The Effect of Insurance Coverage on Preventive Care

Transcription

1 The Effect of Insurance Coverage on Preventive Care Marika Cabral and Mark Cullen March 25, 2011 Abstract Although health insurance products that preferentially price preventive care are growing in popularity, little is known about how preventive care utilization responds to changes in health insurance coverage. Using health insurance claims data from a large company, this paper examines the implementation of an insurance benefit design which differentially increased the marginal price of curative care (non-preventive care) while decreasing the marginal price of prevention. We examine the effect of the differential price change on the use of preventive procedures. In addition, we reveal evidence consistent with an important negative cross-price effect; that is, increases in the price of curative care can depress preventive care utilization. Keywords: health insurance, preventive health care, screening, salience, supplier-induced demand We thank Doug Bernheim, Jay Bhattacharya, Tim Bresnahan, Liran Einav, Peter Hansen, Caroline Hoxby, and Stanford seminar participants for their useful comments. Cabral: Department of Economics, Stanford University 579 Serra Mall Stanford, CA , mcabral@stanford.edu. Cullen: Department of Internal Medicine, Stanford School of Medicine. Cabral gratefully acknowledges support from the Stanford Economics Department and the National Science Foundation. 1

2 1 Introduction One of the major policy challenges in the United States is reducing rising medical costs. Some believe that cost-cutting objectives can be met by increasing patient marginal prices for medical services. Increasing in popularity are health insurance plans that require high patient marginal contributions for most health care procedures but require low or no patient contributions for certain preventive services. The basic idea behind such an insurance design is that high marginal prices for most services will discourage the use of non-essential services while lower preventive care prices will promote health maintenance activities. There are two natural policy questions related to this type of plan design and its potential to cut costs while promoting certain procedures: 1. Which preventive procedures should be encouraged? and 2. Is preferential insurance coverage an effective method to promote usage of these services? In our view, the first question is very important, and increasingly so given the recent debate over the appropriate preventive screening guidelines for breast cancer and prostate cancer. In this paper, however, our investigation is limited to the second question, leaving the first question the subject of other research. We use medical claims data from a large manufacturing company to explore how preventive care behavior responded to an insurance coverage change which increased patient contributions for curative care (all care other than preventive care) while simultaneously decreasing contributions for prevention. We find that increasing curative care prices depresses not only curative care spending but can also discourage the use of free preventive care services. Although traditionally health insurance plans covered curative care and preventive care equally, plans that have different cost-sharing rules for preventive care and curative care, similar to the plans introduced by the company studied, are growing in popularity. The Kaiser Family Foundation reports that in 2007, 66 percent of people on modest deductible, employer-provided PPO plans had a preventive care cost-sharing exemption, and 86 percent of people on employer-provided High Deductible Health Plans had a plan with a preventive care cost-sharing exemption. In 2009, over 90 percent of the people enrolled in employer-provided health insurance plans had preferential coverage of preventive care according to the Kaiser Family Foundation. The benefit change analyzed in this paper is a case study of a company that moved from uniform pricing of curative and preventive care to differential pricing. The increasing popularity of insurance options that differentially price curative and preventive care poses the obvious question to researchers: using a traditional non-differentially priced insurance plan as a starting point, can we increase marginal patient contributions for curative care while encouraging preventive care usage? From the perspective of policymakers, this is a particularly important question as the answer could guide them in their goal of reducing medical costs while promoting prevention. Our results can be interpreted as the effect of such an insurance coverage change on preventive care usage. The company we study has approximately 48,000 employees across the United States. The company introduced new health insurance plans to its employees beginning in The plans available to company employees before this change priced preventive care and curative care at a uniform and low level. The new 2

3 insurance plans required greater patient contributions for curative care while exempting preventive care from any patient charges. The net effect of the benefit change was that the price of a preventive care procedure was reduced from $10-15 to zero while the price of curative care rose substantially through increases in coinsurance payments and the introduction of deductibles up to $1,500. We investigate the effect of the insurance benefit change on preventive care usage rates. Four preventive services are examined: cervical cancer screenings (papanicolaou tests), breast cancer screenings (mammograms), prostate cancer screenings, and child immunizations. We find evidence that the rural and nonrual company populations differed in their response to the benefit change. Among rural enrollees, there was nearly a 4 percentage point reduction in annual cervical cancer screenings and a 6 percentage point reduction in annual early child immunizations. In contrast, nonrural enrollees increased their use of breast cancer screenings by 4.5 percentage points and child immunizations increased by more than 5 percentage points. Prostate cancer screening rates were unaltered by the company benefit change. A central result of our analysis is that rural women and children decreased their use of preventive services despite the fact that the benefit change decreased the price of prevention. We also show that rural enrollees cut back on curative care visits more so than nonrural enrollees in response to the benefit change. One potential explanation for these findings is that there is an important cross-price effect in the sense that the increase in the price of curative care depressed preventive care usage. We investigate the relationship between preventive care and curative care (and the potential for an important cross-price effect) in more depth through a few ancillary empirical tests. The results of these tests indicate that doctor advice during curative care visits may play a meaningful role in reminding and informing patients of recommended preventive care procedures. This channel may explain why policies aimed at discouraging more discretionary curative care visits, like the benefit change studied presently, may have the unintended consequence of discouraging subsidized prevention. 1 This evidence suggests that employers and policymakers may want to consider supplementing doctor-provided preventive care information and reminders when implementing differential pricing policies aimed at discouraging curative care spending. The remainder of the paper proceeds as follows. In Section 2, we describe the expected effect of a differential price change on preventive care usage. Section 3 describes the related literature. In Section 4, we describe the data and the environment. In Section 5, we analyze the effect of the firm s differential price change on the use of preventive care services. Section 6 explores the relationship between curative and preventive care services. Lastly, we conclude in Section 7. 1 In the context studied in this paper, the decline in curative care visits is more pronounced in the rural population suggesting that this cross-price effect may be larger among rural enrollees than nonrural enrollees (perhaps because even after controlling for the observables in the data rural enrollees are likely to be lower educated and earn less income). 3

4 2 Expected Effect of Differential Price Change To understand the expected effect of the company s differential price change, it is important to explore the relationship between preventive and curative care. If one views preventive and curative care as unrelated, then a change in the price of curative care would not affect preventive care usage (ignoring income effects). Under this assumption, the company s differential price change would unambiguously encourage the use of preventive services. This view is perhaps unrealistic because there are a number of reasons why preventive and curative care are related. Before continuing, we will define a useful distinction commonly made between two types of prevention: primary prevention aims to reduce disease incidence (for example, flu shots) while secondary prevention aims to mitigate consequences given a disease will occur (for example, cancer screenings). 2 Both types of prevention are fundamentally related to curative care. Primary prevention is clearly related to curative care as it is done to prevent future curative care usage. On the other hand, secondary prevention is mechanically related to curative care because curative follow-up procedures are often ordered when secondary preventive screenings yield positive test results. Although it may be clear that preventive care and curative care are related, how exactly would the curative care price change affect preventive care usage? Below we outline some channels through which the change in the price of curative care could affect preventive care usage. First, more generous coverage of curative care may deter investment in prevention (Ehrlic & Becker 1972), a concept termed ex ante moral hazard in the literature. When considering a short term change in curative care coverage, ex ante moral hazard most cleanly applies to primary prevention with short run health consequences (for example, flu shots). It is unclear how secondary prevention such as cancer screenings should respond to a short term change in curative coverage because screenings can lead to increased curative care expenditures in the short term in order to avoid more serious (and potentially more costly) curative care sometime in the future. Even though ex ante moral hazard has raised a reasonable amount of theoretical interest, there is little empirical evidence that supports the concept (for a review, see Zweifel & Manning 2000). Since the concept most directly applies to primary prevention, ex ante moral hazard is less applicable to the three cancer screenings we examine in this paper though it is potentially important for child immunizations, the only primary preventive procedure studied in this paper. Second, there may be an indirect effect if doctor advocacy is an important influence on preventive care usage. An increase in the patient cost for curative care generally discourages curative doctor visits. Patients that rationally reduce their use of curative care would interact less frequently with their doctors. If doctor advice and reminders play a central role in preventive care decisions, we might see preventive care usage decline in response to the benefit change studied in this paper. 3 Throughout this paper this effect will be referred to as the doctor interaction effect. Some empirical tests reported in Section 6 present evidence on the presence 2 See Kenkel (2000) for further examples and a more detailed discussion of this classification. 3 Some would call this sort of effect doctor-induced demand while others may interpret this as doctors following best practices. We don t take a stand in this paper as to whether this effect is desirable or not; we just note that this effect may be important when evaluating a differential price change. 4

5 of the doctor interaction effect. Third, imperfect salience of a preventive care cost-sharing exemption may lead people to cut back on prevention when the cost of curative care increases. The preventive care cost-sharing exemption was probably not the most salient feature of the company s benefit change. The majority of care is curative care, and the price of this care increased substantially. If the curative care price increase was the only salient feature of the benefit change, then imperfectly informed enrollees may have believed incorrectly that the price increased for preventive care and rationally reduced preventive care usage as a result. In summary, the doctor interaction effect and imperfect benefit change salience are justifications for a negative cross-price elasticity of preventive care with respect to the price of curative care. In the context of the benefit change and its effect on child immunizations, ex ante moral hazard can be interpreted as an argument for a positive (or less negative) cross-price elasticity of child immunizations with respect to curative care prices. Because the company changed the prices of preventive care and curative care in opposite directions, the effect of the policy change on preventive care usage can be thought of as a combination of two opposing effects: the own-price effect (positive) and the cross-price effect (probably negative). Thus, the net expected effect of the policy change on preventive care usage is ex ante ambiguous. 3 Relation to Literature A number of papers have studied the sensitivity of preventive care behavior with respect to changes in insurance coverage. The gold standard in this literature are those papers based on the Rand Health Insurance Experiment, a large-scale experiment in the 1970s in which families were randomly assigned to health insurance coverage of varying generosity. The experimental design allows researchers to test how preventive care usage varied with insurance coverage through random assignment of insurance plans. While some families were assigned to plans with patient coinsurance requirements of 25 percent or more, other families were assigned to the free care plans that required no patient contribution. According to Newhouse (1993), usage of preventive health services was 7 percent lower for women and 4 percent lower for men in the co-insurance plans as opposed to the free care plan. The difference was larger, approximately 12 percent for women and 10 percent for men, between preventive care usage of those assigned to the free care plan and those assigned to plans with deductibles. Lillard et. al. (1986) report results form the Rand experiment on the differences in preventive and non-preventive usage among those with different insurance coverage. The authors find that the reduction in preventive care usage is smaller than the reduction in non-preventive usage moving from the free plan to the plans with 25 percent or 95 percent coinsurance requirements. Keeler and Rolph (1988) estimate the pure price elasticities for different types of care in the Rand Health Insurance Experiment. They find that the elasticity for well care is a bit lower or a bit higher than for overall care depending on the range of plans one is comparing. They also report that the proportion of episodes that are for well care is not related to insurance coverage, which suggests that preventive and non-preventive care have similar price 5

6 responses. There are several reasons why estimates from the Rand experiment should be interpreted with caution in the context of modern health insurance settings. Cancer screenings are an important form of preventive care, and new innovations in cancer treatments since the Rand experiment have probably affected attitudes toward cancer screenings. Additionally, the Rand plans were likely more salient to experimental subjects than plan details typically are in the context of employer provided health insurance. Unlike the situation studied in this paper, the Rand insurance plans all priced preventive care and curative care the same in terms of insurance contributions. The Rand experiment provides the opportunity to investigate a simultaneous price change of preventive and curative care in the same direction, while this paper examines a price change in opposite directions. The uniform pricing of preventive and curative care in the Rand plans means there was little ambiguity in the cost faced by the experimental subjects for a doctor visit. Additionally, the uniform pricing Rand plans might have encouraged more use of services among those on the free plan thus facilitating more interaction with doctors among these subjects. If we think doctors play a large role in supplying information to patients and reminding patients to do preventive care, the differences in curative care coverage among the Rand plans could have been driving the preventive care results of the Rand experiment. For these reasons, in the Rand experiment the own- and cross-price effects most likely operate in the same direction to encourage more preventive care usage in the free care plans. In the natural experiment studied in this paper, on the other hand, the own- and cross-price effects most likely operate in opposing directions. Thus, the Rand estimates can be viewed as an upper bound on the expected effect on preventive care usage from an insurance policy change that differentially affects curative and preventive care patient prices. There are several more recent studies inspecting the correlation between insurance coverage for prevention and preventive care utilization. Some studies have found that increased cost-sharing is associated with a lower probability of receiving cancer screenings (Friedman et al. 2002; Varghese et al. 2005; Trivedi, Rakowski, and Ayanian 2008; McWilliams et al. 2003), while other studies found no association between cost-sharing and use of preventive care (Cherkin, Grothaus, and Wagner 1990; Tye et al. 2004; Rowe et al. 2008). The majority of these papers simply describe either a correlation based on self-selected groups with varying insurance coverage or a correlation based on constructed comparison groups with varying coverage. 4 In contrast, our paper takes advantage of staggered and plausibly exogenous variation in insurance coverage over time within one company which allows us to cleanly identify the effect of a differential price change on prevention. This natural experiment allows us to examine the enrollee preventive care response to a simultaneous increase in the price of curative care and decline in the price of prevention. Busch et al. (2006) use a subset of the data employed in this paper to investigate the immediate effect of the 4 There are a few exceptions which rely on variation over time. McWilliams et al. (2003) inspect variation in coverage for cholesterol testing, mammography, and prostate examination over time at the age 65 Medicare threshold. The authors look at the effect of increasing coverage for all services on prevention as the coverage for other services (including curative care) also becomes more generous at this threshold. Cherkin, Grothaus, and Wagner (1990) exploit variation in co-payments within one HMO over time. Specifically, they look at the effect of introducing a $5 co-pay for all services (curative and preventive) on the usage of preventive care services. 6

7 benefit change on the use of some preventive care services. The authors present a comparison of means using the two years of data available at the time of their study, and they find no evidence that the policy change affected patient preventive care usage. The study summarized in the present paper extends this analysis in a number of ways. First, the analysis in this paper fully utilizes the staggered natural experiment in the data by using data over five years. Because the company s insurance benefit change was rolled out over a number of years to different company employees, the longer data set allows us to take advantage of this variation. Second, we are able to apply methods that utilize within enrollee information over time using the longer data. This is especially important because the cancer screenings studied are not typically done annually making it critical to have more than two years of data to estimate changes in behavior. Third, we are able to investigate some of the channels through which the benefit change influenced prevention through examining the timing of care. With this richer data, we find significant and economically meaningful heterogeneity of reactions to the benefit change between rural and nonrural enrollees. In addition, we reveal evidence consistent with a meaningful cross-price effect, a channel through which decreasing the generosity of curative care coverage may depress prevention. Despite theoretical interest in the link between curative and preventive care (i.e., Ehrlic & Becker 1972, Zweifel & Manning 2000), there is little empirical evidence on the effect of curative care prices on preventive care behavior. Understanding this link is important for understanding how one might encourage prevention through price and non-price mechanisms. A recent paper by Meeker et al. (2010) finds evidence indicating that reducing the price of preventive services to zero encourages screening among PPO enrollees, across several categories, but this effect is significantly attenuated among individuals that opt for high-deductible plans. The correlation these authors find between the price of curative care (in the form of an annual deductible) and the reaction to a preventive care cost-sharing exemption is consistent with the evidence we find for a negative cross-price elasticity of prevention with respect to the price of curative care. While many studies on preventive care behavior rely on survey data, this paper uses medical insurance claims information from a large manufacturing company. One advantage of using insurance claims data compared to using survey data is that the information in claims data is more detailed and accurate. Another advantage is that the claims data contain information on the same people over a number of years. 5 This allows one to study persistence in preventive care behavior and changes in behavior over time. The drawback of using the company claims data is that the applicability of the estimates outside the company population may be limited. Although the hourly employee population is not nationally representative, the health care consumption of the types of individuals in the hourly company population may be of particular interest to policymakers. In many studies related to consumer protection, researchers want to over-sample the marginal or vulnerable population. Among those with medical insurance in the US, individuals in the hourly company population are relatively marginal in that they are less educated, have less access to care, live in more rural locations, and earn less money than an average insured American. The results of this paper suggest that the differential 5 Over a number of years, we observe the subset of employees (and their associated spouses and dependents) that remain with the company for some time. 7

8 price change had a significant effect on preventive care usage, at least within this marginal population. 4 Background We use data on employees of a large self-insured manufacturing company from 2003 to The company population includes approximately 48,000 employees from numerous work sites across the US. For each worker, the data include information on wages, company tenure, type of job (hourly or salary), age, sex, location, chosen health insurance plan, and medical claims data. The employee population is divided into benefit groups that reflect specifics of the company s business model. Based on benefit group divisions, the company assigned each worker a menu from which to choose a health insurance plan. Each worker chose from this menu of health insurance contracts, and employee decisions and option menus are reflected in the data. Very few employees chose to opt out of health care coverage all together. In addition, each employee could select to insure his/her spouse and children through the various family option pricing offered by the company. When an employee chooses to enroll family members, the data include the age, sex, and medical claims information for these family members. The medical claims data offer a detailed look at the health care behavior of the individuals. For each claim, the data reflect the date of the service, the billed total cost of the service, the out-of-pocket cost of the service, the type of service, the type of facility in which the service was performed, and the specialty of the medical professional that delivered the service. Descriptions of the services vary in the level of specificity. In this paper, we examine the preventive care services that we can unambiguously identify. 4.1 Description of Differential Price Change Prior to 2004, a subset of company employees were offered a standardized menu of options for health insurance which we will refer to as the old menu. The company began to replace the old menu of health insurance plans with a new menu of plans starting in The benefit design change was rolled out to enrollees over a number of years due to staggered expiration dates of union contracts. Both before and after the employees were treated with the insurance benefit change, employees selected health insurance plans from a menu of health insurance policies. It is important to note that we use the exogenous switch from the old menu to the new menu and the exogenous staggered timing of the introduction of the new menu to identify the effect of the policy change on prevention; we avoid using the employees endogenous plan selections within the old and new insurance menus. The plans on the old menu required $10-15 co-pays for both curative and preventive doctor visits. Ninetyeight percent of people on the old menu plans faced no deductible. The new plans differed in their level 6 Some of the company s business divisions had different menus of health insurance contracts and were not eligible for the switch to the new menu contracts. The differences in benefits packages reflect aspects of the subsidiary business model of the company and appear to be uncorrelated with health care utilization (conditional on observed information). We focus on the subset of employees that were undergoing the benefit change from the old menu plans to the new menu plans. 8

9 of cost-sharing for both preventive and curative care. Curative care on the new plans was subject to a deductible and coinsurance. The patients were responsible for 10 percent of charges beyond the deductible, while deductibles ranged from $0 to $1,500. Preventive care was free of charge on the new plans. The health insurance employee brochure highlighted the specific preventive services that were free of cost on the new menu plans. The net effect of the change on the patient cost of prevention was a decline of $ Generally, the cost of curative care increased for people as a result of benefit change. For most services, the 10 percent coinsurance payment alone was larger than the $10-15 co-pays that people paid on the old plans. Sizable deductibles many faced on the new menu plans increased curative prices further. The empirical results capture the effect of moving from uniform pricing of preventive and curative care at quite low levels to differentially increasing the price of curative care while exempting prevention from any financial cost. There are a few additional details about the benefit change that are worth noting. First, the new and old menu plans differed only in their cost-sharing terms and pricing; plans on both menus used the same provider network. Second, the company increased employee contributions to medical insurance premiums with the transition to the new menu plans. Because of this sizable increase in employee premiums, it is likely that employees understood there was a benefit change even if they did not know the details of the change Description of Preventive Procedures Preventive care describes services ranging from nutrition counseling to mammograms to diabetes monitoring. 9 Many primary prevention efforts are simply lifestyle choices that are not recorded by insurance companies. Although some of the most effective preventive measures are not recorded by insurance companies, we are limited to investigating prevention that can be inferred from insurance claims data. However, it should be noted that prevention that is recorded by insurance companies is not only much easier to identify in claims data but also likely much easier to affect through policy than preventive lifestyle choices. This empirical study focuses on a few specific procedures: cervical cancer screening (the papanicolaou test, more commonly the pap smear test), breast cancer screening (mammography), and prostate cancer screening. 7 There is one exception to this treatment: under the old menu cost-sharing for mammograms differed depending on the facility which preformed the screening; at some facilities mammograms had no patient co-pay requirement, while at others there was a $10-15 co-pay. The new menu exempt mammograms, as with other preventive services, from any patient charges. In this sense, the price drop for mammograms was smaller than the price drop for the other preventive care procedures. We might expect the cross price effect to be more dominant in the mammogram analysis since the own price change is smaller. 8 The large increase in employee premiums may indicate that there is an relevant income effect, though it seems the income effect in this case would be second order. Suppose the benefit change is perfectly salient, and preventive care is a normal good. In this case, the income effect would only cause a reduction in preventive services if preventive services were costly after the benefit change. Thus, if the benefit change is salient, then there is no direct income effect on prevention. On the other hand, if the benefit change is not salient, then the income effect could influence prevention. However, this is not a traditional income effect because its presence is due to lack of salience of the price change. In other words, the income effect would amplify the cross-price effect, depressing prevention. 9 According to Coffield et al. (2001), some preventive measures that are most valuable and under-used are: tobacco cessation counseling, vision screening for older adults, counseling young adults about alcohol and drug use, alcohol abuse screening, vaccinating older adults against pneumococcal disease, and screening young women for chlamydial infection. 9

10 Additionally, we examine annual child immunization rates. 10 According to the company s health insurance documentation, all of these services were exempt from payment under the new plans. We focus our empirical investigation on individuals that are eligible for the procedures above. For cervical cancer screening, we define the eligible group as women age 21 and over. Women over the age of 40 are defined as eligible for breast cancer screenings, and men over the age of 50 are eligible for prostate cancer screening for the purposes of this paper. Our examination of child immunizations is limited to children age 3 and younger. Details on the eligibility definition choices for these procedures are contained in Appendix A. 4.3 Description of Samples for Estimation We restrict attention to the company enrollees subject to the benefit change described above. 11 The sample used in the estimation is compared to the total company population in Table 1. Moving from left to right in Table 1, one can see how the sample size decreases with further restrictions. Approximately 29,700 employees had the relevant benefit design. When the company adopted the new menu of plans, it would have liked to switch all 29,700 employees to the new menu in However, because of staggered expiration dates of union labor contracts, the company was only able to switch a subset of these employees (including all non-unionized salary workers) in the first year of implementation. As a result, all salary workers were treated in 2004, while the treatment of hourly workers was staggered from 2004 onward. To keep treatment and control groups as homogeneous as possible, we make the further sample restriction of looking only at hourly employees and their relations. This group is displayed in column 3 of Table 1. Notice this is an unbalanced panel because it contains varying numbers of people across years due to new hires, retirees, and job leavers. We do not know the prior (future) insurance coverage status for those who began at (or parted with) the company during the sample period. The company is quite generous with medical care coverage, both before and after the benefit change, and estimates using the unbalanced panel may be biased. This is of particular concern in this setting since the studied preventive care services are highly optional, meaning some of these services may easily be re-timed. Thus, we further restrict the sample to the balanced panel described in column 4. This sample contains only people that were enrolled in either the old or new plans during the whole period from 2003 to The balanced panel described in column 4 is used to investigate the effect of the benefit change on cancer screenings. However, to study early child immunizations, we employ an unbalanced panel of children 3 years of age and younger Our measure of child immunizations is based on the immunizations that are easily identified in the claims data. We have reason to believe this underestimates actual immunizations children receive for at least two reasons: children may receive immunizations outside the normal medical system and it seems that some doctor visits that included immunizations did not separately bill an identifiable immunization code. However, we have no reason to believe that this underestimation varied systematically with the staggered benefit change so our estimation method is unaffected. This is discussed at length in Appendix A. 11 For some company locations, employees were offered a HMO option in addition to the menus discussed above. The HMO plan was a fundamentally different sort of plan than the PPO-type plans offered on the old and new menus. Very little switching occurs in the data between the HMO option and the new or old menu plans, and there are no claims data available for those employees that select the HMO option. People who opted for the HMO plan during the time period studied are dropped from the data used for estimation. 12 The unbalanced panel is necessary because we only look at children under 4 years old and the time period we consider is 5 years long. The unbalanced panel of small children is less problematic than an unbalanced panel of adults for cancer screening 10

11 Table 2 summarizes some characteristics of the hourly balanced sample. One can see that company s hourly employees earn less money than the average worker in the US, and about half the hourly enrollee sample live in rural areas. The table is organized by the year of introduction of the new plans with the last column representing the untreated group which consists of people for which the new plans had not been introduced as of Because of the staggered introduction dates of the new plans, groups that were treated in later years will serve as controls in earlier years in the empirical estimation. For this reason, we would like the treatment groups to be as homogeneous as possible. There are a total of 14,217 people in the hourly sample including employees, spouses, and dependents. The treatment groups look relatively similar on many observable characteristics. One exception is that those treated in 2006 more often live in rural areas than those in the other treatment groups. We control for the observable differences across the groups in the empirical analysis, and we also show the empirical analysis is robust to excluding those treated in Table 3 describes the eligible sample used in the estimation by procedure type. 5 Analysis of Differential Price Change As discussed earlier, the company introduced new plans starting in 2004 which differentially changed the marginal prices for curative and preventive care. Taking advantage of the exogenous variation in the introduction of the new health plans, we use difference-in-difference regression analysis to identify the effect on annual usage rates of the three cancer screenings and child immunizations. In addition to pooled estimation, we repeat the analysis separately for rural and nonrural enrollees because of documented differences in health care delivery across these types of locations. 5.1 Annual Preventive Procedure Rates To investigate the impact of the policy change on annual preventive care usage rates, a difference-in-difference regression model is used. The effect of the benefit design change is estimated procedure by procedure using variations of the basic equation displayed below: procedure it = β o + β 1 treatment it + β 2 employee i + t α t year t + i δ i treatmentgroup i + γx it + ɛ it. (1) The observations used in estimation are at the individual-year level, where i denotes the individual and t denotes the year. The dependent variable, procedure it, is an indicator variable that equals to one when the relevant preventive procedure was done by individual i in year t. The variable treatment it is a binary variable that takes the value one when the individual was on the new menu plans. Employee i is simply an indicator variable that takes the value one when the individual was a company employee. Year, region, age and treatment group fixed effects are included in all specifications. Because we would expect there to be within-person correlation in the dependent variable, standard errors are clustered at the person level. The analysis because the cause for the unbalance in this sample is primarily age restrictions rather than parental job changes. 11

12 estimation is done on the balanced panel sample for the three cancer screenings described in Table 3, panels a-c. The unbalanced sample of young children described in Table 3 panel d is used to estimate the treatment effect on child immunizations. The OLS regression results are reported in Table Summary statistics of the dependent variables are given at the bottom of each panel in Table 4. The across-person and average within-person standard deviations for the cancer screening dependent variables are displayed. As expected we can see the within-person variation is smaller than the across person variation for all the screenings. However, notice that the within-person variation is not so close to zero. In words, people were not completely regular and persistent in their annual cancer screening behavior. Turning to the regression results in Table 4, we see that there were statistically and economically significant effects on preventive care usage in the rural and nonrural samples. To summarize the findings from this analysis, rural women and children responded to the benefit change by decreasing their use of preventive services, while nonrural women and children meaningfully increased their use of preventive services. The cervical cancer screening rate decreased up to 3.8 percentage points among the rural sample while remaining constant in the nonrural sample. Rural women did not statistically increase their use of breast cancer screenings, but nonrural women did increase their breast cancer screening rate 4.4 percentage points. Annual child immunization rates increased 5.7 percentage points among nonrural children while decreasing 5.9 percentage points among rural children. Prostate cancer screening rates were unresponsive to the differential price change. One limitation of the difference-in-difference empirical approach is that the observations are grouped into year-long intervals which discretizes data that is more continuous. Because the date of service is included in the claims data, the exact number of days between preventive procedures can be calculated. We attempt to address this shortcoming of the difference-in-difference regression analysis by estimating a duration model for the three cancer screenings. 14 Unfortunately, censoring issues in the duration framework put serious limits on the identification that is possible. Details and results from the duration model can be found in the Appendix B. 5.2 Robustness Checks Re-Timing of Preventive Care Re-timing of preventive care would potentially interfere with the identification strategy used in estimation. One could imagine that the benefit change could have caused a surge in usage of preventive services right before or right after the new menu introduction depending on the beliefs of the enrollees about the coverage changes. We find no evidence for any surge in usage for any of the four treatment groups. We find the 13 Probit and logit specifications (results not reported) have qualitatively similar results. 14 Analysis of early child immunizations is excluded because duration analysis seems inferior to cross-sectional analysis for child immunizations. Only a small fraction of the child sample would be usable in duration analysis because we need to rely on within-person variation in treatment status. 12

13 regression results are robust to repeating the analysis ignoring preventive procedures done in the December proceeding the benefit change or the January following the benefit change for each of the treatment groups Heterogeneity in Treatment Groups Inspecting Table 2, we can see that a lot fewer of those treated in 2006 live in rural areas compared to those in the remaining treatment groups. We may be worried that the results are sensitive to the inclusion of this different treatment group. To make sure the results are robust, we repeat the analysis excluding the 2006 treatment group, and we find qualitatively very similar results Correlation in Behavior Within Treatment Groups The standard errors in the baseline analysis are calculated allowing for within-individual correlation over time. One might be worried that preventive care behavior is correlated within treatment groups for reasons other than the treatment which may affect the estimates. In this company, the timing of the treatment and the premiums individuals face on the new and old menus are determined at the benefit group level. There are approximately 50 benefit groups in the relevant sample. We repeat the analysis clustering standard errors at the benefit group level. We see that the size of the standard errors increases when clustering at this higher level, as we would expect with so few clusters. Still, the main results remain fairly precisely estimated even with this level of clustering Relationship Between Curative and Preventive Care As alluded to before, there are a few reasons why we might expect preventive care usage to decline because of the policy change. Although we cannot separately identify how much the salience and doctor interaction effects drive the observed preventive care behavior, the data do allows us to investigate the doctor interaction hypothesis a bit further. In order for the doctor interaction effect to affect preventive screening rates through the benefit change, there must have been a decline in curative care visits because of the benefit change. Table 5 displays evidence about the effect of the benefit change on the annual number of curative doctor visits. 18 The odd columns show evidence from OLS regressions, while the even columns report coefficients from negative binomial regressions. The table shows that cutbacks in curative care visits are sizable. The number of curative visits significantly decreases approximately 7 percent in the rural sample. In the nonrural sample, we see a smaller decline, approximately 1 percent, and this effect is not statistically significant. The estimated coefficients and the 15 Results from this robustness check are displayed in Appendix C, Table C1. 16 The results from this robustness check are displayed in Table C2 of Appendix C. 17 See Appendix C, Table C3 for results. 18 The measure for curative care visits is simply the number of doctors visits that are not coded as one of the preventive services we study. Note that this measure over-counts the curative care doctors visits because we could not identify all the preventive care services in the claims data that qualified for the exemption. However, since the vast majority of care is curative, this probably leads to little noise in the estimation of the effect on curative care visits. 13

14 heterogeneity is consistent with the explanation that the doctor interaction effect played an important role in the observed heterogeneity in the preventive care usage changes. That is, the decline in curative care visits associated with the benefit change is more pronounced in the rural population suggesting that this cross-price effect may be larger among rural enrollees than nonrural enrollees (perhaps because even after controlling for the observables in the data rural enrollees are likely to be lower educated and earn less income). 19 We now turn to more direct evidence of the doctor interaction effect. The doctor interaction effect can be decomposed into short-term and long-term effects. A doctor s visit may serve as a reminder for a patient to return for a preventive screening thereby influencing the short-term behavior of a patient. On the other hand, interacting with a doctor on a regular basis may inform a patient of the merits of prevention, and this information can have a long-term effect on a patient s preventive care behavior. While both short-term and long-term effects may be important, the test we employ to investigate the doctor interaction effect is limited to identifying a short-term doctor interaction effect. To test for the doctor interaction effect without complications from reverse-causality, the effect of urgent curative care visits on cervical cancer screenings is examined. The reason we restrict attention to urgent curative visits is because more discretionary curative visits could be related to personal attributes that may be correlated with differences in preventive care usage unrelated to the doctor interaction effect. We focus on the sample eligible for cervical cancers screenings because it is easy to identify a common and urgent condition among this sample in the insurance claims data: urinary tract infections. Approximately 7 percent of the sample eligible for cervical cancer screenings has at least one urinary tract infection in the period we examine. Urinary tract infections generally require a visit to the doctor as treatment involves prescription antibiotics. Visits associated with urinary tract infections are a good source of conditional random variation that can be used to measure the effect of an urgent curative care visit on the probability of a subsequent cervical cancer screening. It should be highlighted that we flexibly control for individual average annual curative care visits in this analysis to separate out the effect of a recent urgent doctor visit for an infection from the person-specific component of preventive habits that may be associated with frequency of doctor visits. A Single-Spell Cox Proportional Hazard Model is used to test for the impact of urinary tract infection visits on the probability of having a cervical cancer screening soon after. To make the sample as homogeneous as possible, the eligible sample is restricted to those in the first (and largest) treatment group to switch over to the new plans. The spell examined is the time between an individual s latest pap test in 2003 until their next pap test, and individuals without a pap test in 2003 were assigned the start date of Jan 1, The results are reported in Table 6. Infection in previous month is an indicator variable that equals to one for 19 We don t observe the level of education or household income of employees in the data. Still, based on general education and income disparities between rural and nonrural areas, we find it reasonably likely that rural enrollees may be less educated and rural employees may be less likely to have a working spouse (and thus have lower household income) than their urban counterparts. 20 We choose this time period to examine because this sample has already made the transition to the new menu plans and so insurance coverage is constant from 2004 to

15 30 days after an office visit related to a urinary tract infection. In some specifications we flexibly control for an individual s average annual number of curative care doctors visits. The results in columns 1 and 2 indicate that there is a large and significant doctor reminder effect (a short run doctor interaction effect). In the specification which controls for average curative care visits displayed in column 2, we see that an urgent curative care visit triples the hazard rate for having a cervical cancer screening for the following month. 21 Since doctors likely have less time to chat with patients about preventive care during urgent visits than during non-urgent curative visits, the reported effect of urinary tract infection visits likely underestimates the effect of an average curative care visit on the probability of a subsequent cervical cancer screening. For the same reason, these results even further underestimate the effect of a marginal curative care visit as a marginal visit is presumably less urgent than the average visit. This evidence suggests that patients rely on reminders from their doctors in order to make preventive care decisions. Columns 3 and 4 of Table 6 present the results of a placebo test where we replace the variable Infection in previous month with Infection in next month, which indicates when an individual goes to the doctor within the next 30 days for a visit related to a urinary tract infection. With this different variable, we should find no significant effect if the results in columns 1 and 2 are driven by this doctor reminder effect (as opposed to some other unobservables correlated with infections). It is reassuring that we find no significant effects in these placebo specifications. 7 Conclusion This paper examines preventive care behavior of health insurance enrollees using data from a large manufacturing company. The company we study altered its employee health insurance by reducing the price of prevention to zero while increasing the price of curative care substantially. We measure the effect of the benefit change on preventive care usage rates. As discussed earlier, the expected effect on preventive care rates is ambiguous because the own- and cross-price effects of the benefit change probably operate in opposing directions. Within the nonrural sample, women increased their usage of breast cancer screenings and annual child immunization rates increased. In contrast, a statistically meaningful reduction in cervical cancer screenings and annual child immunization rates can be seen in the rural sample. These estimates indicate that the relative importance of the own- and cross-price effects may have varied between nonrural and rural samples. Although the claims data do not allow us to disentangle the effects of the preventive and curative price changes, we look deeper into the hypothesis that doctor advice and reminders play an important role in encouraging preventive care services (making this a channel through with curative price changes can affect preventive care usage). We find the benefit change depressed curative care doctor visitations, and this depression was more pronounced among rural enrollees. In addition, we also find suggestive evidence for the importance of doctor interactions in encouraging prevention through ancillary tests. The results of these 21 As expected, the estimated effect of a urinary tract infection visit is larger when we omit controls for average curative care visits. 15

16 tests suggest that doctors may play an important role in reminding and informing patients of preventive services. A policymaker hoping to encourage prevention, especially a policymaker that also wants to adopt differential pricing, may want to look toward other reforms to encourage prevention according to our results. Options that may be particularly promising include policies aimed at increasing the salience of a preventive care cost exemption and policies aimed at replacing or supplementing doctor-provided information and reminders. Potential policy options include increasing preventive care reminders, assembling preventive care information campaigns, offering preventive services in more convenient locations (at work, at the grocery store, etc.), improving physician preventive care tracking, and risk-rating premiums based on lifestyle choices and past preventive care usage. Future researchers should work to understand which methods, price and non-price, are effective in promoting prevention. References Busch, S., C. Barry, S. Vegso, J. Sindelar, and M. Cullen (2006): Effects of a Cost-Sharing Exemption on Use of Preventive Services at One Large Employer. Health Affairs, 25 (6), Cherkin, D., L. Grothaus, and E. Wagner (1990): The Effect of Office Visit Copayments on Preventive Care Services in an HMO. Inquiry, 27 Coffield, A., M. Maciosek, M. McGinnis, J. Harris, M. Caldwell, S. Teutsch, D. Atkins, J. Richland, and A. Haddix (2001): Priorities Among Recommended Clinical Preventive Services. American Journal of Preventive Medicine, 21(1), 1-9. Ehrlich, I. and G.S. Becker (1972): Market Insurance, Self-Insurance, and Self-Protection. Journal of Political Economy, 80 (4), Employer Health Benefits: 2007 Summary of Findings. The Kaiser Family Foundation and Health Research and Educational Trust. Friedman, C., F. Ahmed, A. Franks, T. Weatherup, M. Manning, A. Vance, and B. Thompson (2002): Medical Care, 40(11): Guide to Clinical Preventive Serivces 2009: Recommendations of the US Preventive Service Task Force, US Dept of Health and Human Services. Keeler, E. and J. Rolph (1988): The Demand for Episodes of Treatment in the Health Insurance Experiment. Journal of Health Economics, 7(4): Kenkel, D. (2000): Prevention. Handbook of Health Economics Vol. I, Chapter 31. Lillard, L., W. Manning, C. Peterson, N. Lurie, G. Goldberg, and C. Phelps (1986): Preventive Medical Care: Standards, Usage, and Efficacy. Rand Publication, R-3266-HCFA. McWilliams, J., A. Zaslavsky, E Meara, and J. Ayanian (2003): Impact of Medicare Coverage on Basic Clinical Services for Previously Uninsured Adults. Journal of the American Medical Association, 290 (6), Meeker, D., G. Joyce, J. Malkin, S. Teutsch, A. Haddix, and D. Goldman (2011): Health Services Research, 46 (1), Newhouse, J., (1993): Free for all?: Lessons from the RAND Health Insurance Experiment. Harvard University Press. 16

17 Parker-Pope, T. (2008): Panel Urges End To Prostate Screening at Age 75. New York Times, Aug Rowe, J., T. Brown-Stevenson, R. Downey, and J. Newhouse (2008): The Effect of Consumer-Directed Health Plans on the Use of Preventive and Chronic Illness Services. Health Affairs, 27 (1), Trivedi, A., W. Rakowski, and J. Ayanian (2008): Effect of Cost Sharing on Screening Mammography in Medicar Health Plans. New England Journal of Medicine, 358, Tye, S., K. Phillips, S.-Y. Liang, and J. Haas (2004): Moving Beyond the Typologies of Managed Care: The Example of Health Plan Predictors of Screening Mammography. Health Services Research, 39 (1), Varghese, R., C. Friedman, F. Ahmend, A. Franks, M. Manning, and L. Seeff (2005): Does Health Insurance Coverage of Office Visits Influence Colorectal Cancer Testing? Cancer Epidemiology, Biomakers, and Prevention, 14, Zweifel, P. and W. Manning (2000): Moral Hazard and Consumer Incentives in Health Care. Handbook of Health Economics, Vol. 1. A Appendix: Background on Preventive Care Procedures The US Preventive Services Task Force (USPSTF), a branch of the US Department of Health and Human Services, sets guidelines for preventive care procedures. According to discussions with medical professionals, these standards are commonly used by practitioners to make preventive care recommendations. According to the USPSTF, women age 21 through 64 should be tested regularly for cervical cancer by a pap smear test. For women below 30 years of age, it is recommended that they go to the doctor annually for a gynecological exam including a pap test. After having three consecutive normal pap smear tests, the guidelines indicate that a woman over 30 can elect to have pap tests every 2-3 years. Otherwise, women over 30 should continue to do annual pap tests. Because positive pap test results are quite common, many women over 30 years of age should continue to do annual screenings. The female company enrollees that were eligible for pap tests are summarized in Table 3 panel a. Prior to 2009, the USPSTF advised women over age 40 to be screened for breast cancer by mammography every 1-2 years. The USPSTF cited evidence that such screening significantly reduces mortality from breast cancer. Over the course of her lifetime, one in eight women will develop breast cancer. The risk of breast cancer increases significantly with age, and family history plays a large role in this type of cancer: percent of women with breast cancer have a close relative who had the disease. Although in November 2009 the USPSTF came out against routinely screening women below 50 years of age, the female population over 40 years of age is examined in this paper. Table 3 panel b describes women that were over 40 years of age in the company population. Men over the age of 50 may elect to be screened for prostate cancer. Prostate cancer screening is a controversial issue. Prostate cancer is the second leading cancer killer among men, behind only lung cancer. However, many studies have shown that prostate cancer is over-diagnosed in the sense that it is detected at a point when the disease is unlikely to affect life expectancy; this happens in 29 to 44 percent of diagnosed cases (Parker-Pope 2008). Prostate cancer is typically slow growing, and many men diagnosed with it will not see any symptoms during their lifetime, particularly if they are older. Diagnosis of and treatment for prostate cancer can significantly reduce a patient s quality of life due to psychological stress and side effects. In August 2008, the USPSTF has come out against screening men over age 75 for prostate cancer. The USPSTF claims that there is no benefit for testing men whose expected remaining lifespan is less than 10 years. The USPSTF has yet to take a stand on the testing of younger men. In the data used in this paper, most all of the men have more than 10 years of remaining life expectancy. As workers retire, they cycle off of the company health plans and onto Medicare. In fact, the average age for men over 50 in the company population is approximately 54 years of age. For the purposes of this paper, we define the eligible group for prostate screening as all men in the data that were over 50 years of age, and these men are described in Table 3 panel c. We also consider a more general category of procedures in the claims data: child immunizations. There are recommended immunizations at all ages, however most recommended immunizations are heavily concentrated in the early years of life. Here the analysis is limited to enrollees that were 3 years of age and 17

18 younger. The enrollee population 3 years of age and younger is described in Table 3 panel d. Identification of immunization claims in the data is more tricky than the identification of cancer screening claims. For the measure of immunizations used in this paper, we count all claims with the description specifying immunization. From inspecting doctor visits for young children, we believe that other claim descriptions were sometimes recorded for doctor visits that included immunizations. Thus, the measure of immunizations used in this paper can be viewed as a proxy that underestimates actual immunizations but should be highly correlated with actual immunizations. Another reason that the immunization measure may not capture all immunizations is that it is fairly common for children to receive immunizations outside the normal medical delivery system. Because noise in the immunization measure is unlikely to be correlated with the staggered introduction of the new menu insurance plans, the noise does not interfere with our analysis as the observed changes in the immunization measure can be interpreted as truly suggestive of changes in the actual child immunization rate even if the immunization measure may not be meaningful in level. B Appendix: Duration Analysis: Time Between Cancer Screenings We examine the time between procedures using a Single-Spell Cox Proportional Hazard Model to investigate how the probability of seeking a preventive procedure changes with the benefit change. To clarify the exposition of the duration analysis, some definitions are given. A failure is defined as having the cancer screening, and the initial state is the state that all people were in every day they are not getting a screening. 22 One common obstacle for duration analysis is censored data. In this situation, there is both right censoring (some unfinished spells) and left-censoring (incomplete information about the start of spells). While right-censoring is easy to deal with, left-censoring can more difficult to handle. The way researchers typically deal with leftcensoring is to assume a distribution of possible start times for the spells and integrate over this distribution in the likelihood function used for estimation. Because there is no straightforward way to use the available data to test which distributional assumption is the most sensible, we pick the most simple, impartial assumption and proceed with the analysis. The assumption used here is simply that everyone s first spell begins on the first observation day, Jan 1, To ensure our results are robust to other assumptions, we re-estimate the hazard model under an alternate left-censoring assumption: that start dates are uniformly distributed over the four years prior to We estimate a Single-Spell Cox Proportional Hazard model using the same eligible samples as were used in the difference-in-difference regression analysis for the three cancer screenings. The single spell considered is the spell beginning on Jan 1, 2003 and ending at the time of the first observed cancer screening. The results for these hazard ratio estimates are displayed in Table B1. The primary covariates of interest are the time-varying treatment term. The time-invariant covariates in the hazard model include age in 2003, an employee indicator, and treatment group indicators. Table B1: Hazard Estimates of Differential Price Change a: Cervical Cancer b. Breast Cancer c: Prostate Cancer All Rural Nonrural All Rural Nonrural All Rural Nonrural treat (0.084) (0.132) (0.111) (0.115) (0.192) (0.147) (0.274) (0.429) (0.363) Subjects % Fail 75.8% 76.8% 74.7% 73.5% 73.9% 73.2% 26.1% 27.7% 24.5% Note: A Cox Proportional Hazard Model is estimated on the duration between the start of the period and the observed preventive care visit. The start of the period is assumed to be January 1, 2003 for all subjects. Coefficients are displayed above with standard errors in parentheses. Additional covariates include age, employee, and region indicators. * p<0.10, ** p<0.05, *** p< The atypical aspect of this application is that there was no end state; in other words, everyone was eligible for a cancer screening (and they were in the initial state) at the beginning of the time period. The dataset is a stock sample which means that the individuals in the data entered the initial state sometime in the interval [0,b] while the data began recording at time b. Fortunately, the observed sample was not selected because there was no absorbing end state, and thus we can treat the sample as if it were a flow sample with a simple likelihood function. 18

19 Table B1 reports the coefficient and hazard ratio estimates for the three screening procedures under the start date assumption of Jan 1, The most notable fact from this table is that the benefit change has little effect on the time between procedures as it is defined for the hazard model. Inspecting Table B2, we see similar results using an alternate start date assumption. Unfortunately, we cannot tell if the overall lack of evidence of an effect on screenings reveals that there was indeed no effect on frequency of screenings. Insignificant estimates may simply reflect the unrealistic assumptions we were forced to make in the face of serious censoring issues. Additionally, much of the data is not used for identification of the hazard estimates because censoring leads us to investigate only a single spell for each person. The limits of identification in the duration analysis seem to be greater than the limits of the difference-in-difference analysis. Table B2: Hazard Estimates of Differential Price Change with Uniform Assumption a: Cervical Cancer b: Breast Cancer c. Prostate Cancer All Rural Nonrural All Rural Nonrural All Rural Nonrural treat (0.084) (0.132) (0.111) (0.115) (0.192) (0.147) (0.274) (0.430) (0.363) Subjects % Fail 75.8% 76.8% 74.7% 73.5% 73.9% 73.2% 26.1% 27.7% 24.5% Note: A Cox Proportional Hazard Model is estimated on the duration between the start of the period and the observed preventive care visit. The start an individual s duration is assumed to be uniformly distributed over the four years before the sample start date Jan 1, Coefficients are displayed above with standard errors in parentheses. Additional covariates include age, employee, and region indicators. * p<0.10, ** p<0.05, *** p<

20 C Appendix: Robustness of Regression Analysis, Tables Table C1: Effect on Differential Price Change by Procedure, Robustness a: Cervical Cancer b: Breast Cancer All Rural Nonrural All Rural Nonrural treat ** (0.013) (0.021) (0.016) (0.016) (0.027) (0.021) N c: Prostate Cancer d: Child Imm All Rural Nonrural All Rural Nonrural treat ** 0.056*** (0.013) (0.021) (0.016) (0.017) (0.024) (0.024) N Note: Columns report results for OLS estimation of equation (1). The dependent variable is an indicator that is equal to one when the person has the procedure in the relevant year and zero otherwise, with the exception of procedures done in December preceding the benefit change or January following the change which are ignored. All robust standard errors are clustered at the person level. Year treatment group, region, employee, and age fixed effects included in all specifications. Age fixed effects are indicator functions that denote five-year age increments. Treat is an indicator variable that is equal to one when the person is on the new menu of plans. Rural is a binary variable indicating that the location of the person is in an area classified by the US Census as non-urbanized. Person std dev is the within person standard deviation. In the child immunization panel, the number of people who are in rural and nonrural add up to more than in the pooled sample because some employees move within the company from (to) rural places to (from) nonrural places. * p<0.10, ** p<0.05, *** p<