Social Learning and Health Plan Choice
Alan T. Sorensen∗
I use data from the University of California to empirically examine the role of social learn-
ing in employees’ choices of health plans. The basic empirical strategy starts with the obser-
vation that if social learning is important, health plan selections should appear to be correlated
across employees within the same department. Estimates of discrete choice models in which
individuals’ perceived payoffs are influenced by coworkers’ decisions reveal a significant (but
not dominant) social effect. The strength of the effect depends on factors such as the depart-
ment’s size or the employee’s demographic distance from her coworkers. The estimated effects
are present even when the model allows for unobserved, department-specific heterogeneity in
employee preferences, so the results cannot be explained away by unobservable characteristics
that are common to employees of the same department. (JEL D12, D83, I11)
Individuals often have incentives to learn from their neighbors before making economic decisions.
In some cases, other individuals’choices serve as signals of private information, so learning comes
from merely observing their actions—for instance, a typical tourist will (rationally) avoid empty
restaurants and prefer those that are crowded with locals. In other cases, information and experi-
ences are shared directly through conversation, as when a consumer planning to buy a car asks her
friends about their experiences with different brands or different dealers. A role for social learning
exists whenever grouped agents use independent information to make parallel decisions involving
∗StanfordUniversityandNBER; email@example.com. I thankMicheleFrenchandtheUniversityofCalifornia
for providing the data, Zhigang Li for outstanding research assistance, the editor and referees for helpful comments,
and the National Science Foundation for financial support (grant SES-0079201). Any errors are mine.
The objective of this paper is to quantify the impact of social learning in a specific context:
individuals’ choices of employer-sponsored health plans. Using data from five University of Cal-
ifornia campuses, I show that health plan choices are “clustered” within departments: the choices
of employees in the same department are too similar relative to what we’d expect based on in-
dividual characteristics and campus-wide patterns. I estimate econometric models of health plan
choice that explicitly allow for social learning, and find large, statistically significant effects that
are robust across campuses and model specifications.
Understanding the demand for employer-sponsored health insurance is increasingly important
given the steep upward trend in health care costs. As with other experience goods, the quality
of a health plan (or one’s match quality with a health plan) is very difficult to know ex ante, and
choosing the “wrong” plan can be costly—especially because the costs of switching can be high.
Thissuggeststhat providinginformation—whichemployersgenerally do bydistributingbrochures
or hosting benefits seminars—is critical. But in some cases the most valuable information comes
from others who have already experienced the experience good. By assessing the significance of
peer influence in employees’ health plan selections, this paper sheds some light on the process
through with information diffuses in a large organization, and also speaks to the importance of
peer recommendations in the demand for experience goods more generally.
While the findings regarding health plan enrollment decisions are interesting in their own right,
this paper’s contribution to the broader literature on social learning extends beyond the specific
context of health plan choice. A growing body of theoretical research has incorporated social
learning into standard models of economic decision-making, showing that social effects can alter
those models’ predictions in important ways.1However, social effects have been notoriously dif-
ficult to quantify empirically, in large part due to identification problems that have been described
in detail by Manski (2000). Of principal relevance is the difficulty distinguishing between what
Manski calls “endogenous interactions,” in which individual decisions are influenced directly by
the decisions of their peers, and “correlated effects,” in which the decisions of individuals within
a group are similar due to shared (and possibly unobservable) characteristics. Previous empirical
studies have attempted to resolve this issue in a variety of ways, with varying degrees of success.
1See, for example, the papers by Banerjee (1992), Bikhchandani et al (1992), and Ellison and Fudenberg (1993).
Relative to the existing empirical literature, the contribution of this paper is to analyze a dataset
in which the presence of social learning can be demonstrated convincingly even while directly
allowing for unobserved heterogeneity. The data permit a rich set of comparisons and tests that
distinguish social learning from correlation in unobserved preferences. Most directly, the panel
nature of the data allows me to simultaneously estimate social effects and department-specific
unobservables. The estimated social effects are somewhat smaller but remain significant when
department-specific unobservables are included. A different robustness check (similar in spirit
to Munshi and Myaux (2000)) examines the own- and cross-group influences among faculty vs.
staff, finding that own-group effects are large and statisticallysignificant, while cross-group effects
are in most cases indistinguishable from zero. Moreover, the own-group effects are no larger
among faculty than among staff, even though the common unobservables problem is probably
more relevant to faculty.
Another well-known difficulty of estimating models with social interactions is that observed
choices are jointly endogenous: coworkers’ choices cannot be regarded as exogenous influences,
since they are in turn influenced by the choice of the employee in question. Rather than incorpo-
rating some notion of equlibrium in the estimation, as some authors have suggested, I handle this
issue by focusing solely on newly hired employees: I assume that new hires are potentially influ-
enced by the observed choices of existing employees, but not vice versa. (I explain in section 3
why this is a reasonable assumption.) This approach has the obvious advantage of simplifying the
estimation problem, and also focuses the model’s attention on employees who are actively making
health plan decisions.
Social learning hypotheses have been studied previously in many other contexts, including
crime (Glaeser, Sacerdote, and Scheinkman 1996), labor supply (Woittiez and Kapteyn 1998),
contraception (Munshi and Myaux 2000), adoption of fertilization technology (Conley and Udry
2000; Munshi 2000), welfare program participation (Bertrand, Luttmer, and Mullainathan 2000),
stock market participation (Hong, Kubik, and Stein 2001), and labor market outcomes (Bayer,
Ross, and Topa 2005). The study most similar to this one is that of Duflo and Saez (2002), who
examine individuals’ decisions about whether to enroll in a university-sponsored retirement plan.
Althoughsimilarin spiritto thepresent analysis, the instrumentalvariablesapproach usedby Duflo
and Saez is fundamentally different from the identification strategies employed here, which rely
primarily on data variation over time and across campuses.2Also, this paper specifically addresses
the issue of quantifying social effects in a polychotomous discrete choice setting.
2Background and Data
The University of California (UC) system is comprised of 9 university campuses plus four ad-
ditional research laboratories. Nearly all full-time employees and some part-time employees are
eligible to enroll in one of the health plans offered through the UC benefits program. The typ-
ical employee at a UC campus can choose from one of three HMOs (Health Net, Kaiser, and
Pacificare),3a point-of-service (POS) plan (UC Care), and a traditional fee-for-service plan (Pru-
dential High option). The HMO plans typically require little or no out-of-pocket payments from
the employee, while the POS plan requires monthly out-of-pocket payments ranging from $17-$50
(depending on the number of dependents to be covered under the employee’s plan). Enrollment
in the fee-for-service plan is extremely rare, since the out-of-pocket payments required are on the
order of $1,000 per month. Employees who choose not to enroll in one of the available plans are
automatically given minimal coverage through a default “Core Medical” option.4Plan enrollments
by campus are shown in table 1 for the year 2000.
The data used in this study were provided by the UC benefits office. The data cover employee
health plan decisions for the years 1995-2000 at each of the nine university campuses; I will focus
attention on the five largest campuses: Berkeley, Davis, Irvine, Los Angeles, and San Diego. For
each employee, the data indicate the plan chosen by the employee, the department in which the
employee works, the date the employee was hired, and the employee’s monthly salary, along with
2Their study utilizes multiple observations on the same individual over time, and uses an instrumental variables
strategy suggested by Case and Katz (1991) to address the endogeneity problem potentially induced by common
unobservables. The salary or tenure composition of the individual’s department is used as an instrument for changes
in coworkers’ participation rates. In contrast, this paper uses only the (one-time) decisions of new employees, and
regards movements over time in a department’s average choices as a direct source of identifying variation.
3One additional HMO, Western Health, was available at UC Davis.
4This default option only covers catastrophic medical care, so virtually all employees choose another plan (as
shown in Table 1). The few who remain in the Core Medical plan are typically employees who are ineligible for one
of the other more comprehensive plans (e.g., because they work less than half time).
additional demographic characteristics including age, sex, and zip code of residence. Family status
can be roughlyinferred fromthe coveragetypechosen withthe healthplan(single-party,two-party,
or family). The availability of relatively rich demographic information is critical in this study, in
particular since we expect individual-level heterogeneity in preferences over health plans to be
driven largely by differences in age, income, and place of residence. Price considerations play
a diminished role in health plan decisions: though the decision of whether to enroll in the POS
plan or an HMO may be driven by cost concerns (the POS plan costs $15-$45 more per month
than the HMOs), decisions among the three HMOs are based on non-price considerations since the
employee premiums are essentially the same (zero in most years), and the required copays (e.g.,
for office visits or prescription drugs) are also generally uniform across plans. The set of available
options and the corresponding pricing structure has remained constant over the sample period for
most campuses, with only a few minor changes.5
Each campus has a benefits office responsible for disseminatinginformationabout the available
health insurance options. Newly hired employees are encouraged to attend orientations in which
the plans (along with other employee benefits) are explained, and brochures with basic information
are typically mailed to employees prior to periods of open enrollment. Most departments in the
university system have a staffperson assigned as the benefits coordinator who may serve as the
point person for information within departments. The information provided is intentionally neutral
(in the sense that it doesn’t favor any particular plan) because presumably different plans will be
optimal for different people. The plan shares shown in Table 1 seem to confirm the idea that no
single option is uniformly superior.6
Most of the institutional arrangements for disseminating health plan information focus on in-
forming employees about the process of enrolling or about the plans’ relative costs and payment
structures (e.g., explaining the difference between a point-of-service plan and an HMO). From an
employee’s perspective, much of the relevant information needed to choose among plans—e.g.,
5A substantial change in the pricing structure occurred just prior to our sample period; see Buchmueller and
Feldstein (1997) for an analysis of the impact of the price change on health plan switching using data very similar to
the data used here.
6This feature of health plan decisions is an interesting contrast to decisions about retirement benefits, where, for
example, it can be argued that every employee should contribute to a 401(k) plan.
Table 4: Maximum likelihood estimates, weighted peer effects
Coworkers’ choices (γ)
Distance weights (ρ):
Age (higher) 0.036
LR test of ρ = 0
Standard errors in parentheses. Coefficients on other covariates are omitted to save space (the estimates are very
similar to those reported in table 3). For the likelihood-ratio test of the hypothesis that the weight parameters (ρ)
are all zero, the table reports the χ2
6test statistics with the corresponding p-values in parentheses.
Table 5: Social effects and department size
(Per-coworker effect is γ0+ γ1N + γ2/N)
Standard errors in parentheses, adjusted for clustering by department. Coefficients on employee
demographics omitted to save space.
Table 6: Relative std. deviations of the components in utility
Based on the estimates reported in Table 3.
Table 7: Estimates of social effects with and without department-specific heterogeneity
Irvine BerkeleyDavisLos Angeles San Diego
Without dept. fixed effects:
Coworkers’ share (γ) 2.056
With dept. fixed effects:
Coworkers’ share (γ) 1.072
# of departments
LR test: no dept. effects
Models estimated using only those departments that hired new employees in all six years of the sample period.
Standard errors are in parentheses, adjusted for clustering by department. Plan-specific coefficients on employee
characteristics are omitted to save space. For the likelihood-ratio tests, χ2test statistics and p-values are shown.
(The degrees of freedom differ depending on the number of department fixed effects estimated at each campus.)
Table 8: Academic departments divided into faculty and staff subgroups
Irvine BerkeleyDavis Los Angeles San Diego
Faculty-Faculty (γff) 1.316
Faculty-Staff (γfs) 0.006
LR test: (γff= γss) & (γfs= γsf)
LR test: (γff= γfs) & (γss= γsf)
Standard errors in parentheses, adjusted for clustering by department. Plan-specific coefficients on employee
characteristics are omitted to save space. “Faculty-Faculty” reflects the impact of faculty coworkers’ choices on
new faculty’s choices, while “Faculty-Staff” indicates the impact of staff coworkers’ choices on a new faculty’s
choice. For the likelihood-ratio tests, χ2
2test statistics and p-values are shown.
Figure 1: Per-coworker social effect as function of department size
Average coworker influence