Doctor Switching Costs in Health Insurance. Gordon B. Dahl (UC San Diego and NBER) and. Silke J. Forbes (Case Western Reserve University)

Transcription

1 Doctor Switching Costs in Health Insurance Gordon B. Dahl (UC San Diego and NBER) and Silke J. Forbes (Case Western Reserve University) Abstract We estimate switching costs in U.S. health insurance coming from three different sources: attachment to one s existing doctor(s), inertia, and the option value of a larger network. We exploit a targeted change in the health insurance offerings of a large employer to estimate the size of these switching costs. Our setting is unique in allowing us to identify the costs that individuals associate with switching doctors separately from inertia and option value. We estimate that, when a new health plan is introduced, 22% of employees initially exhibit inertia, dropping to 16% of employees a year later, even though they could save between $638 and $1,851 annually by switching to an almost identical, but cheaper plan. About one third of individuals are willing to pay higher premiums to keep their doctor, even after controlling for inertia. The option value of a larger network is near zero in our setting. The inertia and doctor switching costs vary by income, age, gender, family composition and prior health care utilization patterns. Our findings have important implications for the design of public health insurance offerings, as well as for private employers who negotiate health plan options on behalf of their employees. In particular, we find that sicker and older individuals exhibit less inertia, but place a greater value on keeping their doctors.

2 1. Introduction Health care costs in the United States have been rising faster than inflation for more than a decade, and the share of GDP spent on health care is far higher in the U.S. than in any other industrialized nation (Blumberg and Holahan, 2014, and Kaiser Family Foundation, 2015). A potential solution to rising prices for health care is to increase competition among insurance and health care providers. However, if consumers face costs of switching among providers, then competition will be less effective as a means to contain health care prices (Klemperer, 1987, Farrell and Klemperer, 2007). Moreover, any policy reform requiring individuals to switch to certain health care providers could impose significant welfare costs on those individuals. A central issue in the policy debate over the U.S. Affordable Care Act (ACA), as well as other health care reform legislation, has been whether people would be required to switch doctors. 1 In the private market, when employers change insurance providers or when employees change jobs, employees and their families are often unable to keep their current network of doctors. 2 While the magnitude of these switching costs is important for evaluating the costs versus benefits of maintaining access to doctors, little empirical evidence on the topic currently exists. Several challenges make it difficult to estimate the cost of switching doctors. The first is that insurance plans involve a large vector of attributes, including which doctors are in a network, covered medical services, copay amounts and prescription drug coverage; hence, it is difficult to separate out willingness to pay to keep one's doctor from other plan characteristics. A related challenge is that premium changes usually occur either in relatively small increments or in combination with substantial changes in other insurance plan attributes. A final issue is that inattention, inertia or confusion by consumers about the details of insurance options makes it difficult to separate out what might broadly be called "inertia" from "willingness to pay". In this paper, we take advantage of a uniquely targeted change in the health insurance offerings of a large employer to estimate how many individuals are willing to pay higher premiums in order to keep their existing doctors. The change mirrored a randomized experiment 1 President Obama outlined the relevant trade off in an interview, say some people are "going to have to make some choices and they might end up having to switch doctors, in part because they're saving money" (WebMD, March 14, 2014). Cox et al. (2015) point out that on the ACA s Health Insurance Marketplaces, individuals may have to switch insurance plans and doctors annually to avoid significant premium increases. 2 An estimated 55 percent of the U.S. population was covered by employer-sponsored health insurance in 2011 (Current Population Survey, September 2012). 1

3 where within the same insurance plan, there were premium increases associated with some doctors (group A) but not others (group B). Prior to the change, employees could choose doctors in group A or B with no price difference. After the change, employees whose existing doctor was in group A were faced with a price increase which could only be avoided if they switched to a doctor from group B. In contrast, employees with existing B-group doctors were not charged more to keep their current doctor, as long as they actively opted in to avoid the price increase. Under the identifying assumption that there are no systematic quality differences between doctors in groups A and B (an assumption we test empirically), one can estimate the fraction of individuals willing to pay a higher premium to avoid switching doctors, as well as the amount of inertia. Importantly, other plan attributes remained identical both before and after the change, eliminating confounding factors. The actual details of the natural experiment are as follows. Prior to 2011, employees of the University of California (UC) system could choose from a variety of health insurance plans, with the most popular plan being Health Net. 3 The available set of primary care physicians (PCPs) individuals could choose from in this plan included UC affiliated doctors, as well as doctors in several other provider networks. Starting in 2011, UC expanded its health insurance offerings to include two versions of the Health Net HMO, one with the previously existing network of doctors ( Health Net or HN) and one with a more limited network ( Health Net Blue and Gold or HNBG). The HNBG network included all UC affiliated physician groups and hospitals, but excluded many of the other provider groups. Overall, about one third of all doctors were excluded from HNBG. The HNBG option was set up to contain premium rate increases, while the HN plan had premiums rise by as much as 100% depending on family type and income level. Importantly, all other aspects of insurance coverage remained the same as before for both insurance plans. The default option for all employees who were previously enrolled in HN remained HN, even if their PCP was in the HNBG network. The assumption that HN and HNBG doctors are viewed as substitutable by employees is central to our identification strategy. UC was able to instigate the creation of the HNBG insurance plan because it wielded strong pressure on UC affiliated doctors and hospitals to not 3 HealthNet is a Health Maintenance Organization (HMO) plan. Under this type of plan, primary care physicians (PCP) play a gatekeeper role. I.e., patients cannot independently seek care from a specialist doctor; they have to be referred by a PCP. 2

4 raise prices beyond the rate of inflation. It did not have the same ability to do this with the other provider networks. We argue that the differential bargaining power exerted by UC was institutionally driven, and unrelated to the relative quality of the two groups. When we test this claim empirically by comparing quality ratings, we find no statistical difference between the two groups of doctors. The behavior of employees whose existing doctors were included in the smaller HNBG network can be used to estimate the amount of inertia. For these employees, actively choosing to change insurance from HN to HNBG would result in a substantially lower premium, without the need to change doctors. But given the default option, an employee not paying attention or confused about the new plan would automatically stay enrolled in the now more expensive HN plan. Since all other insurance plan characteristics besides doctor networks remained the same, the fraction of these employees who do not switch represents the fraction of employees with inertia. In theory, it is possible that these employees do not want to switch to HNBG because they value the option of being able to change from an HNBG doctor to a doctor who is only in the broader HN plan at any time. 4 We argue this option value is close to zero in our setting for several reasons later in the paper. When we include measures of this option value in our empirical estimation, we find that they are mostly statistically insignificant and the size of the effects is always close to zero. Employees whose existing doctors were not covered under the HNBG option after the change in 2011 faced a trade-off: keep their current doctor and pay a higher premium to stay in HN, or switch to a PCP in the HNBG network and save money. Of course, there are two reasons these employees may not switch. They may be willing to accept higher premiums to keep their current doctor, but they may also exhibit inertia. Under the identifying assumption that the rate of inertia conditional on observables is the same across both groups, we can separate out the two factors by comparing the switching rates of these individuals to those whose existing PCPs are in the HNBG network (since this group has no doctor switching costs, but only inertia). Empirically, we find evidence of sizable switching costs, even in this relatively simple environment where the only thing changing was physician networks. Over two thirds of employees who were enrolled in HN prior to the change had PCPs who were included in the smaller HNBG network. Twenty-two percent of these employees exhibited inertia and remained 4 The option to switch is always available during the annual open enrollment period. 3

5 in HN the first year after the change, even though they could have saved between $638 and $1,851 by switching to HNBG. This fraction declines in the second year after implementation, with inertia dropping to 16% of employees. Our main result is that a substantial fraction of individuals are willing to pay higher premiums to keep their doctor. In the first year after the change, 35% of employees are willing to pay an average of over $100 extra per month to retain their doctors. Since it takes time to choose and transition to a new PCP, one might have expected a sizable decline in this fraction the second year after the change. However, the fraction drops only slightly, with 33% of employees willing to pay higher premiums to keep their doctor. Note that while these willingness-to-pay fractions are large, one would have erroneously concluded they were even larger without netting out inertia. Willingness to pay to keep a doctor explains why 61% of individuals do not switch insurance plans, while inertia accounts for 39% of non-switchers in the first year after implementation. Interesting patterns emerge across demographic groups in both inertia and doctor attachment. Multinomial logit models reveal that doctor attachment is a normal good, with high income individuals less likely to switch doctors, even after controlling for inertia. For the average earner whose existing doctor is not in the HNBG network, doubling their salary would reduce the likelihood that they would switch from HN to HNBG by 7.2 percentage points. Given the baseline switching rate of 35%, this represents a sizable income effect. We also find that inertia and doctor attachment are strongly influenced by age, gender, family composition and prior health care utilization patterns. Our paper contributes to two strands of the literature on health insurance choice. The first examines switching between health insurance plans as a result of price changes. Buchmueller and Felstein (1997) study the effect of relative price changes on switching between health plans that are close substitutes for each other, and find that even small relative price changes induce substantial switching between plans. 5 Strombom, Buchmueller and Feldstein (2002) examine how price elasticities vary with employee characteristics and find that younger and healthier employees are more likely to switch away from health plans with relative price 5 Buchmueller and Feldstein also use data from the UC system. Since they do not have information on employees doctors before and after the price changes, they do not study this aspect of switching behavior. Moreover, in the time period they study, many doctors were included in multiple plan networks so that employees could commonly keep their doctors when switching health plans. 4

6 increases. More recently, Gruber and McKnight (2014) examine switching from broad network (mostly PPO) plans to limited network (mostly HMO) plans among Massachusetts state employees, who were offered one-time financial incentives between $800 and $2300 to switch to a limited network plans. Gruber and McKnight find that, overall, about 10 percent of enrollees switch in response to the financial incentive, and enrollees who can keep their existing primarycare doctor are 60 percent more likely to switch than those who would have to change doctors. The second area of the literature we contribute to is the role of bounded rationality and inertia in health insurance choices. This includes a series of papers examining the suboptimal decisions by retirement-age individuals choosing Medicare Part D plans (e.g., Abaluck and Gruber, 2011 and 2013, Ericson, 2014, Heiss, Leive, McFadden and Winter, 2013 and Ketcham, Lucarelli, Miravete and Roebuck, 2012). Examining health insurance decisions of working-age individuals, Handel (2013) finds substantial inertia in insurance choices, and Handel and Kolstad (2013) demonstrate that information frictions and hassle costs play an important role. To our knowledge, our paper is the first to examine how the ability to keep one s existing doctor affects switching between otherwise identical insurance plans with different prices. We estimate this effect while controlling for inertia, something which has proven difficult to do in other settings. Our data also allows us to study inertia directly and estimate how inertia varies with employee demographics and health. The findings from our study are important for health care reforms, since such reforms often place restrictions on provider networks to save on costs. Identifying which individuals are willing to pay substantial sums to keep their doctor is important for determining which groups would suffer the largest welfare losses if they were forced into smaller networks that did not include their existing doctor. Our results also suggest that doctor attachment matters for private insurance options and could contribute to job lock. The remainder of this paper proceeds as follows. We first describe the institutional setting and data which make this study possible. In Section 3 we present our model. In Section 4, we discuss and test the assumptions underlying our identification approach. Section 5 presents descriptive evidence, and Section 6 presents our regression results. A final section concludes. 5

7 2. Institutional Background and Data 2.1 Health Insurance Options and Enrollment In 2010, employees at the University of California (UC) could choose between seven different health insurance options. These included a low-deductible and a high-deductible Preferred Provider Organization (PPO) option, a hybrid PPO-HMO, three HMO options and a high-deductible fee-for-service plan. Among the HMO s, Health Net (HN) offered the largest physician network, Kaiser Permanente offered coverage through its own network of physicians directly employed by Kaiser, and Western Health offered only a regional network, primarily in the Davis and Sacramento area. In 2011, the Health Net Blue & Gold (HNBG) option was introduced, with the intention of containing costs and holding the line on premium rate increases. In the same year, the University also switched from Cigna to Anthem as a provider for the highdeductible PPO plan; however, enrollment in these two plans was small (around 1 to 2 percent of all employees). The UC system sets employer contributions for insurance premiums based on a combination of income and family status. Employees are grouped into four income tiers, and each tier receives a fixed employer contribution which is higher for lower incomes. 6 Employer contributions are also differentiated based on whether the insurance covers the employee only, the employee plus children, the employee plus a spouse or the employee plus a spouse and children. Table 1 shows the annual insurance premiums for each plan that an employee with an income between $47,000-93,000 would pay if she only insured herself (Panel A) or if she insured herself, a spouse and one or more children (Panel B). 7 The table shows that in 2010 the PPO options were substantially more expensive than the HMO options. Among the HMO s, Health Net was the most expensive option, with an annual premium of $614 for an employee insuring herself only, compared to premiums of $470 for both Kaiser Permanente and Western Health Advantage. Thus, HN would have been the preferred choice for an employee who wanted an HMO with a large network of independent doctors and was willing to pay a modestly higher premium. 6 In 2011, the income tiers were set at $47,000 or less, $47,001-93,000, $93, ,000 and $140,001 or more. 7 Premiums for other income groups and for employees insuring a spouse and no children or children and no spouse are available upon request. 6

8 In 2011, the premium for HN increased substantially for this income group it approximately doubled. The new HNBG option was introduced at a small premium increase relative to what HN had cost in the previous year. Employees whose existing doctors were in the HNBG network could thus keep their doctors and all other features of their existing insurance at close to the same price that they had been paying before. Employees whose existing doctors were not in the HNBG network, however, would have to pay substantially higher premiums if they wanted to stay in Health Net and keep their existing doctors. Alternatively, these employees could enroll in HNBG and switch to a new doctor. They could, of course, also switch to one of the other insurance options. Table 2 shows the annual premium differences between HN and HNBG. In 2011, the gaps ranged from $638 for an employee insuring herself only to $1,851 for an employee also insuring a spouse and children. In 2012, these numbers increased slightly. The premium differences were independent of the employee s income (in absolute terms). Every year, UC employees can change their insurance plans during a month-long open enrollment period (usually in November for a change effective the following January). Employees who wanted to switch to HNBG had to make an active choice during open enrollment and fill out a form, either on a website or on paper. The default for employees who made no change during open enrollment was to stay in their existing health care plan. This was true even for individuals who were enrolled in HN in 2010 and whose existing doctor was in the HNBG network. Each employee received information about the HNBG introduction in the form of several s and an annually distributed flyer about UC insurance options. However, some individuals may have chosen not to read this information and thus may have been unaware of the change. Table 3 shows the health plan choices of full-time UC employees under the age of 65 for the years In 2010, Health Net was the most popular insurance option with 44.1 percent of employees choosing this plan. This was followed by Kaiser Permanente (KP) with 30.5 percent. Approximately 23 percent of employees chose one of the PPO options, and less than 3 percent chose the regional HMO, Western Health Advantage (WHA). In 2011, when HNBG was first introduced, 28.4 percent of employees chose this plan and HN s share fell from 8 As we explain below, we exclude employees who live in the zip code for UC Davis or the immediately adjacent zip codes in this and all following tables. 7

9 44.1 percent to 14.9 percent. In the following year, HNBG s enrollment share increased to 31.7 percent while HN s share fell further to 11.3 percent. Interestingly, the combined share of HN plus HNBG remained fairly close to the pre-change HN share. In both years, the enrollment share of PPO s fell slightly, while KP and WHA experienced small increases in enrollments. Appendix tables A.1-A.3 show the transition matrices for health plan choices from , and Data Sources and Estimation Sample Our primary data source is administrative records from the University of California, culled from several sources, for the years This includes the health insurance plan chosen by each employee in each year, demographic information about the employee and all insured family members, the employee s salary (in $5,000 bins), a record of each doctor visit and the treating physician (though not the reason for the visit) and, in the case of Health Net enrollees, each family member s primary care physician. All records were anonymized before they were given to us so that we cannot identify any individuals. We augment this data with information from Health Net. In addition to the medical group ratings described above, we received the doctor and hospital directories for the HN and HNBG networks. Our regression sample consists of full-time staff and faculty employees between the ages of who were enrolled in Health Net in In order to be able to focus on the decisions of existing employees, we require that individuals be in our sample for each year from We examine the health plan choices of new employees separately. We drop individuals with incomes below $25,000 and above $200,000 because we have very few observations in those ranges. Finally, we drop individuals who live in the zip code for UC Davis and immediately adjacent zip codes. UC Davis employees voiced substantial opposition to the HNBG introduction because the HNBG network excluded the largest doctor network in Davis, Sutter Medical Group and Hospital. While HNBG did include the UC Davis medical center and their affiliated doctors, the hospital is located in Sacramento and therefore less accessible to UC 9 Having a doctor in the HNBG network does not predict leaving employment at the university and thus exiting the sample. 8

10 Davis employees. Given these issues, we drop employees in the Davis area because their choices may be systematically different. Our final estimation sample includes 26,359 employees who we observe over three years. We do not include covered family members as separate observations because the insurance plan decision is made at the family level and all family members must be in the same insurance plan. Rather, we include family-member attributes, including their doctors, as characteristics of the employee's household in our regressions. 3. Switching Cost Model 3.1 Theoretical Model Let us assume initially that all consumers are attentive and all doctors are of equal quality. Assume that there are two doctors who are horizontally differentiated by their location on a Hotelling line. Doctor A is located at 0 and Doctor B is located at 1. Consumers are uniformly distributed between 0 and 1. The first time that consumer i chooses a doctor (in time period 0), her indirect utility is: V ii0 = γ i α i p j0 td ii, j = A, B (1) where p j0 is the price of doctor j (j = A, B) in period 0, d ii is the distance between individual i and doctor j, and α i, γ i, and t are parameters. In all future periods, the consumer incurs a fixed switching cost θ i if the doctor she chooses in period t is different from the doctor she chose in the previous period, t-1. Her indirect utility is then: V iii = γ i α i p jj td ii θ i s jj, j = A, B (2) where s jj is equal to 1 if the doctor in period t is different from the doctor in the previous period and equal to 0 otherwise. In all periods, consumers choose the doctor who maximizes their indirect utility. The effect of having switching costs in the model is that, after the first period, the marginal consumer will be willing to pay a higher price for doctor j than for doctor k if she was with doctor j in the previous period. Thus, switching costs reduce the intensity of competition after the first 9

11 period. 10 Our goal in the empirical estimation is to estimate the incidence of these switching costs. 11 Once we add inattention to the model, in each period a fraction φ of consumers choose to evaluate their current doctor choices and change their selection if it is utility-maximizing to do so. The remaining fraction (1 φ) simply stay with the doctor they had in the previous period. Inattention could be exogenously given, or consumers could be rationally inattentive. In the latter case, consumers would be aware that they have to pay a cost c to research the alternatives and make a decision. Consumers will only choose to pay this cost if their expected gain from re-optimizing is greater than c. Empirically, we cannot identify the source of inattention given the information we have in our data set. However, we will present evidence which suggests that a non-trivial proportion of consumers are inattentive. We will also show how this inattention varies with observable characteristics, such as age, gender, income and health status. 3.2 Econometric Model We will estimate a multinomial logit model of health insurance choice for the employees who were with HealthNet in 2010, prior to the introduction of HealthNet Blue&Gold. For these employees, we can get a meaningful estimate of the costs of switching from a doctor who is only in the HealthNet network to a doctor who is also in the HealthNet Blue&Gold network. In specifying the indirect utility to employee i of choosing plan alternative j, we begin with the following specification for indirect utility: V iii = X ii β α i p jj θ i s jj + ε ii (3) where X ii is a vector of employee characteristics, p jj is the plan premium and s jj is a dummy variable which is equal to 1 if the employee and/or the employee s family members would have to switch doctors in order to join the HNBG plan. ε ii is a logit error. However, once we control for family type and income, there is no variation in health care premiums within a given year. We cannot separately identify α i and θ i because p jj and s jj are 10 At the same time, switching costs intensify competition for consumers in the first period (see, e.g., Klemperer 1987a and 1987b). We do not have any data on that first period in our context and, therefore, do not estimate the effect of switching costs on competition in the initial period. 11 Specifically, we will estimate the proportion of consumers for whom the switching cost is greater than the price difference between the HealthNet and the HealthNet B&G plan. 10

12 perfectly collinear. 12 For example, consider employees with incomes between $47,000 and $93,000 insuring a spouse and children with HN in 2010, whose doctors will not be included in the HNBG network after the change in In 2011, these employees have the option of paying $4,330 to stay in HN and keep their doctors or paying $2,479 by switching to HNBG and choosing new doctors. Since there is only one price difference ($1,851), one cannot estimate a price coefficient. Instead, what is identified without further assumptions is the fraction of employees who are willing to pay $1,851 or more to keep their current doctor. Therefore, we use a simplified version of indirect utility: V iii = X ii β θ i s jj + ε ii (4) We can estimate θ i and use it to calculate the proportion of people whose cost of switching doctors is greater than the price difference between the HealthNet and the HealthNet Blue&Gold insurance. In our empirical implementation, we estimate separate regressions for 2011 and We do this because there may be frictions in the first year of the HNBG implementation, which could cause some employees to delay switching by one year. The regressions using 2011 choices will show the employees behavior in the first year of the implementation, while the regressions using 2012 choices will show the extent of switching that has occurred by the second year. To estimate doctor switching costs separately from inertia, we will control for inertia in X ii through our quasi-experimental design. The next section discusses our identification strategy. 4. Identification of Doctor Attachment and Inertia 4.1 Quasi-Experimental Setting As previewed in the introduction, our setting mirrors an experiment where individuals with certain doctors are assigned higher premiums if they want to keep their current doctor (group A), while individuals with other doctors are allowed to keep their doctor without a price increase as long as they make an active choice to do so (group B). For this to be an experiment, the higher premiums need to be randomly assigned to doctors. In this somewhat simplified 12 There is some variation over time, but this variation coincides with the diffusion of the new HNBG plan. 11

13 example, the random assignment of premiums to doctors helps to identify both inertia and doctor attachment. To identify inertia, one can use the fraction of individuals with doctors in group B who do not actively opt in to the lower premium. The fraction of individuals with doctors in group A who do not switch can be used to identify the combination of inertia plus willingness to pay more to keep one's doctor. Since which individuals have group A versus group B doctors is randomly assigned, one can compare the fraction of non-switchers in the two groups to separately identify willingness to pay from inertia. While we do not have an actual experiment, we take advantage of a natural experiment which assigned higher versus lower prices to certain physician groups in a way that appears to be independent of doctor quality. We argue that which provider groups agreed to be part of the HNBG network had little to do with relative quality or other demand factors, and more to do with bargaining power. The central administration at UC instigated the creation of the HNBG insurance plan to contain cost and premium increases. It exerted strong pressure on UC affiliated doctors and hospitals to join HNGB and was successful in the attempt. It did not have the same power to do this with the other provider networks, such as Scripps Health (a large provider in Southern California), which did not join HNBG. 13 We argue the differential bargaining power exerted by UC was institutionally driven, and unrelated to the relative quality of the two groups. 14 The fact that no other aspects of insurance coverage changed also helps to cleanly identify the effects. 4.2 Doctor Substitutability across the Two Networks A first reason to believe that doctors in the broader HN network and the more limited HNBG network are substitutes relates to the fact that both sets of doctors were in the same insurance plan prior to Before the change in insurance plan options, employees choosing HN could choose UC doctors as well as a variety of doctors affiliated with other providers without any distinction in cost. In 2010, before the change, 70% of employees choose doctors who would later be in the HNBG network, while 30% choose doctors who would later not be in 13 Many non-uc providers are also in HNBG. It is difficult to know what negotiations led them to be included in the network, but as we show empirically, their inclusion is not related to observed quality. 14 In a related setting, Grennan (2014) shows that bargaining power matters for how much different hospitals pay for the same product from the same supplier. For medical devices, variation in bargaining ability can explain 79% of the observed price variation and has a large firm-specific component. 12

14 the HNBG network. Hence, HNBG doctors were not only a popular choice, but there was also a sizable network of doctors to choose from. We explore which observable characteristics of employee households explain doctor choice prior to the HNBG introduction. We find that proximity between the home zip code and the doctor s zip code is by far the most important predictor of which doctor is chosen. The second piece of evidence for doctor substitutability comes from medical group ratings which are published by Health Net on its website. Each medical group is given a rating from one (lowest quality) to five stars (highest quality) in three broad categories: member satisfaction, clinical care and preventive health. We construct three measures of aggregate quality across these categories. The first two measures are dummies which equal one if the medical group has at least three or at least four stars in each category, respectively. The third measure adds up the number of stars across the three categories. We merge the medical group ratings to the primary care physician for each employee in our sample who was enrolled in Health Net in We then regress each of the three aggregate rating measures on an indicator for the doctor being in the HNBG network in 2011, controlling for fixed effects for the doctor s five-digit zip code. We cluster the standard errors at the medical group level because that is the level of variation in the ratings data. The results of these regressions are presented in Table 4. In each case, we find that the coefficient on the HNBG doctor dummy is not significantly different from zero, indicating that HNBG doctors do not have a systematically different rating than doctors who are only in the larger HN network. As a final piece of evidence that the HNBG network was high quality, consider the U.S. News and World Report rankings of top hospitals for Out of 440 hospitals in the state of California, the five UC campuses with medical centers rank #1 (UCLA), #2 (UCSF), #5 (UCSD), #9 (UCI) and #16 (UCD). Moreover, in the narrower metro area rankings, the UC hospitals are all the #1 hospitals in their respective geographic areas. 15 While an excellent medical center is no guarantee that affiliated UC doctors are also excellent, and while not all HNBG doctors have a UC affiliation, these rankings are certainly suggestive UCLA and UCI are both in the same metro area of Los Angeles; UCLA is #1 and UCI is #4 out of 145 hospitals in the LA metro area. UC Riverside established a new medical school in 2008, but only started enrolling its first class in We note that if the HNBG network is actually preferred to the HN network, then our estimates of doctor attachment are biased downward. 13

15 4.3 Inertia and Option Value To identify inertia, we take advantage of the non-switching rates of employees whose existing doctors were included in the smaller HNBG network. These employees could save a large sum of money, and still keep their current doctor, by switching from the HN to the HNBG network. Since employees had to make an active choice to achieve these premium savings, we use the fraction of non-switchers to identify the amount of inertia. Inertia could be driven by inattention (i.e., individuals are unaware that they could save a substantial amount of money), or employees may be reluctant to choose the HNBG insurance because of the value they place on the option to switch to a HN doctor mid-year (between open enrollment periods). In order to control for the option value of a larger network, we exploit the fact that the proportion of HN doctors who are also in the HNBG network varies geographically. In areas where all HN doctors are also in the HNBG network, the option value of the HN insurance plan is literally zero since the two networks are exactly the same. In areas where only a small fraction of HN doctors are also in the HNBG network, the option value of the HN insurance plan is potentially larger. We use this variation to identify the size of the option value and we assume that the remaining inertia is driven by inattention. Figure 1 shows the geographic variation in the share of HN doctors who are also in the HNBG network. In our setting, we expect this option value to be quite small. Even before the introduction of the HNBG plan, we rarely observe individuals exercising the option of switching within the HN network in our dataset, even though they can freely do so before For example, almost no employees with HN who had PCPs affiliated with UC medical centers also saw a physician from Scripps Health in the same year. This is partly because few doctors make referrals outside of their narrow provider network and partly because individuals seldom change PCPs outside of the open enrollment period. In our sample, fewer than three percent of individuals ever visit a doctor who is not in the same network as their PCP. 17 Employees who place the highest value on being able to choose different doctors have already self-selected into a PPO, since PPOs impose the fewest restrictions on doctor choice. 14

16 5. Descriptive Evidence Table 5 shows the 2011 and 2012 health plan choices of employees in our estimation sample. All of these employees were enrolled in Health Net in We show the overall distribution of choices in the first column. The second column shows the choices for employees who have at least one family member whose 2010 PCP is not in the HNBG network. The third column shows decisions of employees for whom all of their family s existing PCPs were included in the HNBG network. We find that health plan choices differ strongly depending on whether the employee s (and her family members ) existing doctors were in the HNBG network. More than half of the employees whose doctors are not in the HNBG network stay in the Health Net plan in the first year, while less than a third switch to the cheaper HNBG insurance. There is some additional switching away from Health Net in the second year, but 45 percent of these employees are still in the substantially more expensive Health Net insurance in the second year. Employees who choose neither HN nor HNBG mostly switch to other HMO plans which will also force them to find a new doctor. Among employees who can switch to HNBG and keep their existing doctor(s), on the other hand, almost more than 70 percent do so in the first year and this share increases to more than 76 percent in the second year. Only 22 percent stay in Health Net in the first year and this share falls to about 15 percent in the second year. The gap in switching rates to HNBG between employees whose existing doctors are or are not in the HNBG network is 40.4 percentage points in the first year. It falls slightly to 38 percentage points in the second year. The data in Table 5 give us an unconditional estimate of how much inertia exists in our setting by showing how many employees could have saved a substantial amount of money on their health care premiums while keeping their doctor. In this first year, this applies to 22.4 percent of the sample (or 4117 employees) and in the second year it falls to 15.5 percent (or 2,845 employees). Of the 4,117 employees who did not switch to HNBG in the first year, 1,847 had no family members insured, 515 insured themselves and children, 581 insured themselves and a spouse, and 1,174 insured a spouse and children. In total, these 4,117 employees paid an additional $4.7 million in health care premiums during 2011 because they did not enroll in HNBG. Moreover, in 30 percent of these cases not a single covered family member visited a doctor in 2011 and 13 percent only had a single doctor visit for the whole family. 15

17 Table 6 breaks out this unconditional measure of inertia by insured family members, gender and income. In addition, we also report the share of each group who are willing to change their doctor in order to save on health insurance premiums. A number of patterns emerge from this table. First, employees with children appear to be more attentive than those without children. Second, male employees appear to be less attentive than female employees. However, among those who are attentive, female employees are less often willing to switch to a new doctor in order to save on their health care premiums. Finally, inertia increases with income and so does the willingness to pay higher premiums in order to keep one s doctor. Thus, higher-income employees are more likely to stay with HN than to switch to HNBG. While these patterns are suggestive, they represent unconditional means. In the regression analysis that follows, we control for a number of covariates including demographic and geographic controls. 6. Regression Analysis 6.1 Control Variables and Summary Statistics Table 7 presents summary statistics for the control variables that we include in our regression analysis. These values are for the year 2010 because we will test how the employee s 2010 characteristics affect their health insurance choices in 2011 and Our main variable of interest is the dummy variable Switch, which is equal to 1 if one or more of the employee s family members have a 2010 PCP who is not included in the HNBG network. This is true for 29 percent of our sample. Another key variable of interest is the employee s income. The mean income in our sample is 64,100 per year, with a standard deviation of 28,800. Recall that we dropped incomes below 25,000 and above 200,000. About 16 percent of our observations are faculty and the other 84 percent are staff. The mean age in our sample is 44, and 30 percent of employees are older than percent are male. 13 percent have a spouse but no children insured, 16 percent have children but no spouse insured and 34 percent have a spouse and children insured. On average, there were 10 doctor visits per family in 2010, but 15 percent of families had no doctor visits at all. 31 percent of families had between one and five visits, 21 percent had between six and ten visits and 33 percent had more than ten visits. When we look at doctor visits over the past three years (

18 2010), we see that the average number of visits per family was 19. Still, 11 percent of families did not have a single doctor visit over the past three years. 6.2 Regression Results We estimate multinomial logit models of the insurance choices in 2011 and 2012 for employees who were enrolled in Health Net in We allow for four choices: (i) stay with Health Net, (ii) switch to HNBG, (iii) switch to another insurance that is a PPO, and (iv) switch to another HMO insurance. The level of observation is the employee and our regression sample is as described above. We test for Independence of Irrelevant Alternatives (IIA) using a Hausman test and find that IIA is not rejected. In Table 8, we report results for the employee s health insurance decision in 2011, the first year that HNBG became an option. We show results from three separate specifications. We report marginal effects for continuous variables. For dummy variables, we report the change in choice probability as the value of the dummy changes from zero to one. The base category in our multinomial logit is the choice to stay in Health Net in We include fixed effects for the employee s 3-digit zip code in all regressions. These fixed effects control for geographic variation in average income and household composition, as well as availability and quality of doctors. In the first specification (Columns 1-3), we estimate the effect of our main variable of interest, Switch without any other controls except for the 3-digit zip code fixed effects. We find that having existing doctors outside the HNBG network reduces the likelihood of choosing the HNBG insurance plan by 35.4 percentage points (Column 1). The effect is similar to what emerged from the descriptive data in Table 5. The results from our first regression specification also show that employees whose existing PCPs are not all in the HNBG network are more likely to switch to other non-health Net insurance plans (Columns 2 and 3). The effects are estimated to be 1.5 percentage points and 4 percentage points, respectively. In our second specification (Columns 4-6), we add controls for income. Specifically, we compute the natural logarithm of the employee s salary, and then we demean this variable. We do this because it simplifies the interpretation of the other covariates in the regression. We also include and interaction of the income variable with Switch. Because we have demeaned logged income, we can interpret the coefficients on the dummy for having HNBG doctors as the effects 17

19 for an employee with mean income. This makes it easier to compare these coefficients to the previous specification. In fact, we find almost no change in these coefficients after we add the controls for income. For employees with HNBG doctors, we find that logged income has no statistically significant effect on switching to the HNBG insurance plan (Column 4). Since the primary reason why this group would not switch to HNBG insurance is inertia, we interpret this as inertia not varying significantly with income at least in this specification without additional demographic controls. For the interaction of Switch with our income variable, we find a statistically significant effect on the likelihood of switching to HNBG of This means that doubling an employee s salary would reduce the likelihood that this employee switches from Health Net to HNBG insurance by 7.2 percentage points. Given that, on average, only about 31 percent of these employees switch to HNBG insurance, this is a sizable effect. The other results for our second specification show how income affects switching to other insurance plans that are PPO s (Column 5) and HMO s (Column 6). We find that the effect of income on switching to PPO s is positive and statistically significant, but there is no difference for employees with and without HNBG doctors. This implies that doubling an employee s income would increase switching to PPO plans by 1.7 percentage points. Since the overall rate of switching to PPO s is 1.9 percent, this suggests that income is a very important driver of the choice to enroll in a PPO. Given the relatively high premiums and co-pays associated with the PPO options, this is not surprising. The fact that the effect of income is the same for employees with and without HNBG doctors suggests that the decision to enroll in a PPO plan is not driven by the desire to keep one s primary care doctor but by the other aspects of PPO plans, such as an even wider network of doctors or the ability to see a specialist without being referred by a primary care physician. The results in Column 6 show that the propensity to enroll in other HMO s declines with income. This is not surprising since the other HMO s (Kaiser Permanente and Western Health Advantage) have lower premiums than HNBG and are thus attractive options for families who want to save on their health insurance. Again, the effects are large relative to the overall rate of switching to HMO s. The interaction effect of Switch and income is positive. This suggests that employees who could keep their existing doctors when enrolling in the HNBG plan are less 18

20 likely to switch to the even cheaper HMO s as their income increases, compared to employees who would have to switch doctors to enroll in HNBG as well as in another HMO. Our final specification in Table 8 adds additional demographic controls (Columns 7-9). Each variable is interacted with Switch. Because we add a large number of interacted explanatory variables, the coefficient on the direct effect of Switch is no longer directly comparable to the two previous specifications. With the additional covariates, the direct effect of logged income remains statistically insignificant. This suggests that in our sample inertia is not affected by income. For families with some non-hnbg doctors, the effect of logged income falls in magnitude to Assuming that conditional on covariates the level of inertia is the same for both groups of employees, we would infer that the willingness to pay higher premiums in order to keep all of the family s existing doctors increases with income. This effect would reduce the rate of switching to the HNBG insurance from employees without HNBG doctors by 6.2 percentage points as the employee s income is doubled. The next covariate in this specification is a dummy for whether the employee is at least 50 years old. We find that employees with HNBG doctors are more likely to switch to HNBG insurance if they are in this age category suggesting that inertia is lower for older employees. This may be due to the fact that they tend to have more interactions with the health care system and thus are better informed about their doctor and/or their health care costs. Older employees with some non-hnbg doctors, however, are less likely to switch to HNBG insurance than younger employees. This implies that, while the older employees are more attentive and thus likely more aware of the HNBG option, they are less likely to be willing to change their doctor in exchange for a lower premium. This is therefore one of the groups that would suffer a greater welfare loss if they were forced into a smaller network that did not include their existing doctor. The magnitude of the coefficients implies that among older employees the propensity to enroll in the HNBG insurance is 8.3 percentage points higher if the family s existing doctors are in the HNBG network than when they are not. Thus, the effect of being in the older age group is larger than the effect of doubling the employee s income. For the two other categories ( other PPO and other HMO ), we find that older employees are less likely to enroll in either of these. This is true whether or not their existing doctors are in the HNBG network. 19