Selection on Moral Hazard in Health Insurance

The paper introduces the novel concept of "selection on moral hazard" in health insurance markets. It posits that individuals' choices of health insurance plans are influenced not only by their risk assessment (adverse selection) but also by their anticipated behavioral response to the chosen plan's cost-sharing features (moral hazard). This 'selection on moral hazard' significantly impacts efforts to mitigate both adverse selection and moral hazard. The study uses individual-level panel data from a single firm to explore this interaction. A key aspect is the identification of the behavioral response to health insurance coverage and the heterogeneity within this response. The authors aim to develop and estimate a model of plan choice and medical utilization to formalize their analysis and understand the implications of selection on moral hazard. They highlight that ignoring this selection effect could lead to substantial overestimation of the spending reduction associated with introducing a high-deductible health insurance option, a common strategy for cost control. The study notes the lack of prior empirical work on this specific topic within insurance markets.

The paper discusses how the existence of selection on moral hazard challenges standard approaches to managing adverse selection and moral hazard in insurance markets. Risk adjustment, a common method for mitigating adverse selection by pricing insurance based on observable risk characteristics, is considered. However, the study suggests that monitoring techniques designed to curb moral hazard (like differentiated cost-sharing across claim categories) may offer additional benefits in combating adverse selection. The typical approach of offering plans with higher consumer cost-sharing to reduce moral hazard is also analyzed. The study argues that if individuals anticipate their behavioral response affecting their plan choice, the actual behavioral response could be drastically different than if these were independent. The authors acknowledge that both the existence and the sign of the relationship between anticipated behavioral response and the demand for higher coverage are uncertain a priori and require empirical investigation. Alcoa Inc., a large US aluminum producer, provides the data source for this empirical investigation.

The research leverages data from Alcoa Inc., focusing on US workers' health insurance choices, subsequent medical utilization, and related demographic information. This includes individual-level details on health insurance options, choices made, and medical expenditures. Crucially, the data includes variation in health insurance options offered to different employee groups. Alcoa introduced a new set of options in 2004 designed to encourage higher cost-sharing, which would have increased the average out-of-pocket share of spending from 13% to 28% if employee behavior remained unchanged. The phased introduction for unionized employees provides a quasi-experimental setting, as the new options were rolled out when existing union contracts expired. This staggered implementation creates a source of variation crucial for identifying the causal impact of health insurance on medical care utilization (moral hazard). The analysis focuses primarily on 2003 and 2004 data (7,570 employee-years, 4,477 unique employees) to avoid complexities introduced by pricing changes and inertial behavior in later years (2005-2006).

Descriptive analysis suggests both heterogeneous moral hazard and selection on it. Individuals choosing more comprehensive coverage also exhibit a greater behavioral response to that coverage. To formally analyze these observations, the researchers develop and estimate a model of plan choice and medical utilization. Results align with the descriptive evidence, facilitating investigation into the selection and moral hazard interaction. The study points out that ignoring selection on moral hazard could lead to a significant overestimation of spending reduction when introducing high-deductible plans. The analysis considers a counterfactual scenario of moving from the most comprehensive to the least comprehensive new option (a no-deductible plan to a high-$3000 family deductible plan). This highlights the significant heterogeneity in moral hazard and the importance of selection, finding that selection on moral hazard is roughly as important as selection on health risk when choosing between these two plans, exceeding the importance of selection on risk aversion.

The research connects to existing literature in health economics, particularly studies examining the impact of higher consumer cost-sharing on spending using experimental and quasi-experimental designs. The study's difference-in-differences approach aligns with this literature but adds the crucial element of accounting for endogenous plan selection. The authors' central difference-in-differences estimate of -0.14 aligns with existing findings (range -0.1 to -0.4), but the heterogeneity and selection effects identified underscore the limitations of estimates ignoring endogenous choice. The research further builds upon the growing literature showing that insurance market selection extends beyond risk, encompassing preferences like risk aversion, cognition, and bequest motives. The study introduces selection on moral hazard as a significant additional dimension, particularly relevant to contract design. This element is especially important when considering contract design to reduce selection and its impact on spending, unlike other selection dimensions. The authors draw parallels to existing econometric work examining treatment effects and selection based on anticipated treatment responses, termed "essential heterogeneity".

The study's primary data source is individual-level panel data from Alcoa Inc., a large multinational aluminum producer. The data covers a substantial number of employee-years (7,570) and unique employees (4,477) during the years 2003 and 2004. The dataset includes comprehensive information on health insurance options available to each employee, their individual coverage choices, and detailed, claim-level data on medical care utilization and expenditures. Crucially, the data includes rich demographic information beyond typical claims data, such as union affiliation, employment type (hourly or salary), age, race, gender, annual earnings, job tenure at Alcoa, and importantly, individual health risk scores. This richness allows for a detailed analysis of employee characteristics and their influence on health insurance plan selection and subsequent health spending. The inclusion of health risk scores enables the researchers to control for pre-existing health conditions, thereby improving the accuracy of their analysis by isolating the impact of cost sharing. The study also notes that data from 2005 and 2006 was excluded due to substantial pricing changes and the resulting inertial behavior in plan choices.

A key feature of the Alcoa data is the variation in health insurance options offered to different groups of workers. This variation is instrumental in identifying and estimating moral hazard. Alcoa introduced new health insurance options in 2004, designed to encourage employees to move to plans with substantially higher consumer cost-sharing. This change is central to the study's design, allowing for a quasi-experimental approach. Notably, for unionized employees, the introduction of these new options was phased in gradually, coinciding with the expiration of existing union contracts. This staggered timing provides a valuable source of exogenous variation, crucial for the difference-in-differences identification strategy employed. The data also includes information on different coverage tiers (employee only, employee plus spouse, employee plus children, family), which are treated as given in the analysis. The study clarifies that there were three Preferred Provider Organization (PPO) options under the original benefits and five distinct PPO options under the new benefits, preventing passive enrollment in existing plans. The analysis focuses on the in-network features of the plans as they account for more than 95% of spending, simplifying the model and avoiding the need to model the in-network/out-of-network decision.

The primary econometric method used is difference-in-differences, exploiting the staggered introduction of the new health insurance options. This approach allows researchers to compare changes in health spending between groups that received the new options at different times and control for time-invariant differences between groups. The initial difference-in-differences specification uses total spending as the dependent variable, considering both level and proportional moral hazard effects. To account for the endogeneity of plan choice and the heterogeneity of moral hazard, the researchers develop a more sophisticated model of plan choice and medical utilization. This model incorporates latent variables representing individuals' expectations about future health risks, their moral hazard types (price elasticity of demand for healthcare), and their risk aversion levels. The model accounts for the endogeneity of plan choice by allowing for the anticipated behavioral response to affect plan selection. The estimation employs Gibbs sampling to handle the complexities introduced by this rich model structure, allowing for efficient estimation despite the numerous parameters and the inherent difficulty of this model. The study explains that using simpler models which do not account for these complexities could lead to biased results, such as overestimation of spending reductions.

The empirical findings strongly support the existence of substantial heterogeneity in individuals' responsiveness to cost-sharing in health insurance. This heterogeneity is a necessary condition for selection on moral hazard to be a significant factor. The study notes that prior research largely ignored this heterogeneity, focusing on average effects. The researchers' analysis reveals that the standard deviation of the spending reduction achieved by switching from a no-deductible plan to a high-deductible plan is more than double the average reduction. This significant heterogeneity in the response to cost-sharing highlights the limitations of analyses focusing solely on average effects and underscores the importance of accounting for individual variations in the response to cost sharing. The study's findings also suggest that the effects of cost-sharing might be additive rather than multiplicative, meaning that the impact of cost-sharing on spending doesn't vary proportionally with individuals' underlying health status. The authors acknowledge that this issue is complex, as changes in health insurance alter individuals' budget sets, and their responsiveness to cost-sharing varies depending on where they are on this budget set. These complexities motivate the detailed modeling approach undertaken in the paper. The difference-in-differences approach provided an estimate of the arc elasticity of medical spending with respect to average out-of-pocket cost share of about -0.14, which is broadly similar to findings from previous experimental studies but it suggests that these prior estimates may overstate the impact of cost sharing because they do not account for endogenous plan selection.

The study finds strong evidence for selection on moral hazard, showing that individuals who exhibit a greater behavioral response to insurance coverage (higher moral hazard) are more likely to select higher coverage plans. This confirms the hypothesis that individuals' anticipated behavioral response to the plan’s features plays a role in their choice of health plan. The analysis directly compares the estimated behavioral response (changes in spending associated with moving from the original to the new options) between those who initially chose higher versus lower coverage. The results reveal a spending reduction associated with the move to the new (higher cost-sharing) options that is more than twice as large for those initially selecting higher coverage. This observation supports the concept of selection on moral hazard, even though the treatment (reduction in cost-sharing) was larger for those initially having lower coverage. While not precise enough to decisively reject the null hypothesis of equal spending reductions across both groups, this descriptive evidence highlights the need for a formal model to account for endogenous plan choice and the variation in marginal prices influenced by health status differences. Further analysis confirms that older and sicker workers, who tend to have larger moral hazard effects, also tend to select more comprehensive insurance plans, further supporting the findings. This correlation, however, doesn't directly isolate whether insurance choices are driven by health expectations or by anticipated behavioral response.

The research findings have significant implications for policies aimed at reducing health care spending and managing adverse selection in health insurance markets. The study highlights that ignoring selection on moral hazard leads to overestimation of spending reductions associated with introducing high-deductible health plans. This is because individuals who choose high-deductible plans tend to exhibit lower levels of moral hazard (lower responsiveness to cost-sharing). The paper suggests that policies focused solely on risk adjustment may be less effective than anticipated. Risk adjustment, a common method for mitigating adverse selection by adjusting premiums based on observable risk characteristics, is presented as an insufficient approach when selection on moral hazard is present. The paper proposes that investing in improved monitoring technologies—a method usually associated with mitigating moral hazard— could also be highly effective in mitigating the adverse selection problem. By carefully considering the different types of claims and their susceptibility to moral hazard, more nuanced cost-sharing strategies could be implemented. This would allow for a more effective approach to both cost containment and managing the complexities of adverse selection.

The study utilizes counterfactual analyses to explore the welfare implications of different contract designs, focusing on the choice between a no-deductible and a high-deductible plan under the new Alcoa benefit options. One counterfactual scenario simulates "perfect screening," where insurers can observe and price on all determinants of healthcare utilization (including moral hazard types). Eliminating adverse selection through this perfect screening scenario leads to lower average premiums for the no-deductible plan, increased coverage of this plan, higher overall welfare, and lower expected spending. The welfare gain per employee is estimated to be around $52. Another counterfactual explores "perfect monitoring," assuming that insurance only covers health-related spending regardless of actual spending. This eliminates moral hazard by aligning incentives, again leading to improved welfare. The research emphasizes that the quantitative estimates are highly context-specific; however, the findings broadly illustrate the importance of accounting for selection on moral hazard in understanding the welfare consequences of selection and moral hazard, as well as informing the design of policies aimed at mitigating these costs. The research suggests that efforts to reduce spending through high consumer cost sharing may produce smaller reductions than expected if selection on moral hazard is ignored. Furthermore, improvements in monitoring technology could provide an ancillary benefit of reducing adverse selection.

The study concludes that selection on moral hazard is a significant phenomenon in health insurance markets. The research demonstrates substantial heterogeneity in individuals' responses to cost-sharing (moral hazard), and importantly, that this heterogeneity significantly influences health insurance plan choices. The magnitude of moral hazard is found to be roughly as important as traditional measures of health risk in determining plan selection. Ignoring this selection effect leads to a substantial overestimation of the spending reduction that would result from introducing high-deductible plans. The estimated arc elasticity of medical spending with respect to the out-of-pocket cost share (-0.14) is comparable to previous research but requires caution given the endogenous nature of plan selection. The study's findings challenge conventional wisdom on the effectiveness of policies designed to address adverse selection and moral hazard, suggesting that alternative approaches, such as improved monitoring technologies, may be more effective in mitigating both adverse selection and moral hazard than previously thought.

The authors advocate for further research to explore selection on moral hazard in diverse contexts and using different methodologies. They emphasize that limited work exists on heterogeneity in moral hazard effects, let alone the selection of insurance based on this heterogeneity. The approaches used in this study, along with others suggested for estimating heterogeneity in moral hazard effects and their correlation with demand, could be usefully applied in other settings. The research has focused primarily on the spending and welfare implications of selection on moral hazard under existing contracts. Future research could broaden this scope by incorporating a theoretical and empirical examination of this selection's implications for a wider range of contract designs, considering more complex scenarios and the interaction of numerous factors. This includes the theoretical development of richer models and empirical investigations into how this phenomenon manifests in other markets. The importance of considering both the level and the slope of spending in the context of insurance plan selection is highlighted as a crucial aspect for future research, as it shows that the extent to which selection is based on moral hazard can be critical for evaluating policy interventions. This work serves as a foundation for more comprehensive studies of health insurance markets and the dynamics of consumer behavior in this context.