methods in some subgroups of the population,
what the relative weight of each outcome is in the contraceptive decision-making process, and ultimately
what factors would lead at-risk women to adopt fail-safe methods. In the paper, I use the expectations data
and the inferred preferences parameters to simulate the impact of various policies on contraception use: cost
subsidy, information campaign and technological innovation.
Recently, economists have increasingly undertaken the task of asking probabilistic expectations to survey
respondents about important events, such as the chance of survival or losing one’s job (e.g., Hurd and
McGarry, 1995; Dominitz and Manski, 1996, 1997; Fischho¤ et al., 2000). However, while expectation
data are becoming available, few studies have until now employed them to conduct inference on behavior.6
3 Examples
of contraception myths are provided by Brown and Eisenberg (1995).
include Tanfer and Rosenbaum (1986), and Rosenzweig and Schultz (1985). Tanfer and Rosenbaum evaluate the
4 Exceptions
e¤ects of women’s rating on a three points scale about the characteristics of one method on the decision to use that particular
method, but do not estimate preferences parameters. Rosenzweig and Schultz estimate a couple-speci…c measure of fecundity
and …nd that couples with above-average fecundity use ine¤ective methods less frequently.
5 Walker (2002) uses pregnancy and fertility expectations from the National Longitudinal Survey of Youth (NLSY) 1997 to
measure teenage girls’ ability to assess their conception risks, and concludes that most young women can accurately predict
their risk of pregnancy; yet the ones living in poverty systematically underpredict it.
6 See Manski (2004) for an overview and discussion on the state of knowledge about expectations data.
3
Recent studies incorporating expectations into econometric models include Nyarko and Schotter (2002),
Lochner (2007), Hurd, Smith and Zissimopoulos (2004) and Erdem, Keane and Strebel (2005) to study
various decisions such as strategies in games, committing a crime, the timing of Social Security claiming and
computer purchase.7
Another innovative aspect of this paper is the use of stated preferences. In order to get a direct measure
of respondents’preferences toward getting pregnant, I elicited respondents’Willingness-To-Pay (WTP) for
a hypothetical 100% e¤ective birth control method. To assess the overall coherency of the two types of
subjective data, the elicited WTPs are compared with the monetary cost that generates the same utility loss
as getting pregnant, implied by the estimates of the model. Remarkably, the median elicited WTP is found
to be very close to this monetary value. Furthermore, I use the WTPs in a novel way by incorporating them
directly in the estimation to fully account for heterogeneity in preferences for getting pregnant.
The methodology developed in this paper is far from being limited to the study of fertility and could
be applied to other household production functions such as investment in health (Rosenzweig and Schultz,
1983) or in children’s education and cognitive achievement (Todd and Wolpin, 2003). For example, the
decision to smoke, drink alcohol or seek prenatal care during pregnancy depends importantly on subjective
parental beliefs about the health consequences of those decisions for the baby. Given available data, it is
extremely di¢ cult to identify the parameters of both the utility function and the household production
function.8 Existing studies have thus to rely on the assumption that there is no heterogeneity in beliefs
about home production technology. It is quite likely however that individuals di¤er in their beliefs about
those technologies, which would explain a (large?) part of the heterogeneity in choices and outcomes.
Eliciting expectations data on household technology such as health production and incorporating them into
the estimation of behavioral models would improve our ability to guide policy.
The paper is organized as follows. Section 2 develops a theoretical framework of women choosing a contraception method. Section 3 describes the data collection methodology and respondents’answers. Section
4 presents the choice model estimation results and Section 5 focuses on the policy experiments.
2
Theoretical Framework
In this Section, I describe how sexually active women who are not trying to become pregnant make their
choice of a contraception method. Moreover, I discuss the tensions between generality of the model and
tractability of the data collection.
7 Van
der Klaauw (2000) and Van der Klaauw and Wolpin (2005) also employ expectations data to improve the precision
of the estimates of their structural dynamic model of teacher career decisions and retirement behavior, respectively, but they
maintain the conventional assumption of rational expectations to identify their model.
8 Rosenzweig and Schultz (1983) stress this point for the context of health production function.
4
Woman i’s utility function Ui (e1 ; :::; en ; c) is a function of a vector of outcomes (e1 ; :::; en ) and the monetary cost c of the contraception method she uses: Examples of outcomes (e1 ; :::; en ) include getting pregnant,
contracting an STD, and partner’s disapproval.9 Both the outcomes (e1 ; :::; en ) and the cost c are subject to
uncertainty at the moment of woman i’s choice of a birth control method. The uncertainty about the cost
is introduced to re‡ect respondents’ potential lack of knowledge about the cost of some of the methods.10
Although each method m entails an objective probability for the realizations of (e1 ; :::; en ; c), each woman
i possesses beliefs Pim (e1 ; :::; en ; c), or subjective probabilities, about the occurrence of the method-related
outcomes associated with method m. Those beliefs may evolve through time as woman i receives additional
information.
Women are assumed to maximize their current subjective expected utility.11 Thus, letting Mi denote
i’s birth control method choice set (including “no method” as an option12 ), woman i solves the following
problem:
max
m2Mi
Z
Ui (e1 ; :::; en ; c)dPim (e1 ; :::; en ; c):
(2.1)
While (2.1)’s generality is attractive, it may be di¢ cult to elicit from respondents the joint probability
distribution Pim (e1 ; :::; en ; c): I thus make several assumptions to ease the elicitation of the entire distribution
Pim (e1 ; :::; en ; c) from relatively simple survey questions. Assuming that the utility function is additively
separable in outcome, linear in cost and identical for all women with observable characteristics zi ; up to a
n
P
random term "im not observable to the econometrician, one can write Ui (e1 ; :::; en ; c) =
uj (ej ; zi ) + azi c +
j=1
"im ; where uj (ej ; zi ) denotes the utility associated with outcome ej for a woman with characteristics zi ; and
azi is a constant. The terms f"im g can be thought of as random taste variations that are individual and
method speci…c. Therefore, (2.1) becomes:
max
m2Mi
n Z
X
uj (ej ; zi )dPim (ej ) + azi
j=1
Z
cdPim (c) + "im :
(2.2)
The maximization problem presented in (2.2) is equivalent to the maximization under budget constraint
of a quasilinear utility function depending on the outcomes (e1 ; :::; en ) and a numeraire composite good.13
9 Rather
than assuming a joint decision process between male and female partners, the model assumes for simplicity that
the woman makes the contraception decision alone, but cares about her partner’s approval of the method.
1 0 Respondents who know the cost have a subjective distribution, allocating probability one to this cost.
1 1 By assuming that agents maximize their current expected utility, I rule out that women may …nd optimal to use a method
in order to learn about it. Experimentation may be important but is beyond the scope of this paper.
1 2 Women always have the possibility not to rely on any method. But the focus is on sexually active women who do not intend
to become pregnant, hence no-method is not chosen in order to achieve a pregnancy.
n R
P
1 3 Consider the following maximization problem:
max
uj (ej ; zi )dPim (ej ) + bzi X + "im ; subject to the budget
m2Mi ;X j=1
R
constraint that cdPim (c) + px X = Y; where X denotes a composite good of price px ; and Y denotes income. Plugging the
budget constraint into the objective function, we get the maximization problem presented in (2.2) with azi =
5
bzi
px
:
The speci…cation implies that (as long as the methods’prices are less than income) the birth control method
chosen is independent of the optimal amount of the composite good.14
The additive separability of the utility function implies that only the marginal distributions of beliefs
enter the expected utility, while the linearity in cost implies that only the expected value of the cost of each
R
method m; Ei (cm ) = cdPim (c); matters, rather than the whole probability distribution Pim (c): Therefore,
the marginal e¤ect implied by the change in one of the outcomes is independent of the level of the other
outcomes, e.g., an increase in the probability of getting pregnant yields the same utility loss whether the
risk of side e¤ects is low or high.
Finally, I consider only outcomes that are binary, so that Pim (ej ) reduces to Pim (ej = 1): This is not
a strong assumption for some of the outcomes, such as getting pregnant or contracting an STD, which are
clearly binary but may be restrictive for others, such as weight change or interference with romance.15 One
could alternatively allow some of the outcomes to be continuous and ask respondents for example 3 points
in their subjective cumulative distribution of beliefs, which could then be used to …t a whole distribution.16
This implies, however, two additional questions per method for each continuous outcome. For a respondent
whose choice set contains 13 methods, two continuous outcomes would require 52 additional questions. To
reduce the respondent’s burden, all outcomes considered are restricted to being binary. A similar argument
would follow if one were to relax the assumption about linearity in cost.
The subjective expected utility (SEU) maximization assumption implies that the precision of beliefs (i.e.,
the underlying probability distribution of beliefs about the values that the probability Pim (ej = 1) can take
in [0; 1]) does not in‡uence the decision process. Departing from the SEU framework, one could introduce
the notion of ambiguity aversion (Ellsberg, 1961). Incorporating it into the current framework would require
data not only on the subjective probability of the binary event, but also on the dispersion of that belief which
would increase respondents’burden substantially. It would also require assumption on how the dispersion
enters in the utility function. I will therefore abstract from it in the empirical analysis.17
With this speci…cation, the probability that a woman with given characteristics zi ; choice set Mi and
1 4 Suppose
n R
P
j=1
that the choice set has two methods.
The decision-maker would choose method 1 over method 2 if
n R
R
R
P
uj (ej ; zi )dPi1 (ej ) + "1m
uj (ej ; zi )dPi2 (ej ) "2m > azi
cdPi1 (c)
cdPi2 (c) ; and would then allocate the
j=1
remaining income to good X:
1 5 Focusing on preferences over binary events also implies that risk-aversion cannot be de…ned in the standard fashion through
a concave utility function: a “demi-pregnancy” cannot be preferred over a non-pregnancy with probability 1/2 and a pregnancy
with probability 1/2.
1 6 See for example Dominitz and Manski (1997), who elicit 3 points in the subjective distributions of future earnings and …t
these points to an individual-speci…c log-normal distribution using a least-squares criterion.
1 7 Appendix B2 presents some robustness checks on the role of the precision of beliefs. See also Delavande (2006) for a new
method to elicit precision of beliefs.
6
subjective expectations fPim (ej ); Ei (cm )gj2f1;:::;ng chooses method m is as follows:18
m2Mi
Pr m zi ; fPim (ej ); Ei (cm )gj2f1;:::;ng ; Mi
m2Mi
!
0
B
B
B
B
B
= Pr B
B
B
B
B
@
"im
"im <
1
) C
C
C
uj (ej = b; zi )Pim (ej = b) + azi Ei (cm )
C
C
j=1 b2f0;1g
(
) C:
C
n
P
P
C
uj (ej = b; zi )Pim (ej = b) + azi Ei (cm ) C
C
j=1 b2f0;1g
A
8m 2 Mi ; m 6= m
(2.3)
n
P
(
P
Equation (2.3) serves as the basis for estimation. When looking at the results, one should keep in mind that
if any of the imposed assumptions is invalid (e.g., if women based their decisions on continuous outcomes
rather than binary ones), it would lead to misspeci…cation biases.
3
Data Collection and Description
This Section describes the data and the methodology employed to collect them. The participants in the
study are mostly college students from Northwestern University and the Truman College, a 2-year Chicago
city college. Young US women of traditional college age (18-24 years) are among the people at high risk of
pregnancy and STD. They experience the highest rate of unintended pregnancy and represent the largest
group to seek abortion (Alan Guttmacher Institute, 2000). Moreover, the Centers for Disease Control and
Prevention (1997) estimate that 89% of female college students have had sexual intercourse and 20% of
those who had intercourse in the preceding month report not using any contraception method during the
last intercourse.
Respondents were recruited through advertising ‡yers posted on the two campuses. Flyers explicitly
asked for sexually active women between 18 and 40 years old to participate in a one-hour interview about
birth control, in exchange for $15.19 Ads were also posted in supermarkets of a Chicago neighborhood,
Rogers Park, close to a public library where I was able to conduct the interviews. I personally conducted all
the interviews face-to-face from January to May 2003. Of the 100 interviews, 77 were done at Northwestern
University, 15 at Truman College and 8 at the public library.20
1 8 It
will be assumed that the random terms f"im g have a continuous distribution, hence ties between utility levels associated
with di¤erent methods occur with probability zero.
1 9 The ‡yers did not specify any pregnancy intention, but none of the respondents was trying to get pregnant at the time of
the interview (respondents who reported not using any method were asked the reasons for not using, and being pregnant or
currently trying to get pregnant were among the possible answers).
2 0 The di¤erence in the number of participants from the two colleges stems from the fact that students at Truman were
much less willing to participate, perhaps because most of them work part-time while attending college and spend little time on
campus. Most of the interviewees from Northwestern were undergraduate students, but some graduate and MBA students or
7
Table 1 shows that respondents are young (70% are less than 25 years old), mostly single, though most
of them have a regular sexual partner, educated (three-quarters have completed more than 14 years of
schooling) and a majority are still going to school. The most-represented race is white-non Hispanic, which
constitutes almost half of the sample. Three-quarters of the respondents belong to a religious group, and a
third of them attend religious services at least once a month.21
{Table 1 about here}
The questionnaire collects data that are relevant for the estimation of the random utility model: respondent’s birth control choice, respondent’s birth control choice set, subjective beliefs for the probabilities of
outcomes and the cost of the birth control methods in her choice set, as well as respondent’s WTPs for
certain properties of contraception methods.
3.1
Current and Past Contraceptive Choice
Respondents were asked their current and past contraceptive choices. The …rst column of Table 2 presents
the distribution of methods currently used by the respondents. More than three-quarters of the women
interviewed rely on the birth control pill, condom, or the combination of both. Interestingly, methods that
were recently made available on the US market are already in use in the sample.22 Three respondents do
not use any “modern” contraception method.23 Considering the young age of most respondents, it is not
surprising to have so few women relying on non-reversible methods such as sterilization. All respondents
who undertook the sterilization surgery, except one, were older than 30 years of age and already had at least
2 children.
Even if the sample is far from being representative of US women, the distribution of methods is similar
to that of women aged 20-24 from the 1995 National Survey of Family Growth (NSFG).24 Among the
contraceptors aged 20-24 interviewed in the NSFG, 52% relied on the pill, 26% on condom (women using
both condom and birth control pill were counted as pill users), 6% on injectable, 0.3% on IUD and 4% used
traditional methods.
In my sample, 58% of the respondents had used one or several other methods in the past. The second
column of Table 2 presents the distribution of ever-used methods. Less common barrier methods, such as
their spouses participated as well.
2 1 Respondents recruited on the street are quite di¤erent from the rest of the sample; most of them are older, black and did
not graduate from high-school.
2 2 Five women rely on the injectable Depo-Provera (the “Shot”), which has been available since 1992, two on the weekly
birth control patch Ortho-Evra available since September 2002, and one on the monthly vaginal ring Nuvaring, available since
October 2001.
2 3 Non-modern methods refer to traditional methods, such as rhythm, periodic abstinence or withdrawal.
2 4 See Abma et al. (1997).
8
the cervical cap, diaphragm or female condom, had been used previously by the interviewed women.
{Table 2 about here}
3.2
Choice Set
Most analyses of choice behavior lack data on choice sets. This survey elicits the individual speci…c choice set
Mi . Respondents were asked whether they were aware of the existence of several birth control methods.25 To
elicit the choice set, I follow the strategy of the Demographic and Health Surveys (DHS) and present a list of
methods to ensure better recall.26 Comparing an open-ended versus a probing format for several countries,
Curtis and Kneitzel (1996) found probing to be quite e¤ective at ensuring the recall of some methods, such
as female sterilization. A list of 17 methods was provided, including two combinations (birth control pill +
condom and spermicide + condom) because they are frequently used jointly. A respondent’s choice set is
composed of all the methods that she states being aware of.27 Respondents are knowledgeable about the
existence of various birth control methods (see column 3 of Table 2). On average, they know approximately
13 methods among the 17 proposed. All the respondents are aware of the existence of the birth control
pill and condom, and most of them know less-frequently-used methods such as the female condom, the
diaphragm and Depo-Provera. Few are familiar with new methods like Lunelle, a monthly injectable, or
Nuvaring. Nevertheless, many respondents had heard about the new birth control patch.
“No-method” is assumed to be part of every respondent’s choice set. In the terminology no-method,
I include traditional methods, such as rhythm, periodic abstinence or withdrawal, as an option to reduce
pregnancy risk in case none of the “modern”existing birth control methods is used. Even though ideally one
would like to di¤erentiate among traditional methods, as di¤erent pregnancy risks may accompany their use,
they are aggregated into one alternative in order to limit the size of the choice set, which is already large.
In the empirical analysis, I will conduct several robustness checks to assess the sensitivity of the results to
the choice set (see Appendix B1).
2 5 The
survey elicits the choice set of each respondent at the time her most recent contraception choice was made.
the DHS elicit the knowledge of contraception methods in two steps (i.e., …rst by using an open-ended format and
2 6 While
second by probing), as of 1997, methods elicited from steps 1 and 2 are recorded in the same way for high contraceptive
prevalence countries (DHS, 2001).
2 7 In the empirical analysis, the choice set is assumed to be exogenous. One concern might be that women who have more
disutility for getting pregnant search for information more actively and end up with a larger choice set. This may not be such
a problem for most respondents, as the women’s health center of Northwestern University encourages students to attend a
Reproductive Health Educator session before their …rst gynecological exam or new prescription of birth control. See also the
discussion in Appendix B1, which suggests that heterogeneity in preference towards pregnancy is not correlated with the size
of the choice set.
9
3.3
Subjective expectations about birth control methods
3.3.1
Elicitation Methodology
n
The theoretical model described in Section 2 highlights the central role of the subjective beliefs fPim (ej = 1)gj=1
in the decision-making process. Relying on existing surveys and birth control prescription information, I
identify a set of outcomes and properties associated with birth control methods that may a¤ect women’s
utility. In a global survey on birth control organized by Pharmacia & Upjohn, women aged 18-49 who
have used birth control or plan to use it have identi…ed the following as characteristics of their ideal birth
control method: easy to use (91%), safe (90%), e¤ective (89%), few or no side e¤ects (85%), and allows
the user to become pregnant soon after stopping use (52%). Emans et al. (1987) report that in a study of
US adolescents, 45% of the teens were very concerned about weight gain while using oral contraception. In
focus-group discussions from Mexico, e¤ectiveness appears as the most desirable property for a birth control
method, followed by duration of protection and return to fertility.28 The obstacles to method acceptance
most frequently cited include bleeding problems, partner’s objections, fear of not being able to become pregnant soon after stopping use, concerns of side e¤ects and discomfort with having to interrupt intimacy or
inserting a method.
For each birth control method in the respondent’s choice set, I thus elicit the subjective probability
Pim (ej = 1) of the realization of six binary outcomes, the subjective disapproval from partner and religion,
the expected monthly cost and subjective beliefs about mode of administration. The six outcomes are: e1 :
pregnancy; e2 : side e¤ects (nausea, vomiting, headaches, menstrual irregularities, vaginal infection); e3 :
weight gain / loss; e4 : contracting an STD; e5 : interference with romance / sex play; e6 : being able to
get pregnant within 12 months after discontinuing the method.29 e7 and e8 refer to partner and religion’s
disapproval, respectively.
The question format to elicit the subjective probabilities is based on the use of percentages. A short
introduction to percentages, similar to that of the Survey of Economic Expectation (SEE) (Dominitz and
Manski, 1997), was read and handed out to the respondents. The probability questions were then worded
as follows:
What do you think is the percent chance that you would get pregnant during the next twelve
months if you were using [...] during that period?
and What do you think is the percent chance that you would contract a sexually transmitted
disease during the next twelve months while using [...] during that period, taking into account
2 8 See
Garcia et al. (1997).
the model focuses on decision-makers who are not trying to get pregnant and maximize their current expected utility,
2 9 Since
desired future fertility level enters only through e6 . See Heckman and Willis (1976) for an analysis of birth control strategy to
determine household fertility.
10
your own life-style?
The questions inquiring about partner and religion’s disapproval for the use of a given method were also
worded in percentage form as follows:
What do you think is the percent chance that your husband / partner would disapprove
your use of [...]?
Regarding weight gain and side e¤ects, probabilistic questions were asked only if the method had never
been used. For current or past contraception methods, the user had already experienced side e¤ects and
weight change, so these outcomes were asked directly using a yes/no format to capture 0% and 100%,
respectively (i.e.: Do you experience nausea, vomiting, headache, menstrual irregularities or vaginal infection
due to the use of [...]? ). The expected cost was elicited using the following question:
How much do you think you would spend per month if you were to use [...], taking into
account health care provider visits if necessary?30
As mentioned earlier, no-method is assumed to be a possible choice for every respondent and may include
“natural”methods such as rhythm, withdrawal or periodic abstinence. A smaller set of questions was asked
for this option, since it is assumed that the probability of some of the outcomes, like experiencing weight
change or nausea while “using” no-method, is zero. More speci…cally, respondents were asked about the
chances of getting pregnant, contracting an STD, partner’s disapproval, interference with sex play/romance
and being able to get pregnant within 12 months when trying to have a baby, if they were not using any
modern birth control methods.
Respondents were also asked how they think each method in their choice set is administered.31 For
methods such as the pill or the condom, every respondent knew the correct mode of administration. However,
some respondents thought that the IUD is administered locally by self or that the diaphragm has to be
administered locally by a doctor.
3.3.2
Description of Subjective Expectations
Respondents appear to be comfortable with the percentage format of the questions, as attested by a zero nonresponse rate. There has been some concern that the open-ended response mode, requiring respondents to
provide their own numbers, may increase the rate of respondents using “50%”to express epistemic uncertainty
3 0 The
elicited expected cost incorporates potential switching costs associated with the need to see a doctor to obtain a
prescription. It however does not incorporate the costs from switching to an over-the-counter method (which could be, for
example, an extra trip to the drugstore). Those are likely to be relatively small.
3 1 The following modes of administration were presented: (1) no female administration, (2) oral administration, (3) local
administration by self, (4) injection and (5) local administration by doctor. The …rst group refers to male condom and nomethod; the second group to the birth control pill, birth control pill + condom, and the morning after pill; the third group to
the diaphragm, cervical cap, female condom, spermicide, spermicide + condom, the sponge, Nuvaring and the patch; the fourth
group to Lunelle and Depo-Provera, and the last group to the IUD, Norplant and sterilization.
11
(Bruine de Bruin et al., 2000). However, the interviewed women use extreme as well as intermediate values.
To illustrate this, Table 3 presents the distribution of subjective probability that a given method would
interfere with sex play and romance, pooling all methods and all respondents. The answer 50% was provided
by 6% of the respondents and was less often used than 0, 5 or 10%.32 Table 3 shows respondents’willingness
to use the scale from zero to 100, although percentages rounded to the nearest 5 are more frequent. 95%
of the answers are rounded to the nearest 5, which is similar to the SEE, the Health and Retirement Study
(HRS) and the NLSY97.33 The introductory text about percentages does not seem to create anchoring, as
the numbers presented in the introduction are not more frequently given by respondents.
{Table 3 about here}
It is of great interest to compare respondents’ subjective expectations with objective realities. First, a
close correspondence suggests that respondents are able to answer probability questions in a coherent manner
and are aware of the risk they face. Second, it is useful to evaluate assumptions commonly made on expectations. Typically, expectations are assumed to equal average outcomes within a population.34 In existing
analysis of sexual behavior and fertility, the probability of pregnancy or conception is usually estimated using
panel data as a function of observable characteristics such as education and age, and then implicitly assumed
to be known by the decision-makers (e.g., Heckman and Willis, 1976; Hotz and Miller,1993; Arcidiacono et
al., 2005). As a consequence, respondents sharing the same demographic characteristics are assumed to have
the same beliefs. Rather than considering all possible birth control methods, economic agents are usually
assumed to be choosing between a limited numbers of actions.35 Other studies have abstracted for tractability from the uncertainty associated with contraception use, i.e., agents choose between having a child or not
(e.g., Wolpin, 1984).
Regarding the chance of getting pregnant while using a particular contraception method, one can compare
respondents’ answers to existing statistics available from medical reports. Two types of failure rates are
available: the percentage of pregnancies during typical use and during perfect use during …rst year of use
(See Appendix A for details). Table 4 presents the distribution of respondents’ subjective probability of
3 2 As
a comparison, the proportion of respondents in the Health and Retirement Study (HRS) 1998 providing 50-50 for the
subjective probability of being alive at age 85 or having income keep up with in‡ation is 22 and 24% respectively; and the
proportion of respondents in the labor force from the SEE providing 50% for the subjective probabilities of having no health
insurance or being victim of burglary is 7 and 10% respectively.
3 3 For example in the 1998 HRS, 98% of the respondents provided a percentage rounded to the nearest 5 for the probability
of being alive at age 85. In the SEE, this proportion is 91% for the probability of having no health insurance. In the NLSY97,
this proportion is 97% and 93% for the chance of being in school next year and being a parent before age 20.
3 4 See Dominitz and Manski (1996) for examples of assumptions about beliefs on the return to schooling.
3 5 For example, Heckman and Willis (1976) analyze the probability of conception for women who contracept versus women
who do not. Hotz and Miller (1993) estimate the probability of getting pregnant while using no method, contracepting or after
sterilization. Arcidiacono et al. (2005) focus on no contraception, episode-speci…c contraception or scheduled contraception.
12
getting pregnant while using the birth control pill and stresses the heterogeneity in beliefs with answers
ranging from zero to 100%. The most common values are 1, 2 and 5%. More than two-thirds of the
respondents think that the percent chance they would become pregnant while on the pill is between 0.1%
and 5%, which are the rates for perfect use and typical use respectively, but several of them provide answers
such as 10%, 40% or 60%. One may wonder whether respondents who have ever used the method in the past
have di¤erent beliefs than non-users. Table 4 highlights the answers of the respondents who are currently
using the pill or have used it in the past, and there is no striking di¤erence.36
This heterogeneity in beliefs can stem from di¤erent factors. First, respondents may have misperceptions
about actual population outcomes. Second, respondents may have private information about themselves.
For example, two of the respondents who provide a very high probability for the birth control pill mentioned
spontaneously during the interview that they got pregnant while using this particular method. Moreover,
additional available statistics point out that failure rates decrease with length of use and vary according to
some demographic characteristics.37 Finally, the heterogeneity in beliefs may re‡ect heterogeneity in future
behavior. For example, some women may expect to have more sexual intercourse than others, or some may
expect to properly and consistently use a method while others might expect to be less careful.
In contrast to pregnancy, respondents provide answers relatively di¤erent from actual population outcomes for common side e¤ects associated with some known methods. Table 5 presents the subjective beliefs
about the chances of experiencing weight change due to the use of the injectable Depo-Provera. More than
65% of the respondents think that they have a probability less than 25% of gaining or losing weight. These
numbers contrast with the fact that two-thirds of the women in clinical trial experienced weight gain due to
this method, the average weight gain being 5.4 pounds in the …rst year.
{Tables 4 and 5 about here}
Appendix A describes respondents’beliefs in greater detail, parallels them with available statistics and
stresses the heterogeneity in expectations. The correspondence of most of the median answers with the
statistics found in medical studies emphasizes that respondents answer seriously and meaningfully when
queried about their beliefs in percentage format. They do not exhibit regular misperceptions about their
risk regarding pregnancy and side e¤ects for most methods, an exception being Depo-Provera.
3 6 For
both pill and condom, one cannot reject the hypothesis that the samples of beliefs for users and non-users are drawn
from populations with the same distribution at a 5% level (Mann-Whitney test).
3 7 Ranjit et al. (2001) provide evidence that failure rate decreases with length of use of a particular method and present
the following typical failure rates for the …rst year and second year of use, respectively, among women aged 20 to 24:
IUD/injectable/implant, 5.1 and 4.2; pill, 9 and 5.8; diaphragm, 21.2 and 16.8; condom, 17.8 and 11.9; spermicide, 37.2
and 21.5. Fu et al. (1999) …nd that single, younger and poorer women experience higher failure rates. Rosenzweig and Schultz
(1989) …nd that more educated women use methods more e¢ caciously.
13
4
Choice Model Estimation
The primary objective of the collection of expectations data is to improve inference on choice behavior. Existing studies are limited to investigating the relationship between observable characteristics and contraceptive
choice (e.g., Tanfer et al., 1992; Bankole et al., 1999).38 However, they do not identify the structural relationship between demographic characteristics and choice behavior. What is it about more educated women
that makes them more likely to use the pill? Schooling may modify fertility and contraception choice because
it induces a change in preferences (Easterlin, Pollak and Wachter, 1980) or lowers the cost of acquiring information about contraception (Rosenzweig and Seiver, 1982). We can better understand contraceptive choice
if we model it as women’s decision to in‡uence the occurrence of important outcomes. In this paper, the
availability of actual contraception choice and data on subjective probabilities for many outcomes is essential
to determine women’s preferences toward each outcome considered within a structural utility framework.
4.1
Estimation Results with Homogeneous Preferences
I initially assume that the utility function of the decision-makers depends solely on the eight outcomes de…ned
in Section 3.3.1 and monthly cost but not on demographic characteristics zi : Therefore, the expected utility
maximization presented in (2.2) becomes
max
m2Mi
8
X
uj (ej = 1)Pim (ej = 1) + uj (ej = 0) [1
Pim (ej = 1)] + aEi (cm ) + "im ;
j=1
which is equivalent to
max
m2Mi
where
8
X
uj Pim (ej = 1) + W + aEi (cm ) + "im ;
j=1
uj = uj (ej = 1)
uj (ej = 0) represents the di¤erence of utility levels between experiencing the
8
P
outcome j and not experiencing it, and W =
uj (ej = 0):
j=1
I further assume that the random terms f"im g are i.i.d and have a Type I extreme value distribution,
implying that the f"im
"im g have a standard logistic distribution (the i.i.d assumption is discussed in
Appendix B2). Thus, the probability that woman i chooses method m presented in (2.3) can be rewritten
as:
Pr m fPim (ej ); Ei (cm )gj2f1;:::;8g ; Mi
m2Mi
!
exp
!
uj Pim (ej = 1) + aEi (cm )
j=1
=
P
m2Mi
3 8 For
8
P
exp
8
P
j=1
!:
(4.1)
uj Pim (ej = 1) + aEi (cm )
example, Tanfer et al. (1992) …nd that family structure at age 15, education, work status and religious a¢ liation
in‡uence the choice of a particular method.
14
Under the parametric assumption for the random terms, the constant a and the di¤erence of utility levels
8
f uj gj=1 are identi…ed. The heterogeneity in beliefs is critical for the identi…cation of the parameters.39 I
employ the elicited expectations and the birth control choice of the respondents to estimate the structural
8
preferences parameters a and f uj gj=1 . The knowledge of these structural parameters is crucial to determine
the e¤ects of various policies on the use of contraception (see Section 5).40
Table 6 presents maximum likelihood estimates based on equation (4.1). The relative magnitude of the
8
di¤erence of utility levels f uj gj=1 re‡ects the importance of each outcome in women’s decisions. The
di¤erence of utility levels between getting pregnant and not getting pregnant is negative and the largest
in absolute value, showing that the outcome that matters the most in contraception choice is pregnancy.
The second most important outcome is contracting an STD, whose coe¢ cient is also negative and large in
absolute value. The third most important factor that a¤ects choices is partner’s disapproval, whose coe¢ cient
is approximately half of the e¤ectiveness one. Religion’s disapproval also has a negative coe¢ cient, of a
magnitude of one-…fth compared to partner’s disapproval. It is, however, not signi…cantly di¤erent from
zero. The di¤erence in utility levels for being able to become pregnant within one year of discontinuing the
method is positive, consistent with the fact that women are more likely to choose a method that allows them
to get pregnant soon after stopping its use. The probabilities that the method interferes with romance and
is associated with side e¤ects both have negative coe¢ cients, of similar magnitude as the coe¢ cient attached
to religion’s disapproval. The coe¢ cients attached to the probabilities of weight change is positive, which
may be surprising. One explanation could be that treating weight change as a binary outcome is a strong
assumption and that it may be more informative to distinguish between gaining 1 or 10 pounds. The cost
coe¢ cient is negative and a $10 increase per month of the method price a¤ects the utility as much as a
5% increase in the probability of getting pregnant, which suggests that price is relatively important. The
relative ranking of outcomes implied by the estimates is similar to the …ndings of Grady et al. (1999), who
rank methods’characteristics according to the proportion of women who rate it as “very important.” They
obtain the following ordering: e¤ective in preventing pregnancy is the most important, followed by e¤ective
in protective women from STDs, no health risk and does not interfere with sexual pleasure.
In order to gain insight into the magnitude of the estimated parameters, one can translate the di¤erences
of utility levels into the monetary amount that would make a woman indi¤erent between paying that amount
and experiencing the outcome considered. Let W T Pj (P1 ; P2 ) denote the monthly willingness to pay to
experience outcome j with probability P1 instead of probability P2 , other things being equal. The model
3 9 If
everybody has the same beliefs in equation (4.1), only the realizations of the " would lead di¤erent individuals to make
di¤erent choices.
4 0 Note that abstinence is not a possible choice, as I consider only sexually active women. The preferences parameters may be
di¤erent if one includes women who choose abstinence as a way to avoid pregnancy or STD. Including abstinence would require
including the utility from sex in the choice model.
15
described in Section 2 implies the following indi¤erence condition:
P1
uj (ej = 1) + (1
P1 )
uj (ej = 0) + a
W T Pj (P1 ; P2 ) = P2
uj (ej = 1) + (1
P2 )
uj (ej = 0): (4.2)
In particular, the W T P1 (0; 0:05) to reduce the chance of getting pregnant by 0.05 satis…es the following
indi¤erence condition:
W T P1 (0; 0:05) =
0:05
u1
a
:
(4.3)
Table 7 presents the monthly monetary value that a respondent would be willing to pay to reduce the
probability of getting pregnant by 0.05, as well as the willingness to pay to reduce the probability of all other
outcomes by 0.05, using the di¤erence in utility levels
uj and the constant a presented in Table 6. It shows
that respondents would be willing to pay about $10 per month to reduce the chance of getting pregnant by
0.05, $8 to reduce the chance of contracting an STD and around $5 to reduce the probability of partner’s
disapproval. To reduce the interference with romance or the occurrence of side e¤ects by 0.05, the women
interviewed would be willing to pay approximately $1 per month.
{Tables 6 and 7 about here}
In order to assess how changes in beliefs a¤ect women’s decisions, Table 8 presents the alteration of the
predicted probability of choosing a speci…c contraception method as the subjective beliefs for that method
are modi…ed, everything else equal. The table describes how the predicted probabilities of choosing the pill,
the condom and Depo-Provera, computed using the sample average answers for every method and outcome
and the estimates of Table 6, change when the expectation about the chance of experiencing outcome j with
that method increases by 0:05, keeping constant all other beliefs (including beliefs about the combination
pill+condom). The predicted probabilities are quite responsive to a small increase in the chance of getting
pregnant: adding 0.05 generates a substantial decrease, around 40%, of the predicted probabilities.
{Table 8 about here}
While it is hoped that the outcomes considered in Table 6 cover the most important factors for contraception choice, others might also be relevant. I investigate this possibility in Table 9. The …rst column of Table
9 presents maximum likelihood estimates based on equation (4.1) with the addition of a dummy for each of
the methods to capture all observed and unobserved characteristics that describe the method (normalizing
the dummy for no-method to zero), i.e. the expected utility associated with method m is assumed to be
8
P
given by
uj Pim (ej = 1) + W + aEi (cm ) + Dm + "im , where Dm denotes the coe¢ cient associated with
j=1
the method’s dummy. The obtained coe¢ cients are very similar to those of Table 6 (coe¢ cients associated
with the dummies not shown). Using a likelihood ratio test of size 0.1, one cannot reject the hypothesis
16
that all constant terms are equal to zero. Therefore, throughout the remainder of the paper I do not include
those dummies.41
Column 2 of Table 9 presents a speci…cation that includes dummies for the perceived mode of admin8
P
istration, i.e., the expected utility associated with method m is assumed to be given by
uj Pim (ej =
j=1
1) + W + aEi (cm )+ dmode
1i [modem ]+ "im ; where 1i [modem ] is a dummy variable for method m0 s perceived
mode of administration according to individual i. The coe¢ cient for oral administration has been restricted
to zero. The coe¢ cient of the dummy for no female administration is positive, while those for local administration by doctor or self and injection are negative and signi…cantly di¤erent from zero. The magnitude of
the coe¢ cients reveals the following ordering of preferences, from the most preferred to the least: no female
administration, oral administration, injection, local by self and local by doctor.
Other factors such as frequency of administration or doctor’s visit might also be important to the contraceptive decision. Column 3 investigates this possibility by adding to the speci…cation of Column 2 dummies
for frequency of administration (take everyday or every week, take every month to every three months, and
take less often than every three months) and whether the method requires a doctor’s prescription. Note
that administration before intercourse and local administration by self pertain to the same set of methods,
so we cannot di¤erentiate their e¤ect. The need for a prescription decreases utility, though the coe¢ cient
is not signi…cantly di¤erent from zero. The coe¢ cients on frequency of administration suggest that women
prefer methods that are administered frequently. Including all these attributes does not change the WTP
presented in Table 8 for pregnancy, but the WTP to avoid contracting an STD decreases slightly.42
{Tables 9 and 10 about here}
Mode of administration seems to be an important factor in the decision. Table 10 presents the monthly
willingness to pay for using the mode of administration X instead of oral administration, other things being
equal, implied from the coe¢ cients of Table 9, Column 3. Respondents would be willing to pay $30 per
month to use a method without female administration rather than an oral method, and would pay $10 to
use an oral method rather than an injection. Note that the WTP for local administration by self may re‡ect
both disutility for local administration as well as regular administration before intercourse.
4.2
Estimation Results with Heterogeneous Preferences
In the previous subsection, the preferences parameters are assumed to be identical for every individual.
However, women with di¤erent demographic characteristics may have distinct preferences. For example, age
4 1 Note
that given the small sample size, the standard errors of the 17 estimated constants are fairly large.
the coe¢ cients of Column 3, we obtain that respondents would be willing to pay about $9.2 per month to reduce
4 2 Using
the chance of getting pregnant by 0.05, $5.8 to reduce the chance of contracting an STD and $4 to reduce the probability of
partner’s disapproval.
17
may in‡uence the relative price for getting pregnant or the outcome e6, “being able to get pregnant within
one year after discontinuing the method,” as older women are more likely to intend childbearing in a near
future. This heterogeneity in preferences may bias the estimated parameters, and should be accounted for to
8
use the results for policy simulation.43 It is straightforward to allow the f uj gj=1 to vary by demographic
characteristics. One could either separate the sample into subsamples of respondents de…ned according to
8
demographic characteristics and estimate di¤erent f uj gj=1 for each subsample, or introduce a functional
8
form to specify how demographic characteristics a¤ect the parameters f uj gj=1 . For example, imposing the
linear assumption uj (ej = 1; zi )
uj (ej = 0; zi ) =
max
m2Mi
where Wzi =
8
P
8
X
j zi Pim (ej
j zi ,
the expected utility maximization problem becomes:
= 1) + Wzi + aEi (cm ) + "im ;
(4.4)
j=1
uj (ej = 0; zi ):
j=1
Due to the limited size of my sample, I allow heterogeneity in the estimation for two outcomes. Table
11 presents estimation results using di¤erent speci…cations of the utility function. All the speci…cations
include dummies for prescription and frequency and mode of administrations (coe¢ cients not shown).44
Speci…cation (1) focuses on the most important outcome, namely getting pregnant, using the speci…cation
(4.4). It allows preferences for pregnancy to depend on the following zi : a constant term, marital status, the
number of children and the square of the number of children. Becoming pregnant might be less costly for
married women because these (especially in the age group of the sample) may intend to become pregnant
soon, and do not bear the social cost that their single counterpart may bear from having a child out of
wedlock. The current number of children might also in‡uence preferences for getting pregnant, though in a
non-linear way. For example, women without children might not be ready to bring up a child and might have
higher disutility for getting pregnant than women with one child. Alternatively, women with three children
might have reached their desired number of children and might have higher disutility from an additional
pregnancy than women with one child. In Column 1 of Table 11, we see that aside from the constant term,
the coe¢ cients estimating preferences for pregnancy associated with the zi are not signi…cantly di¤erent from
zero. The small proportion of married women and women with children in the sample might explain the lack
of precision of these coe¢ cients.45 Looking at the magnitude of the coe¢ cients, we see a non-linear e¤ect
of the number of children on
4 3 While
u1 : women with two children are those with the least disutility of becoming
the model as speci…ed is static, the preferences parameters can be thought of as being the result of a dynamic
optimization. Thus, state variables in a dynamic model such as age or the number and age of existing children are expected to
in‡uence preferences parameters.
4 4 The coe¢ cients associated with the dummies are similar to those obtained in Table 9, and are excluded in order to conserve
space.
4 5 Fewer than 20% of the respondents are married, and only 10% have children.
18
pregnant.46
{Table 11 about here}
The rate of unintended pregnancy in the US varies across race, the highest rate being among Black
women (Henshaw, 1998). Heterogeneity in preferences might explain this gap. Speci…cation (2) allows for
heterogeneity in preferences regarding becoming pregnant by race (White versus non-White) by estimating
a di¤erent preference parameter
u1 for each subsample. The di¤erence in preferences between the two
subsamples is striking: White have more than twice as high di¤erence in utility levels between getting
pregnant and not than do non-White.47 While we cannot generalize this result to the whole population,
given the nature of the sample, this …nding suggests that looking into heterogeneity of preferences across
race is worth investigating in a larger sample.
Speci…cation (3) focuses on heterogeneity in preferences for delay of return to fertility. Women who intend
to get pregnant in a near future are likely to have strong preferences for methods that allow fast return
to fertility once the method is discontinued. I can evaluate how intention regarding childbearing a¤ects
preferences for delay of return to fertility because each respondent i was asked her subjective probability
preg;i
that she will try to become pregnant within three years. About half of the women provide an answer
of less than 2%, while the remaining half provide answers approximately uniformly distributed between 3 and
100%. Preferences for “being able to get pregnant within one year after discontinuing the method”may not
be linear in the intention of getting pregnant, so a power parametric form is introduced. More speci…cally,
speci…cation (3) presents the estimated parameters for the following maximization problem:
max
m2Mi
8 5
P
>
>
uj Pim (ej = 1) + (
>
>
< j=1
>
>
>
>
:
+ u7 Pim (e7 = 1) +
pm + dmode
b
preg;i + 1%)
u8 Pim (e8 = 1) + W + aEi (cm )
1i [modem ] + df req
where the variable b is a constant to be estimated,
9
g
>
u6 Pim (e6 = 1) >
>
>
=
preg;i
1[f reqm ] + "im
>
>
>
>
;
;
(4.5)
is i’s subjective probability that she will try to
become pregnant within three years, pm is the parameter associated to the dummy for whether method m
requires a prescription and f reqm denotes method m’s frequency of administration. Results from Table 11,
column 3, highlight that the percent chance of getting pregnant within one year of discontinuing the method
is important to women who intend to get pregnant in the next three years. With the new speci…cation,
4 6 With
this speci…cation, the preference parameter
u1 is given by
c o n sta nt
+
married 1i [married]
+
children childreni
+
2
0
children2 (childreni ) ; where 1i [married] equals 1 if respondent i is married and zero otherwise, and childreni denotes i s
number of children. For married respondent with one child,
u1 =
8:5: For married respondent with two children
u1 =
7:5
while for married respondent with three children u1 = 8:9:
4 7 Using a likelihood-ratio test, I can reject the hypothesis that the two coe¢ cients for White and non-White are equal at a
5% level.
19
the di¤erence in utility levels for being able to get pregnant within one year after discontinuing the method
versus not is given by (
preg;i
+ 1%)b
g
u6 : For women whose belief that they will try to get pregnant
within the next three years is above 65%, delay of return to fertility is a more important outcome than
contracting an STD (65% + 1%)b
g
u6 = 6:42 : However, women who do not plan to get pregnant soon
do not attach much importance to a possible delay to fertility: the preference for this outcome is of the same
magnitude as interference with romance for women who have a 15% probability of trying to get pregnant in
the next three years (15% + 1%)b
4.3
u6 = 1:79 .
Estimation with Elicited WTPs
In the survey, a set of questions directly elicits respondents’ preferences by asking WTP for birth control
method’s e¤ectiveness using counterfactual scenarios. Stated preference (SP) methods are useful to conduct
inference on behavior.48 WTPs have often been elicited in studies that attempt to measure valuation for
environmental resources, such as air-quality, oil spill or river cleaning (See Hanneman, 1994 and references
therein). One of the major criticisms of assessing environmental utility through contingent valuation is
the lack of a market to generate prices for these public goods, and the resulting di¢ culty for people to
give coherent WTP for such goods. However, as pointed out by McFadden (1998), it is plausible that the
more realistic a hypothetical market setting, the more likely that SP will look like real preferences. From
reviewing existing work, McFadden (1998) suggests that two conditions must be met for SP to be successful:
(i) the consumer must be fully informed about the attributes of the commodity and experienced in making
decisions; (ii) the exchange must parallel a real exchange su¢ ciently closely. These conditions are met in the
present context: respondents who have already made their contraceptive decision are asked to express their
preferences for product attributes, such as the e¤ectiveness of a birth control method in terms of a price.
4.3.1
Eliciting WTP with a Scenario
Collecting WTP is informative for several reasons. First, WTPs are useful in assessing preference heterogeneity among individuals and can be incorporated into the estimation procedure to fully account for this
heterogeneity, as will be done in this Section. Second, they can be used to conduct robustness checks about
the collected subjective data. Unlike the situation with many other survey variables whose accuracy can
potentially be veri…ed if matched with administrative data, there is no formal way to evaluate whether
respondents’stated subjective beliefs and WTPs re‡ect their actual expectations and preferences. Manski
(2004) describes several ways that researchers have used to evaluate the coherency of elicited expectations:
4 8 For
example, Swait, Louviere and Williams (1994) …nd that estimates of preferences based upon SP data are more robust
than estimates based upon revealed preference data, because the SP design is more e¢ cient in terms of the variation in the
characteristics of the alternatives in the choice set.
20
comparison of individual expectations with realizations, comparison of mean expectations with realizations,
and comparison of mean expectations with historical realizations. Appendix A is based on the last approach.
This Section presents a new evaluation method based on the use of additional subjective data. If stated
WTPs have the same magnitude as the computed WTP implied by the parameters estimated using elicited
expectations, it shows coherency and robustness in the way respondents answer subjective questions. This
use of WTP is comparable to the use of subjective expectations about future occupation by van der Klaauw
(2000).49 Along with the methods described by Manski (2004), the evaluation presented in this Section o¤ers
comforting evidence but does not provide a formal test about the validity of the elicited data.
Preference Heterogeneity
Respondents were asked their WTP for e¤ectiveness by using the following
scenario:
Suppose now that there are only two birth control methods available on the market: the 100%-M and the
85%-M. They both involve that you take a pill once per week. They are completely identical in every aspect
(side e¤ ect, return to fertility, protection against STD...) except regarding the e¤ ectiveness:
- 100%-M is 100% e¤ ective, i.e. you have 100% chance of not getting pregnant if you use 100%-M or
equivalently a 0% chance of getting pregnant
- 85%-M is 85% e¤ ective, i.e. you have 85% chance of not getting pregnant during the next twelve months
if you use 85%-M during that period or equivalently a 15% chance of getting pregnant during the next twelve
months.
While the second method is presented as 85% e¤ective and available for free, I asked respondents their
monthly WTP for the …rst one, which is 100% e¤ective. Table 12 shows the sample distribution of WTPs
and illustrates preferences heterogeneity among respondents. Every woman in the sample would be willing
to pay at least $10 per month for the fully e¤ective method. The two most common stated values are $30
and $50. More than half of the respondents provide a value less than $30. Eleven respondents provide a
high WTP by giving amounts greater than $100 per month. This large variance may re‡ect the fact that
respondents would react di¤erently if they were to get pregnant: some may choose to end their pregnancy
while some may decide to keep and raise the child. Obviously, these two alternatives imply di¤erent costs
for a woman.
{Table 12 about here}
Coherence between stated preference and probabilistic beliefs The elicited WTPs enable one to
determine how each agent values e¤ectiveness and can thus be compared with the preferences parameters
4 9 As
in the present paper, van der Klaauw (2000) …rst tests for coherency of the expectations data by comparing them to the
predicted value from a model (albeit one estimated without subjective data) and then incorporates the subjective expectations
to gain e¢ ciency when estimating the model.
21
estimated previously. Using equality (4.2), the model implies that the elicited WTPs are given by
W T P1 (0; 0:15) =
Using the estimates of
0:15
u1
a
:
u1 and a presented in the third column of Table 9, one …nds that
(4.6)
0:15
u1
a
= 27:9:
Hence, the estimation results from the previous Section imply a monthly WTP of $27:9 for a di¤erential of
15% in e¤ectiveness. $27:9 is strikingly close to the median stated WTP, $30, which is also the most common
answer.
Once can also compare the heterogeneity in stated WTP with the heterogeneity in preferences for pregnancy by race described in Table 11. Using the estimates from Table 11 and equation (4.6), the implied
WTP to reduce the chance of getting pregnant by 15% for White and non-White respondents is $58 and
$23, respectively, while the mean stated WTP is $52.6 and $37.7. The di¤erential in preferences from the
stated WTP is slightly smaller than the one implied by the computed WTP, but both emphasize that White
respondents have more disutility for getting pregnant than do non-White.
The fact that the preferences parameters implied from two types of totally unrelated subjective questions
are so close is quite remarkable and strongly suggests that respondents provided coherent and meaningful
answers when asked about their beliefs and their preferences. Note that here, it is implicitly assumed that the
heterogeneity in WTP for e¤ectiveness is driven solely by the heterogeneity in preferences for the pregnancy
outcome. Heterogeneity in a could also explain variation in stated WTPs, but this possibility is ignored for
simplicity in this application.50
4.3.2
Using WTPs to Account for Heterogeneity: Estimation Results
The elicited WTPs clearly suggest that heterogeneity in preferences regarding pregnancy is important.51
As argued earlier, this heterogeneity should be taken into account for policy simulation. While observable
variables such as demographic characteristics or pregnancy intention can be used to allow heterogeneity in
8
the parameters f uj gj=1 ; they may not capture it fully. Elicited WTPs, however, provide direct evidence on
heterogeneity of preferences across individuals (under the assumption of homogeneity in the cost parameters
a). The previous estimations stress that pregnancy outcome is crucial in the contraception decision, and
the elicited WTPs for e¤ectiveness can be used directly in the estimation to relax the assumption that the
di¤erence of utility level
5 0 Respondents
u1 between getting pregnant and not getting pregnant is identical for all women.
provided the bracket of their household income, but some of them gave answers based on their own income
while others gave answers based on their parents’income, making this information hard to use to allow heterogeneity in a: As an
experiment, I re-estimate (4.1) allowing two values for a depending on whether the respondent is a student from Northwestern.
The estimates of a for Northwestern and non-Northwestern respondents are the following, respectively: aN U =
anon N U = 0:034. The estimates of the other parameters are similar to those of Table 6.
5 1 And thus that the common preference assumption of (4.1) might be invalid.
22
0:060 and
Using (4.6), the preferences parameter
u1i of individual i is given by:
u1i =
aW T Pi
;
0:15
(4.7)
where W T Pi refers to i’s stated WTP for the 100% e¤ective method as described in Section 4.3.1. Thus,
one can plug (4.7) into (4.1) and estimate the following equation:
Pr m fPim (ej ); Ei (cm )gj2f1;:::;8g ; Mi ; W T Pi
m2Mi
!
exp
=
8
P
uj Pim (ej = 1) + a Ei (cm ) + Pim (e1 =
j=2
P
exp
m2Mi
8
P
j=2
T Pi
1) W0:15
T Pi
uj Pim (ej = 1) + a Ei (cm ) + Pim (e1 = 1) W0:15
(4.8)
Table 13 presents the maximum likelihood estimates of (??), including in addition, dummies for prescription and mode and frequency of administration (coe¢ cients not shown). These results yield estimates
fairly similar to the previous tables but show a higher relative importance of the outcomes “contracting an
STD” and “experiencing side e¤ects” than found without taking into account the heterogeneity in taste for
e¤ectiveness and modes of administration. The estimates of Table 13 imply that women would be willing to
pay about $30 per month to reduce the chance of contracting an STD by 9.5%.52 So, comparing with the
elicited WTPs presented in Table 12, one can conclude that half of the respondents would rather decrease
their chance of contracting an STD by 9.5% than decrease the chance of getting pregnant by 15%. This
estimation illustrates how stated WTPs can be used in a new way to account for unmeasured consumer
characteristics that determine preferences.
{Table 13 about here}
The elicited WTPs allow me to account for heterogeneity in preferences toward the pregnancy outcome,
but preferences toward other outcomes may be heterogeneous as well. Unobserved heterogeneity in preferences regarding STD may generate correlations between a given individual’s random terms for the methods
o¤ering more protection against STDs (condom, female condom and combinations of condom with another
method). Since I did not elicit WTP for this outcome, I use a random coe¢ cients (or mixed) multinomial
logit model (see for example McFadden and Train, 2000) to investigate this possibility.53 I now assume that
u4 , which measures di¤erences of utility levels for the outcome “contracting an STD,”is a random variable,
and present two di¤erent speci…cations for its distribution:
(i)
u4 is de…ned such that
u4 = u4 + v; where u4 and
are real parameters and v is a random variable
distributed according to a speci…ed parametric family G(:; ): With this new formulation, the probability
the estimates from Table 13, we have 0:095a u4 = 30:6:
principle, one could similarly allow heterogeneity in preferences for all the outcomes. However, considering the small
5 2 Using
5 3 In
sample size, I focus on one of the most important.
23
!
!:
that a woman with given choice set Mi and subjective beliefs fPim (ej ); Ei (cm )gj2f1;:::;ng chooses method m
m2Mi
is as follows:
Pr m fPim (ej ); Ei (cm )gj2f1;:::;8g ; Mi
m2Mi
where V (m) =
8
P
j=2;j6=4
(ii)
!
=
Z
exp ((u4 + v) Pim (e4 = 1) + V (m))
P
dG(v; );
exp ((u4 + v) Pim (e4 = 1) + V (m))
(4.9)
m2Mi
T Pi
uj Pim (ej = 1) + a Ei (cm ) + Pim (e1 = 1) W0:15
:
u4 is a discrete variable and is distributed as follows:
8
>
>
> u41 with probability
<
u4 =
:::
>
>
>
: u
with probability
4k
k
In practice, to ensure that the f i gi=1 2 [0; 1] and sum up to 1;
i
1
:
k
is de…ned as follows:
i
=
exp( i )
:
k
P
exp( j )
j=1
Table 14 presents the estimation results with the assumption that, on the one hand, v is normal with
mean zero and variance one and lognormal with parameters zero and one and, on the other hand,
u4 has
a discrete distribution which takes two values.54 The coe¢ cients of the probabilities other than contracting
an STD are similar for every speci…cation and are also comparable to the results in Table 13. However,
the three speci…cations have somewhat di¤erent implications concerning the heterogeneity for preferences
toward the STD outcome. All speci…cations put most of the probability mass between -3 and -20. However,
the values of
i
imply that P ( u4 =
13:743) = 0:69 and P ( u4 =
4:082) = 0:31 in the discrete case, and
hence allocate most of the mass point to a slightly higher number in absolute value compared with the other
speci…cations. This di¤erence underscores the usefulness of collecting data eliciting preferences directly, as
di¤erent assumptions on the distributions of preferences in the population may lead to di¤erent results in
terms of heterogeneity.
{Table 14 about here}
4.4
Elicited Expectations versus Assumptions on Expectations
Without data on expectations, a typical assumption would be that individual expectations match population
outcomes, e.g., the expectations for getting pregnant would be assumed to equal population failure rates
of contraception methods. As emphasized in Section 4.1, the heterogeneity in expectations is crucial for
the identi…cation of the preferences parameters. So, in order to be able to identify the parameters of the
model under this maintained assumption, one needs data on population outcomes disaggregated according
5 4 Considering
the small sample size, one cannot allow k to be too large.
24
to some observable characteristics (yielding di¤erent outcomes in each subgroup). For all outcomes outside
of pregnancy there are no such data available, so one cannot estimate women’s preferences for those characteristics without data on expectations. Fu et al. (1999), however, provide contraceptive failure rates that
vary according to age, marital status and poverty level (below or above 200% of the federal poverty level)
based on the NSFG. For example, the failure rate of the birth control pill for single women between 20 and
24 years old with income greater than 200% of the poverty level is 7.4%, while it is 24.3% for cohabitating
women in the same age group with income below 200% of the poverty level.
As an experiment, I illustrate the advantage of collecting data on expectations with respect to imposing
assumptions on expectations by re-estimating the same model but replacing the elicited expectations with
the failure rates obtained by Fu et al. (1999).55 I classify each respondent in one of the subgroups speci…ed
by Fu et al. and assume that her expectations are equal to the failure rates of her own subgroup. For some
subgroups, the failure rates of certain methods are unavailable, due to infrequent use. Moreover, Fu et al.
provide information for eight methods only. For all failure rates that are not available from Fu et al., I
employ the failure rates during typical use presented in Table 16 (see Appendix A). For outcomes other than
pregnancy, I use the elicited expectations.
Table 15 presents the estimation results using the failure rates instead of the elicited expectations regarding pregnancy. All the speci…cations include dummies for prescription, and mode and frequency of
administration. The …rst column deals with the case of homogeneous preferences and can be compared to
the third column of Table 9. The coe¢ cient that measures the di¤erence in utility levels between getting
pregnant and not getting pregnant is negative and signi…cantly di¤erent from zero but is now much smaller
in absolute value than in Table 9. The implied WTP for reducing the probability of getting pregnant by
5% is approximately $3 while it was $9.2 with the expectations data. With this new speci…cation, getting
pregnant is suggested to give much less disutility than contracting an STD, which contradicts the ordering
of preferences presented in Grady et al. (1999). The second column presents the estimates when preferences
for getting pregnant depend on race, and can then be compared to the second column of Table 11. Again,
assuming that expectations are equal to failure rates implies less disutility for getting pregnant than using
the expectations data.
{Table 15 about here}
The last column of Table 15 provides the results of a speci…cation including both the failure rates and the
elicited expectations for getting pregnant as independent variables. The motivation for this speci…cation is
5 5 This
speci…cation ignores the possibility of heterogeneity in health endowment or fecundity known to the decision-maker
but not to the researcher, as in Rosenzweig and Schultz (1983), which could bias estimates. Given that most of the women
in the sample do not have children, it may be reasonable to assume that they do not have private information on their own
fecundity.
25
to assess whether the expectations data provide additional information compared to the failure rates that are
relevant for predicting behavior. The results show that the coe¢ cient attached to the expectations data is
signi…cantly di¤erent from zero and negative while the coe¢ cient for the failure rates is now not signi…cant.
These results highlight that the elicited expectations have more predictive power than the failure rates
from Fu et al. (1999) to explain contraceptive choice. Several reasons could explain this …nding. First,
assuming that individuals’expectations match population outcome may be incorrect (the heterogeneity in
beliefs presented in Appendix A tends to support this hypothesis). Second, using population failure rates
may not be adequate considering the nature of my sample. Note however that a similar problem would
arise whenever conducting inference on the behavior of a subgroup of individuals (e.g., black teens), for
which average outcomes might not be available or cannot be computed due to small sample size in existing
surveys. Which population shall then be considered to evaluate average outcome? How recent do the data
have to be to re‡ect current beliefs? How can one deal with expectations about a new technology, such
as the birth control patch? These questions and the results of Table 15 illustrate the di¢ culty of making
reasonable assumptions on beliefs, and emphasize the relevance of collecting expectations data rather than
making those assumptions.
5
Policy Experiments
The methodology presented in this paper can be used by policy-makers interested in fertility control and
family planning. First, eliciting subjective expectations about contraception and estimating women’s preferences regarding the methods’properties is useful to determine the nature of the policies that would have an
impact on contraception decision. For example, if respondents are found to exhibit regular misconceptions
about some birth control methods, information campaigns could be an appropriate action. Furthermore, the
elicited beliefs and the estimated preferences parameters can be used to assess precisely by simulation what
the impact would be of possible policy interventions on the use of contraception.
As an illustration, I discuss below the impact of three types of policies on my sample: cost subsidy,
technological change and information campaign. I report for each experiment the percentiles of the predicted
probabilities of choosing a method before and after the considered policy. In the following examples, I employ
the estimates of Table 13 to compute the predicted probabilities, thereby accounting for heterogeneity in
preferences toward pregnancy outcome.56
One way of reducing unwanted pregnancy is to encourage the use of contraception, and a price subsidy
could be implemented to make all birth control methods free. One can assess the impact of such a policy
5 6 This
illustration thus presents the e¤ect of the policies for sexually active women. With observations including non-sexually
active women, one could either model …rst the decision to have sex, and then, conditional on having sex, the choice of a birth
control method; or one could model abstinence as another method in the choice set.
26
by looking at the predicted probabilities of choosing no-method, and quantify how much incentive such a
price subsidy would provide to the use of contraception. Figure 1 presents the distribution percentiles of the
predicted probabilities of choosing no-method before and after a price subsidy has been implemented. The
predicted probabilities of “No subsidy”are computed using the elicited expected costs for every method, while
the predicted probabilities of the “Cost subsidy”are based on a cost of zero for each contraception method.
The percentiles of the initial predicted probabilities of choosing no-method are already low, but subsidizing
contraception prices makes them 10% - 35% smaller. The median predicted probabilities decreases by 22%
with the cost subsidy.
Another way to encourage the use of contraception is to subsidize the cost of a given method. I examine
the incentive provided by subsidizing the birth control patch. Figure 2 presents the quantiles of the predicted
probabilities of choosing the patch before and after a price cut of this speci…c method. The initial predicted
probabilities are computed using the patch’s elicited expected costs, while the predicted probabilities of the
“Cost subsidy” are based on a patch’s price of zero. The cost subsidy leads to a substantial increase in
predicted use of the patch. The median predicted probability of using the patch increased by about 40%
when a price of zero has been assumed.
{Figures 1, 2, 3 and 4 about here}
One can think of other shocks, like a technological improvement, that could lead to the increase of the
use of certain methods. Pharmaceutical companies are now developing new vaginal gels which are both
spermicide and microbicide, and hence protect against pregnancy and STDs.57 These would give women
protection against STDs without a partner’s knowledge or cooperation. Figure 3 examines the e¤ect on the
use of spermicides before and after such a technological improvement. The initial predicted probabilities are
computed using the elicited subjective beliefs of contracting an STD. For the “After technological shock,”
I assume that agents believe the new contraceptive gel would be as e¤ective against STDs as condom, and
the subjective beliefs of contracting an STD while using spermicides are replaced by the elicited beliefs of
contracting an STD while using condom. This technological improvement would substantially encourage the
use of vaginal gels. It leads to an increase of the percentiles of predicted probabilities of using these gels of
between 20% and 150%.
Finally, Figure 4 shows the e¤ect of a hypothetical information campaign that would lead everybody to
believe that the failure rate of Depo-Provera is 0.3%. Such an information campaign is credible since the
typical and perfect use failure rates are both 0.3%. But approximately 40% of the respondents perceived
that their chance of becoming pregnant if they were using the Depo-Provera injection was 5% or higher (see
Table 17 in Appendix A). The initial predicted probabilities are thus computed using the elicited subjective
5 7 See
Schwartz and Gabelnick (2002) for details.
27
beliefs regarding the risk of pregnancy while using Depo-Provera, whereas the predicted probabilities for after
the “campaign” are computed by setting this belief equal to 0.3% for every woman. The change of beliefs
leads to a rise of the predicted probabilities percentiles of 3% - 46%. The increase would be much sharper
if one considered an information campaign on methods for which there is more misconception regarding
e¤ectiveness, such as the IUD.
6
Conclusion
In this paper, I present a new methodology to uncover women’s preferences about contraception based on the
collection of unique data on expectations and their use in structural model estimation. The heterogeneity
in expectations in my relatively homogenous sample is indicative that there exists tremendous heterogeneity
in beliefs in the population. I combine expectations data with actual choice to estimate a random utility
model of contraception decision. The results show that women in my sample mostly consider e¤ectiveness,
protection against STD and partner’s disapproval when choosing a contraception method. The availability of
expectations data is critical in conducting inferences under weaker assumptions than what is usually imposed
under revealed preferences analysis.
Accurate knowledge of expectations and preferences is indispensable for developing e¤ective policy interventions aimed at reducing unintended pregnancies and instances of STDs. While this paper has shown the
validity of the methodology, the size, homogeneity and selected nature of the sample prevent general policy
recommendation. This paper clearly calls for similar data collection of contraception-related expectations at
a larger scale to inform policy. The important role of the male partner’s approval suggests that it may be
useful from a policy viewpoint to elicit subjective expectations from a male population as well.
The role of subjective beliefs is far from being limited to contraception, and expectations data about
various domains can improve our understanding of other important behaviors. Some large household surveys,
such as the HRS and the NLSY, are now collecting individuals’ subjective expectations. Yet, it would be
worthwhile to expand current data collections in two directions. First, one might improve data quality when
eliciting expectations about more complicated outcomes (e.g., non-binary) by using visual and interactive
design in new survey modes, such as Internet interviewing. Second, while most of the current data collection
has been undertaken in developed countries, one would bene…t from collecting data in developing economies
where uncertainty about health and economic outcomes is likely to be substantial.
Appendix
A
Detailed Description of Subjective Expectations
Regarding pregnancy, one can compare respondents’answers to existing statistics about …rst year of use
available from medical reports. Two types of failure rates are available: the percentage of pregnancies during
28
typical use and during perfect use. Pregnancy rates during typical use re‡ect how e¤ective a method is for
the average person who does not always use it correctly or consistently, while perfect use refers to consistent
use according to a speci…ed set of rules, which for many methods requires use at every act. Table 16 presents
the estimates of pregnancy rates based on Hatcher et al. (1998), as well as the mean and median answers of
the survey respondents for every method. For spermicide, diaphragm, male condom and pill, the estimates
for typical use are based on the experience of women from the NSFG. In the NSFG, a woman is “using”
a contraceptive method if she considers herself to be using that method, so a typical use of condom could
include actually using a condom occasionally. The estimates for the other methods are based on clinical
investigations. Pregnancy rates for perfect use are based on a series of clinical trials measuring method
failure rather than user failure. For almost all methods, the median subjective probability lies between the
rates during perfect use and typical use, showing a good understanding of both the percentage wording of
the questions and pregnancy risk. The median answer overestimates slightly the pregnancy risk associated
with the IUD, Norplant and Depo-Provera, which are methods that leave little room for errors.
{Table 16 about here}
While Table 16 shows that median answers are comparable to actual failure rates, it does not emphasize
the wide range of answers provided by respondents. Table 17 presents the distribution of beliefs for various
methods. For a widespread method such as condom, half of the respondents express a subjective probability
greater than 3% but less than 15%, which are the failure rates for perfect and typical use, respectively. In
contrast, a quarter of them give an answer lower than the perfect use failure rate and the same proportion
provides an answer higher than the typical use failure rate. For less common methods, the stated subjective
beliefs for pregnancy are mixed as well, but not exceedingly di¤erent from available statistics for a large part
of the respondents. For example, more than 60% of the respondents think that the risk of getting pregnant
is less than 2% while relying on Depo-Provera, and almost 90% think it is less than 5%, while the rate for
both typical and perfect use is very low (0.3%). Respondents show some optimism about the diaphragm,
with a third of them giving a probability less than 6%, which is the perfect use rate.
{Table 17 about here}
One can make similar comparisons regarding the risk of experiencing side e¤ects. Table 17 presents the
subjective beliefs’ distributions for the IUD and female condom, and shows that a large proportion of the
distribution has beliefs close to population outcomes. Approximately half of the respondents think they
have a 10% to 30% chance of experiencing one of the presented side e¤ects if they were to use an IUD. As a
comparison, medical reports document that 10% to 15% of copper IUD users will have their IUD removed
because of abnormal bleeding, and that insertion of the IUD can also increase the risk of infections.58 For
5 8 See
chapter 21 of Hatcher et al. (1998).
29
female condom, 60% of the respondents perceive a risk of side e¤ects less than 3%, which is compatible with
population outcomes, as problems associated with female condom are objectively infrequent.59
Nevertheless, respondents provide answers quite di¤erent from actual population outcomes for common
side e¤ects associated with known methods. Regarding weight gain associated with the pill, respondents
seem to overestimate their risk, with more than half providing answers above 20% while empirical research
has failed to document weight gain as a side e¤ect of pill use.60 Moreover, after discontinuing Depo-Provera,
women have a 6 to 12 months delay in return of fertility, but the respondents do not show awareness of
such a long delay. A little under half of the respondents may be knowledgeable about the consequences on
fertility by providing a percentage less than 50%, but more than a third seems totally unaware of it, and
provide answers higher than 80% (Table 17).61
Most of the respondents consider themselves at low risk of contracting an STD. Although a quarter
of the respondents perceive a risk greater than 30% of contracting an STD during the next 12 months
while using no-method, Table 17 shows that almost half of them provide a zero probability.62 Among those
respondents, all except three state having a “regular”sexual partner. For outcomes that cannot be compared
with population outcomes, respondents exhibit an even greater heterogeneity in beliefs. For example, a third
of the respondents expect no disapproval from their partner for the use of female condom while the others
have beliefs distributed between 1 and 100% (Table 17).
B
B1
Robustness Checks
Choice set
It is interesting to evaluate whether the estimated preferences are sensitive to the self-reported choice
set. To this end, I conduct two experiments. First, I exclude from a respondent’s choice set the methods
that she reported with an incorrect mode of administration. Incorrect response about administration might
be an indication that the method is not really considered by the decision-maker. If one excludes those
methods, respondents know on average 12 methods (compared to 13 without the exclusion). Second, I
restrict individual i’s choice set to be composed of the methods known to all respondents (i.e., pill, condom,
pill + condom and no-method) and i0 s chosen method. In both cases, the coe¢ cients estimated and the
implied WTP are very similar to those using the elicited choice sets.
One may also wonder whether women who have more disutility for getting pregnant may end up with a
larger choice set. To investigate this, I test whether the elicited WTP is correlated to the size of the choice
set. Using a Pearson correlation test, I cannot reject the hypothesis that there is no correlation between the
5 9 Among
the 360 women from the e¢ cacy study, only one discontinued the method because of discomfort.
Rosenberg (1988).
6 1 As a comparison, the percent chance of getting pregnant within a year for women who want to get pregnant and did not
6 0 See
use contraception associated with delay to fertility is estimated to be 85% (Hatcher et al. 1998).
6 2 In Illinois, 6% of the women aged 20-24 who visited family planning clinics were infected with chlamydia in 2001.
30
elicited WTP and the size of the choice set at a 5% level.
The random terms f"im g
The assumption that the f"im g are independent for every i and every m deserves further investigation.
Heterogeneity in tastes for observed attributes may yield correlation between the random terms. Section
4.3.2 investigates this possibility by allowing for heterogeneity in preference toward pregnancy and STD.
Heterogeneity in preferences for unobserved methods’characteristics may also generate correlation. I estimate
the choice model using a nested logit speci…cation allowing correlation between two groups of methods (either
barrier versus non-barrier methods or condom versus non-condom methods). The results (not shown) reveal
that there is no correlation between the error terms of the barrier group, the non-barrier group and the
condoms group. The coe¢ cients of the other outcomes remain basically unchanged with respect to Table
13.
Another concern might be that the random terms f"im g are correlated with the subjective probabilities.
Suppose for example that all women start with the false perception that a relatively new method that is
easy to use has no side e¤ects. Women who greatly like easy-to-use methods will be more likely to try the
new method and update their beliefs about the side e¤ects. As a consequence, the preference heterogeneity
about ease of use captured by the errors "im will be correlated with the subjective beliefs associated with that
method.63 While I cannot test for this particular example, I can use the elicited WTPs to evaluate whether
preference for pregnancy is correlated with subjective beliefs, which would suggest that heterogeneity in
preference a¤ects beliefs through learning. Using a Pearson correlation test, I cannot reject the hypothesis
that there is no correlation between the elicited WTPs and the subjective beliefs (all methods, all outcomes)
at a 5% level.
Finally, if the precision of beliefs matters in the decision process (see discussion in Section 2) and enters
the utility function directly, it will be captured in the error terms of relatively unknown methods and might
thus create correlation between the f"im g of unknown methods. One might expect a priori that beliefs
about widespread methods, such as the pill and condom, and about the currently used method are precise.
The fact that, as pointed out in Appendix B1, the results obtained when only widespread methods are
included in the choice set are very similar to those using the elicited choice set shows that whether methods
for which respondents are likely to exhibit more uncertainty are considered in the choice set or not yields
similar estimates of preferences, which is indicative that there is no systematic bias coming from ignoring
the precision of beliefs.
Universidade Nova de Lisboa, Portugal and RAND Corporation, U.S.A
63 I
thank an anonymous referee for pointing this out.
31
References
[1] Abma, J., A. Chandra, W. Mosher, L. Peterson and L. Piccinino, “Fertility, Family Planning, and Women’s Health:
New Data from the 1995 National Survey of Family Growth,”
Vital and Health Statistics 23 (1997), 1-114.
[2] Alan Guttmacher Institute, Facts in Brief: Induced Abortion (2000).
[3] Arcidiacono, P., A. Kwaja and L. Ouyang, “Habit Persistence and Teen Sex: Could Increased Access to
Contraception have Unintended Consequences for Teen Pregnancies?,” mimeo, Duke University, 2005.
[4] Bankole, A., J. Darroch and S. Singh, “Determinants of Trends in Condom Use in the United States,
1988-1995,” Family Planning Perspectives 31 (1999), 246-271.
[5] Becker, G. and N. Tomes, “Child Endowments and the Quantity and Quality of Children,”
Journal of Political Economy 84 (1976), S. 143-62.
[6] Brown, S. and L. Eisenberg, The best Intentions: Unintended Pregnancy and the Well-Being of Children and Families.
(Washington: National Academy Press, 1995).
[7] Bruine
bal
de
and
Bruin,
W.,
Numerical
B.
Fischho¤,
Expressions
of
S.
Millstein
Probability.
and
L.
‘It’s
a
Halpern-Felsher,
Fifty-Fifty
“Ver-
Chance’,”
Organizational Behavior and Human Decision Processes 81 (2000), 115-31.
[8] Centers for Disease Control and Prevention. “Youth Risk Behavior Surveillance: National College Health
Risk Behavior Survey,” Surveillance Summaries 46 (1997), 1-54.
[9] Curtis,
S.
and
K.
Neitzel,
“Contraceptive
Knowledge,
Use
and
Sources,”
Demographic and Health Surveys Comparative Studies 19 (1996).
[10] Delavande, A., “Measuring Revisions to Subjective Expectations,” mimeo, Universidade Nova de Lisboa/RAND, 2006.
[11] Demographic and Health Survey, “Model A questionnaire, with Commentary for High Contraceptive
Prevalence Countries,” Measure DHS+ Basic Documentation, Number 1.
[12] Dominitz, J. and C. Manski, “Eliciting Student Expectations of the Returns to Schooling,”
Journal of Human Resources 31 (1996), 1-26.
[13] Dominitz, J. and C. Manski, “Using Expectations Data to study Subjective Income Distribution,”
Journal of the American Statistical Association 92 (1997), 855-67.
32
[14] Easterlin,
R.,
R. Pollak and M. Wachter,
ity Determination:
“Towards a More General Model of Fertil-
Endogenous Preferences and Natural Fertility,” in R. A. Easterlin, ed,
Population and Economic Change in Less Developed Countries (Chicago: University of Chicago Press,
1980), 81-135.
[15] Ellsberg, D., “Risk, Ambiguity, and the Savage Axioms,” Quarterly Journal of Economics 75 (1961),
643-69.
[16] Emans, S., E. Grace, E. Woods, D. Smith, K. Klein and J. Merola, “Adolescents’Compliance with the
Use of Oral Contraceptives,” Journal of the American Medical Association 257 (1987), 3377-81.
[17] Erdem, T., M. Keane and J. Strebel, “Learning About Computers: An Analysis of Information Search
and Technology Choice,” Quantitative Marketing and Economics 3 (2005), 207-47.
[18] Fischho¤, B., A. Parker, W. Bruine de Bruin, J. Downs, C. Palmgren, R. Dawes and C. Manski, “Teen
Expectations of Signi…cant Life Events,” Public Opinion Quarterly 64 (2000), 189-205.
[19] Fu, H., J. Darroch, T. Haas, N. Ranjit, “Contraceptive Failure Rates: New Estimates from the 1995
National Survey of Family Growth,” Family Planning Perspectives 31 (1999), 56-63.
[20] Garcia, S., R. Snow and I. Aitken, “Preferences for Contraceptive Attributes: Voices of Women in
Ciudad Juarez, Mexico,” International Family Planning Perspectives 23 (1997), 52-8.
[21] Grady, W., D. Klepinger, A. Nelson-Wally, “Contraceptive Characteristics: The Perceptions and Priorities of Men and Women,” Family Planning Perspectives 31 (1999), 168-175.
[22] Grady, W., J. Billy and D. Klepinger, “Contraceptive Method Switching in the United States,”
Perspectives on Sexual and Reproductive Health 34 (2002), 135-145.
[23] Guzman, S., R. Snow and I. Aitken, “Preferences for Contraceptive Attributes: Voices of Women in
Ciudad Juarez, Mexico,” Family Planning Perspectives 23 (1997), 52-58.
[24] Hatcher, R., J. Truessell, F. Steward,
W. Cates, G. Stewart, F. Guest and D. Kowal,
Contraceptive Technology, 17th revised edition, (Ardent Media Inc., New York, 1998).
[25] Hanemann,
M.,
“Valuing
the
Environment
Through
Contingent
Valuation,”
Journal of Economic Perspectives 8 (1994), 19-43.
[26] Heckman, J. and R. Willis, “’Estimation of Stochastic Model of Reproduction: An Econometric Approach,” in N. E. Terleckyj, ed., Household Production and Consumption (New York: Columbia University Press, 1976).
33
[27] Henshaw, S., “Unintended Pregnancy in the United States,” Family Planning Perspectives 30 (1998),
24-29 & 46.
[28] Hurd, M. and K. McGarry, “Evaluation of the Subjective Probabilities of Survival in the Health and
Retirement Study,” Journal of Human Resources 30 (1995), S. 268-92.
[29] Hurd, M., J. Smith and J. Zissimopoulos, “The E¤ects of Subjective Survival on Retirement and Social
Security Claiming,” Journal of Applied Econometrics 19 (2004), 761-75.
[30] Lochner, L., “Individual Perceptions of the Criminal Justice System,” American Economic Review 97
(2007), 444-460.
[31] Manski, C., “Measuring Expectations,” Econometrica 72 (2004), 1329-1376.
[32] McFadden, D., “Measuring Willingness-To-Pay for Transportation Improvements,” in T. Gärling, T.
Laitila, and K. Westin, eds, Theoretical Foundation of Travel Choice Modeling (Amsterdam: Elsevier
Science, 1998), 339-64.
[33] McFadden,
D.
and
K.
Train,
“Mixed
MNL
Models
for
Discrete
Response,”
Journal of Applied Econometrics 15 (2000), 447-470.
[34] Nyarko,Y. and A. Schotter, “An Experimental Study of Belief Learning Using Elicited Beliefs,”
Econometrica 70 (2002), 971-1005.
[35] Pharmacia & Upjohn. “What women want:
international survey of birth control methods,”
http://www.birthcontrolresources.com/what.htm, accessed in 2000.
[36] Ranjit, N., A. Bankole, J. Darroch and S. Singh, “Contraceptive Failure in the First Two Years of Use:
Di¤erences across Socioeconomic Subgroups,” Family Planning Perspectives 33 (2001), 19-27.
[37] Rosenberg, M., “Weight Change with Oral Contraceptive Use and during the Menstrual Cycle. Results
of Daily Measurements,” Contraception 58 (1998), 345-9.
[38] Rosenzweig, M. and P. Schultz, “Estimating a Household Production Function: Heterogeneity, the
Demand for Health Inputs, and their E¤ects on Birth Weights,”Journal of Political Economy 91, (1983),
723-746.
[39] Rosenzweig, M. and P. Schultz, “The Demand for and Supply of Births: Fertility and Its Life Cycle
Consequences,” American Economic Review 75, (1985), 992-1015.
[40] Rosenzweig, M. and P. Schultz, “Schooling, Information and Nonmarket Productivity: Contraceptive
Use and Its E¤ectiveness,” International Economic Review 30 (1989), 457-77.
34
[41] Rosenzweig, M. and D. Seiver, “Education and Contraceptive Choice: A Conditional Demand Framework,” International Economic Review 23 (1982), 171-98.
[42] Schwartz,
J.
and
H.
Gabelnick.
“Current
Contraceptive
Research,”
Perspectives on Sexual and Reproductive Health 34, (2002), 310-16.
[43] Swait, J., J. Louviere and M. Williams, “A Sequential Approach to Exploiting the Combined Strengths
of RP and SP Data: Application to Freight Shipper Choice,” Transportation 21 (1994), 135-52.
[44] Tanfer, K., L.A. Cubbins and K. Brewster, “Determinants of Contraceptive Choice Among Single
Women in the United States,” Family Planning Perspectives 24, (1992), 155-61.
[45] Todd, P. and K. Wolpin, “On the Speci…cation and Estimation of the Production Function for Cognitive
Achievement,” Economic Journal 113, (2003), 3-33.
[46] United Nations, Population, Development and HIV/AIDS with Particular Emphasis on Poverty. The Concise Report,
(New York: United Nations Educational, 2005).
[47] Van der Klaauw, W., “On the Use of Expectations Data in Estimating Structural Dynamic Models: An
Analysis of Career Choices,” mimeo, Federal Reserve Bank of New York, 2000.
[48] Van der Klaauw, W., and K. Wolpin, “Social Security and the Retirement and Savings Behavior of Low
Income Households,” mimeo, Federal Reserve Bank of New York, 2005.
[49] Walker, J., “Pregnancy and Fertility Expectations: Estimates of Bounded Rationality and Unintended
Births,” mimeo, University of Wisconsin, 2002.
[50] Wolpin,
K.,
“An Estimable Dynamic Stochastic Model of Fertility and Child Mortality”
Journal of Political Economy 92 (1984), 852-874.
35
Table 1 : Demographic characteristics of the respondents - Sample Distribution (N=100)
Age
18-20
21-24
25-41
40
31
29
Marital Status
Single
Married
Living in couple
Separated
72
19
7
2
Years of Schooling
9-13
14-15
>15
Go to school
Regular partner
27
36
37
Religion
Catholic
Christian non-Catholic
Jewish
Hindu / Buddhist
Muslim
No religion
37
19
10
4
1
30
Race / Ethnicity
White non-Hispanic
Black non-Hispanic
Hispanic
Asian
American Indian
49
17
15
16
3
82
82
Table 2: contraception methods: choice and choice set
Current method
(in %)
Ever-used
methods (in %)
Percentage of
respondents
having method in
choice set
Birth Control Pill
32
26.8
100
IUD
0
1.5
75
Diaphragm (+ Spermicides)
0
1.5
89
Methods
Cervical Cap (+ Spermicides)
0
0.5
63
Condoms
29
35.9
100
Female Condoms
1
2.5
86
Spermicides
1
2.0
94
Condoms + Spermicides
0
1.5
94
Birth Control Pill + Condoms
22
13.1
100
Sponge
0
1.0
64
Norplant
0
0.0
59
Depo-Provera
5
6.1
86
Lunelle
0
0.0
13
Nuvaring
1
0.5
23
Birth Control Patch
2
1.5
58
Morning after pill
0
1.0
91
Sterilization
4
2.0
90
No "modern" method
Number of observations
3
2.5
100
100
198
100
36
Table 3: Subjective probability that using a method would interfere
with sex play and romance - All respondents and all methods
Subjective beliefs
0
1
2
3
4
5
6
7
8
10
12
15
19
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
98
100
N
Percent
38.48
0.58
2.4
0.8
0.5
7.5
0.1
0.2
0.3
11.0
0.1
4.3
0.1
5.9
2.0
4.9
0.1
2.3
0.3
6.6
0.1
1.7
0.1
1.4
0.9
1.9
0.1
0.9
0.1
4.5
1 385
37
Table 4: Percent chance of getting pregnant Birth Control Pill
All respondents
Ever-user
Subjective beliefs
Percent
Percent
0
0.1
0.2
0.3
0.5
1
2
3
4
5
6.5
7
9
10
15
20
30
40
50
60
70
100
N
9
1
1
1
2
18
15
11
1
18
1
2
1
5
1
3
2
2
2
2
1
1
100
7.55
1.89
1.89
1.89
3.77
24.53
11.32
13.21
18.87
5.66
1.89
1.89
1.89
1.89
1.89
53
Table 5: Percent chance of experiencing weight change - DepoProvera
Subjective beliefs
Percent
0
1
2
3
4
5
8
10
12
15
18
20
25
30
35
40
50
60
70
75
90
95
100
N
14.0
2.3
2.3
3.5
1.2
10.5
1.2
8.1
1.2
9.3
1.2
9.3
1.2
1.2
2.3
3.5
14.0
2.3
1.2
1.2
1.2
1.2
7.0
86
38
Table 6: Choice model estimation - homogeneous preferences
Coefficients
SE
∆u for getting pregnant
-10.466
2.175
∆u for side effects
-0.967
0.541
∆u for weight change
0.977
0.490
∆u for contracting an STD
-8.229
1.806
∆u for interference with romance
-1.315
0.690
∆u for getting pregnant within 1 year
after stopping
1.235
0.274
∆u for partner disapproval
-5.657
1.092
∆u for religion disapproval
-1.467
1.029
a - expected monthly cost
-0.053
0.011
100 observations - Subjective probabilities are between 0 and 1
Table 7: Implied monthly willingness to pay to reduce the
probability of experiencing outcome j by 0.05 computed with
estimates of Table 6
willingness to pay in $
Getting pregnant
Side effects
Weight change
Contracting an STD
Interference with romance
Getting pregnant within 1 year after
stopping
9.88
0.91
-0.92
7.76
1.24
Partner disapproval
Religion disapproval
5.34
1.38
-1.17
Table 8: Marginal effects: predicted probabilities of choosing methods when beliefs change
Initial predicted probabilities of choosing method
Pill
Condom
Depo-Provera
0.131
0.126
0.099
New predicted probabilities of choosing method
Pr. of getting pregnant + 5%
Pr. of side effects + 5%
Pr. of weight change + 5%
Pr. of contracting an STD + 5%
Pr. interference with romance + 5%
Pr. of getting pregnant within 1 year
after stopping + 5%
0.082
0.126
0.137
0.091
0.124
0.079
0.121
0.132
0.087
0.119
0.061
0.095
0.103
0.068
0.093
0.138
0.133
0.105
Pr. partner disapproval + 5%
Pr. religion disapproval + 5%
Expected monthly cost + $5
0.102
0.123
0.104
0.098
0.118
0.100
0.076
0.093
0.078
39
Table 9: Choice model estimation - additional methods' characteristics
∆u for getting pregnant
∆u for side effects
∆u for weight change
∆u for contracting an STD
∆u for interference with romance
∆u for getting pregnant within 1 year
after stopping
∆u for partner disapproval
∆u for religion disapproval
a - expected monthly cost
Methods'
dummies *
Mode of
administration
Mode and
frequency of
administration
-11.338
-9.722
-9.364
2.327
2.274
2.223
-1.182
-1.409
-1.305
0.571
0.546
0.573
0.788
0.147
-0.398
0.496
0.546
0.579
-7.914
-6.334
-5.936
1.793
1.405
1.223
-2.339
-1.958
-1.889
0.747
0.738
0.738
1.099
0.262
0.265
0.483
0.860
0.844
-6.175
-4.474
-4.113
1.227
0.999
0.839
-0.489
-0.556
-0.080
1.269
1.047
1.227
-0.054
-0.043
-0.050
0.013
0.013
0.012
0.195
1.498
0.435
0.779
(ref)
(ref)
-2.587
-1.525
0.495
0.591
-1.269
-0.524
0.499
1.147
-3.130
-0.758
0.733
1.092
∆u for No female administration
∆u for Oral administration
∆u for Local administration by self /
administration before intercourse
∆u for injection
∆u for Local adm. by doctor
Prescription
-1.335
1.037
Take everyday or every week
3.072
1.040
Take every month/every 3 months
2.204
1.007
Take less often than every 3 months
(ref)
* Coefficients associated with the dummies not shown.
Standard Error in Italics
40
Table 10: Implied monthly willingness to pay to use mode of
administration X rather than oral administration
No female administration
Oral administration
Local administration by self
Injection
Local administration by doctor
Willingness to pay in $
29.7
0.0
-30.3
-10.4
-15.1
Table 11: Choice model estimation with heterogeneous preferences
(1)
-10.383
Beta for pregnant - constant
(2)
(3)
2.612
Beta for pregnant - married
-2.023
Beta for pregnant - number of children
5.059
8.510
5.808
Beta for pregnant - (number of children)
2
-1.299
2.091
∆u for pregnant - White
-19.385
5.656
∆u for pregnant - non-White
-7.840
2.155
∆u for pregnant
-10.420
2.230
∆u for side effects
∆u for weight change
∆u for contracting an STD
∆u for interference with romance
∆u for getting pregnant within 1 year
after stopping
~
-1.209
-1.042
-1.037
0.583
0.556
0.581
-0.420
-0.454
-0.289
0.590
0.586
0.592
-5.953
-5.726
-6.302
1.156
1.153
1.250
-1.897
-2.011
-1.850
0.757
0.766
0.747
0.316
0.047
0.931
0.940
∆u6
∆u for getting pregnant within 1 year
after stopping
9.338
4.025
0.901
b
0.277
∆u for partner disapproval
∆u religion disapproval
a - Expected monthly cost
-4.139
-4.321
-4.452
0.926
0.845
0.917
-0.170
0.043
0.088
1.396
1.214
1.144
-0.050
-0.050
-0.052
0.012
0.013
0.013
Standard Error in Italic
All the specifications include dummies for prescription and mode and frequency of administration
41
Table 12: Elicited Willingness To Pay for
effectiveness (100% effective method rather than a
free 85% effective method)
WTP
Frequency
Cum.
10
15
20
25
27.5
30
35
40
50
60
75
80
90
100
120
150
200
300
9
5
10
7
1
21
1
7
21
2
2
1
2
6
1
1
2
1
9
14
24
31
32
53
54
61
82
84
86
87
89
95
96
97
99
100
Table 13: Choice model estimation using elicited WTP for
effectiveness
Coefficients
SE
∆u for side effects
-1.731
0.584
∆u for weight change
-0.145
0.561
∆u for contracting an STD
-6.317
1.206
∆u for interference with romance
-2.300
0.698
∆u for getting pregnant within 1 year
after stopping
0.229
0.474
∆u for partner disapproval
-3.974
0.785
∆u for religion disapproval
0.166
1.025
Expected monthly cost
-0.020
0.005
Includes dummies for prescription and mode and frequency of administration
42
Table 14: Random coefficients model : heterogeneity in preferences for "contracting
an STD"
Contracting an STD (∆u4bar)
Contracting an STD (rho)
Normal
Lognormal
-9.330
-2.860
3.541
3.134
-2.729
-6.812
1.549
10.007
Discrete - k=2
Contracting an STD (∆u41)
-13.743
6.229
Contracting an STD (u42)
-4.082
2.360
Contracting an STD (lamba1)
0.396
1.057
Contracting an STD (lamba2)
-0.396
1.058
∆u for side effects
∆u for weight change
∆u for interference with romance
∆u for getting pregnant within 1 year
after stopping
∆u for partner disapproval
∆u for religion disapproval
Expected monthly cost
-1.750
-1.727
-1.705
0.551
0.595
0.601
-0.107
-0.112
-0.022
0.57373
0.562
0.588
-2.341
-2.348
-2.323
0.745
0.726
0.768
0.262
0.253
0.262
0.4636
0.439
0.462
-3.981
-3.985
-3.922
0.796
0.857
0.907
0.201
0.208
0.154
1.0952
1.070
1.089
-0.019
-0.019
-0.019
0.004
0.005
0.005
Standard Error in Italics
Lognormal and normal specifications computed using simulated likelihood with 1000 draws
All the specifications include dummies for prescription and mode and frequency of administration
43
Table 15: Choice model estimation with failure rates
(1)
Failure rates
(2)
(3)
-3.207
-0.010
0.996
0.013
Failure rate - White
-3.684
1.627
Failure rate - non-White
-2.997
1.125
Elicited expectations
-9.103
2.359
∆u for side effects
∆u for weight change
∆u for contracting an STD
∆u for interference with romance
∆u for getting pregnant within 1 year after stopping
∆u for partner disapproval
∆u8 religion disapproval
a - expected monthly cost
-1.239
-1.240
-1.361
0.537
0.533
0.576
-0.221
-0.217
-0.395
0.562
0.560
0.588
-6.472
-6.447
-5.867
1.317
1.216
1.474
-1.957
-1.983
-1.799
0.702
0.696
0.737
-0.432
-0.470
0.212
0.920
0.918
0.967
-4.032
-4.002
-4.101
0.776
0.756
0.970
-0.724
-0.661
-0.343
0.987
1.097
1.079
-0.051
-0.052
-0.052
0.014
0.015
0.015
Standard Error in Italics
All the specifications include dummies for prescription and mode and frequency of administration
44
Table 16: Percent chance of getting pregnant: medical statistics and respondents' answers by method
% of Women experiencing an
unintended pregnancy within the
first year of use
Methods
Birth Control Pill
IUD
b
Diaphragm (+ Spermicides)
Cervical Cap (+ Spermicides)
c
Typical Use
Perfect Use
5
0.1 a
Respondents
Mean Answer
Respondents
Median Answer
N
8.61
3
100
1
0.73
8.88
5
75
20
6
14.79
10
89
20
9
13.07
10
63
Condoms
14
3
14.03
10
100
Female Condoms
21
5
15.55
10
86
Spermicides
26
5
24.96
20
94
Condoms + Spermicides
NA
0.30
9.84
5
94
Birth Control Pill + Condoms
NA
NA
2.82
1
100
Sponge c
20
9
20.69
15
64
Norplant
0.05
0.05
7.02
2
59
Depo-Provera
0.30
0.30
3.91
2
86
Lunelle
d
Nuvaring
d
Birth Control Patch
d
Morning after pill
Sterilization
No modern method
<1
7.15
5
13
1.50
7.67
5
23
1
8.01
5
58
NA
e
0.50
0.50
19% to 80%
f
9.70
5
91
1.21
0
90
55.6
60
100
Source: Hatcher et al. (1998) except otherwise noted
c
Combined birth control pill
Rate for nulliparous women
b
d
Average rate for Progesterone T, Copper T 380A and LNg 21
Source: clinical trials results from patient prescribing information
e
Treatment initiated within 72 hours of unprotected sex reduces the risk of pregnancy by at least 75%
a
f
The failure rate is 19% for withdrawal, 25% for periodic abstinence and 80% for no method at all.
45
Table 17: Distribution of subjective beliefs in percentage
Getting pregnant
Side effects
Subjective beliefs
0
0.1
0.5
1
2
3
4
5
6
7
8
10
12
15
17
19
20
25
30
35
40
45
50
55
60
65
70
75
80
85
88
90
95
97
98
99
100
N
Condom
Depo-Provera
Diaphragm
IUD
6.0
16.3
1.2
4.7
16.3
23.3
4.7
1.2
19.8
4.5
10.7
3.4
3.4
1.1
2.3
19.1
1.3
2.7
1.3
2.0
11.0
4.0
4.0
16.0
1.0
1.0
2.0
16.0
44.0
1.2
4.0
1.2
7.0
17.3
2.3
1.2
12.8
4.7
4.0
5.8
9.3
1.2
1.2
3.0
3.5
1.2
2.3
4.7
2.3
3.5
3.0
3.0
3.0
8.1
3.5
1.2
2.3
13.3
2.7
5.3
1.3
1.3
6.7
1.2
1.1
2.7
1.0
1.0
9.0
1.0
1.0
1.2
3.4
4.7
1.2
19.8
1.1
5.3
2.7
1.3
1.2
2.3
1.2
1.1
12.4
4.5
5.6
86
2.3
1.2
2.3
12.0
2.0
3.0
100
No-method
1.0
4.0
7.0
2.3
3.0
Depo-Provera
Partner's
disapproval
Female
Condom
32.6
1.2
10.0
1.0
1.0
5.0
STD
5.8
9.3
2.3
1.2
16.3
2.3
4.5
14.6
1.1
12.4
5.8
Female
Condom
43.0
Return to
fertility
10.7
1.3
2.7
75
89
46
7.0
2.3
4.7
8.1
11.6
1.2
86
5.8
1.2
1.2
2.3
1.2
7.0
86
8.1
2.0
1.2
1.2
1.2
1.2
4.7
2.3
1.2
4.7
1.0
1.0
4.0
100
7.0
86
1.0
2.0
Figure 1. Predicted Probabilities of
choosing "no-method"
Figure 2. Predicted Probabilities of
choosing the birth control patch
Figure 3. Predicted Probabilities of
choosing spermicides
Figure 4. Predicted Probabilities of
choosing Depo-Provera
47