Health insurance in general, and prescription drug insurance in particular, has widely been considered by the economics literature to induce moral hazard. Moral hazard in this context means that when a beneficiary’s drugs are paid for by a prescription drug plan, the beneficiary does not factor in the full drug costs in their consumption decision, and utilizes a sub-optimally high quantity (see Chandra et al., 2007 for a review of this literature). Traditionally, moral hazard has been counteracted by requiring beneficiaries to pay for a portion of the drugs’ cost out-of-pocket (OOP), in an effort to align their incentives when choosing among treatment options. OOP cost sharing requirements can results in beneficiaries bearing considerable out-of-pocket costs for their medical treatments if they end up in a health state requiring costly care. Economically, this reduced generosity of insurance coverage amounts to restricting health insurance to offer only partial insurance due to moral hazard concerns.2

The past decades have seen a growth in formulary-based utilization management in prescription drug insurance design. A prescription drug formulary is a document listing a limited set of drugs covered by the insurance, whereby drugs excluded from the list will be denied coverage; it segments drugs into tiers, imposing differential OOP cost sharing by drug tier; and it can impose additional utilization management mechanisms on select drugs. These standardized mechanisms include step therapy (requiring beneficiaries be treated with a less costly drug before being treated with the costly drug), prior authorization (requiring the provider to submit a request that a beneficiary’s prescription be covered by the insurer before being dispensed) and quantity limits (restricting the quantity of the drug that can be dispensed within a certain timeframe).

Fig. 1 shows that these restrictions, and in particular prior authorization, have become more prevalent in Medicare Part D over time.3 This rise of prior authorization has resulted in backlash from providers, who are concerned with the amount of administrative burden created by the prior authorization process. For example, the American Medical Association (AMA) advocacy efforts regarding prior authorization include providers attesting that “the amount of time that it takes myself and my office staff to go through that process now for almost every prescription we write has got to be an enormous problem both for us and for our patients.”4

Coscelli (2000) has found that the choice of medical treatment is determined jointly by the patient and the provider. Neither the patient nor the provider are liable for the full costs of the medical treatment, which is paid by the insurer, or in the case of Medicare Part D, ultimately the taxpayer. OOP costs discourage utilization of high cost drugs by imposing costs on the patient. In contrast, prior authorization amounts to imposing administrative costs on the provider. Even if, as the AMA claims, prior authorization increases providers’ administrative costs without medical benefit, such an approach may be preferable to imposing cost-sharing requirements on patients in order to address moral hazard. Prior authorization would be preferable if for a small amount of costs incurred for processing prior authorizations, a large amount of OOP costs could be waived. Therefore, whether the administrative costs incurred for processing prior authorizations are justified in the Medicare Part D design hinges on the relative effectiveness of OOP costs vs prior authorization at shifting demand.

This paper develops a method for estimating the elasticity of demand for prescription drugs with respect to OOP costs and to some commonly employed utilization management strategies, including prior authorization, step therapy and quantity limits. I develop a demand model, that highlights a couple of important econometric endogeneity concerns:

If higher OOP costs and utilization management mechanisms are placed on drugs which provide higher utility,5 this would create a bias for underestimating the effect of these mechanisms on demand. I.e., even if higher OOP costs and/or utilization management cause beneficiaries to utilize these drugs less than they otherwise would have, because these mechanisms are applied to drugs which would have otherwise been more commonly utilized this would attenuate the association between OOP costs and/or utilization management and reduced demand.

If beneficiaries select into plans which impose less OOP costs and utilization management on drugs they need, this would create a bias for overestimating the effect of these mechanisms on demand. I.e., if beneficiaries with need for certain drugs enroll in plans with low OOP costs and/or utilization management on these drugs, this will result in a negative association between OOP costs and/or utilization management with demand, even if these variables have no effect on demand per se.

I propose an instrumental variable (IV) method for estimating the model and addressing these endogeneity concerns. The instrumental variable strategy is implemented within a structural model akin to Berry et al. (1995), where the estimated demand model incorporates many of the features of Medicare Part D, including treating drugs as differentiated products, the specifics of Medicare Part D benefit phase design, and switching between plans. I use fixed effects to account for the unobserved utility from each drug. I use an approach similar to Abaluck et al. (2018) to address selection, transforming the data into a panel-like structure of beneficiaries who are members of the same plan over consecutive years. The existence of inertia in plan selection guarantees that most beneficiaries stay in the same plan even when those plans impose additional utilization management mechanisms. The IV strategy isolates the demand response to variation in OOP cost sharing and utilization management within the same plan over time, holding selection into plans fixed. The approach is explained in more detail in Section 4.

Using this estimation approach, I find that the prior authorization and step therapy both reduce demand in a magnitude comparable to a large increase in OOP costs. In contrast, quantity limits have a negligible effect on demand. This does not imply that quantity limits do not have other effects, perhaps reducing abuse, but I find that they do not have an aggregate effect on demand. I also find that the reduction in demand induced by prior authorization is not weaker than the reduction in demand induced by step therapy. Finally, I find considerable variation in demand responses among different drug therapeutic classes.

Using the estimates of how OOP costs and prior authorization impact demand, I assess whether usage of prior authorization is desirable by Medicare beneficiaries. In order to assess this, I consider a counterfactual world in which prior authorization was cancelled and replaced with increased OOP costs to achieve the same level of moral hazard mitigation. The demand estimates allow me to compute the level of counterfactual OOP costs that would have to be imposed on beneficiaries to reach the same target level of demand without using the prior authorization tool. The counterfactual distribution of expenditures among beneficiaries will have a lower mean expenditure, due to not needing to finance the costly prior authorization process, but a higher variance, due to increased OOP costs imposed on beneficiaries who utilize high-cost drugs.

In order to asses whether the counterfactual of cancelling prior authorization is desirable, I conduct a veil-of-ignorance type analysis, in which I consider a beneficiary who needs to choose between a lottery of incurring the expenditures of a randomly selected Medicare Part D beneficiary in Medicare’s existing design with prior authorization, versus a randomly selected beneficiary in the counterfactual with no prior authorization. This transforms the decision of the preferred system into a decision under uncertainty, analogous to a young person needing to decide which design of Medicare Part D they would prefer, prior to knowing what their realized health status would be once they become eligible for Medicare. To prefer the current system with prior authorization over the counterfactual where prior authorization is cancelled, the beneficiary needs to prefer an expenditure distribution with higher mean but lower variance, which means they need to be sufficiently risk averse. I back out how risk averse beneficiaries need to be in order to prefer the current prior authorization system. I find that if a Medicare beneficiary was willing to pay a single extra cent per month to remove the entirety of risk from their Medicare Part D expenses (including both premiums and OOP costs), they would prefer to keep the current system with prior authorization rather than cancelling it. I view this as strong evidence that the current prior authorization system is desirable despite the administrative costs incurred to operate it.

This paper contributes to the research on Medicare Part D, one of the main funding sources for prescription drugs in the United States, by providing demand elasticity estimates for a number of utilization management mechanisms, and assessing the desirability of prior authorization. The insights and methods developed in this paper may apply more generally to other health insurance contexts, where similar utilization management mechanisms are becoming increasingly common.

An extensive literature has estimated demand elasticity for prescription drugs. Goldman et al. (2007) provide an overview of the healthcare literature on the topic. A somewhat older literature, including Ellison et al. (1997), Cleanthous (2002), Chaudhuri et al. (2006) and Dunn (2012) has estimated demand for pharmaceuticals as a classical differentiated goods demand estimation problem, but has largely only addressed endogeneity caused by unobservable drug quality by using drug fixed effects, and has not tackled setups with endogeneity caused by selection into plans. A newer literature, including Einav et al. (2015), Abaluck et al. (2018), Einav et al. (2018) and Dalton et al. (2020) has addressed endogeneity caused by selection but has treated drugs as a single good without accounting for substitution effects. The method developed in this paper will both address selection endogeneity and will treat drugs as differentiated products.

Importantly, these previous studies have focused on estimating the elasticity of demand with respect to OOP costs, but did not estimate the elasticity of demand with respect to commonly employed utilization management mechanisms including prior authorization while addressing endogeneity concerns caused by unobserved drug quality and selection into plans. Most of the papers cited above identified their elasticity estimates using variation in cost sharing requirements between the Medicare Part D benefit phases, a dimension along which there is no variation in the non-OOP utilization management mechanisms. Therefore, their identification strategies would not allow for estimating these additional utilization management elasticities. An exception is Abaluck et al. (2018), whose approach will be adapted to the setup in this paper.

The topic of prior authorization for prescription drugs has been relatively under-explored in the health economics literature. A number of studies (Delate et al., 2005, Dillender, 2018, Fischer et al., 2004, Hartung et al., 2004, Law et al., 2008, Siracuse and Vuchetich, 2008, Stein et al., 2014) have documented event studies of the introduction of prior authorization requirements on specific drugs, mostly by pre-post comparison of utilization, and mostly in the Medicaid context. These studies have generally found meaningful effects of prior authorization on demand.

This paper further contributes to the literature of estimating structural selection models. Previous papers, such as Cardon and Hendel (2001), Handel (2013) and Einav et al. (2013) have typically relied on full simulation methods, which address selection by simulating a fully specified plan choice model. This paper’s approach will not require making assumptions regarding the selection process itself, but will rather derive an estimation equation with an explicit residual representing selection and use an IV strategy to estimate the parameters of interest using variation that is orthogonal to the selection residual. The benefit of this approach is that it allows direct assessment of the validity of exclusion restrictions. This approach may be useful in other structural selection models.

Section 2 begins with institutional background and introduces the data used in the analysis. Section 3 derives a demand model. Section 4 discusses the endogeneity challenges and introduces the empirical identification strategy. Section 5 presents the results, and Section 6 presents the counterfactual. Section 7 concludes.