Effects of utilization management on health outcomes: evidence from urinary tract infections and community-acquired pneumonia

Abstract

Background:

Utilization management policies are pervasive in the Medicare Part D program. We assess the effect of utilization management restrictions in the Medicare Part D program on the quality of care in two clinical areas—community-acquired pneumonia (CAP) and urinary tract infections (UTI).

Methods:

In this study, we identified new cases of CAP and UTI from Medicare claims data from 2010 to 2016. We assessed the relationship between exposure to utilization management for antibiotic medications suitable for treating these conditions and adverse health outcomes, based on the Agency for Healthcare Research and Quality prevention quality indicators.

Results:

We identified 147526 cases of CAP and 632407 UTI cases in our data. In these samples, the adverse event rate varied from 3.6 to 5.7 percent. The probability of an adverse event increased by 0.75 (p=0.061) percentage points for each ten percentage point increase in exposure to quantity limits (one form of utilization management) among people with CAP. There was no relationship between utilization management and adverse events in the UTI cohort.

Conclusions:

In some circumstances, exposure to utilization management policies–particularly quantity limits–may adversely affect health.

1. Introduction

From its inception in the year 2006, the Medicare Part D program has provided prescription drug coverage through an array of competing private insurance plans with a total cost, as of 2018, of $900 billion [1]. To contain prescription drug spending, many Part D plans use both financial, including deductibles and other forms of cost-sharing, and non-financial methods, policies such as prior authorization—a requirement that the plan approve a prescription in advance—and quantity limits—a fixed limit on the units of a drug that may be dispensed at one time, to limit the use of prescription drugs. These methods are effective at reducing prescription drug use in other settings [2], but may also have the unintended consequence of increasing spending on related healthcare services [3,4]. In essence, prescription drugs may be a substitute for other healthcare services. Previous research has demonstrated that prescription drugs, in the aggregate, can offset other healthcare spending [5,6]. However, these studies have used variation in out-of-pocket costs, which may not generalize to non-financial barriers to accessing a particular prescription drug such as prior authorization and quantity limits.

In this paper, we test for adverse effects of utilization management in the Medicare Part D program using two common acute clinical conditions—urinary tract infection (UTI) and community-acquired pneumonia (CAP)—which are treatable with antibiotics. We measure treatment (in)effectiveness for UTI (CAP) using avoidable emergency room visits and inpatient hospitalizations for UTI (CAP), as defined in the AHRQ Prevention Quality Indicators [7], which we can identify in Medicare claims data. We study the association between utilization management and adverse events in two ways. First, we use exposure to utilization management, which reflects the share of antibiotics that are subject to utilization management. Second, we use the utilization management policies attached to the drugs that are dispensed. Because people select into plans, we use instrumental variable methods to estimate the causal relationship between utilization management and adverse outcomes, using variation in exposure to prior authorization and quantity limits that does not depend on a person’s own plan choice, but rather arises from the interaction of lagged health status, time, and location.

Antibiotics prescriptions are especially relevant to the Medicare Part D population, and with an average of 1,400 prescriptions per 1,000 beneficiaries per year and costing more than $1.5 billion in total, “antibiotic stewardship efforts” such as utilization management have the potential to affect outcomes at both the extensive and intensive margin [8]. It is also the case that antibiotics have a short course of treatment [9], unlike drugs taken for chronic or long-term illnesses. This characteristic gives strength to the causality argument since it mitigates the effects of user-specific behaviors that may wash out treatment effects. Furthermore, antibiotics are the primary means of treatment for both CAP and UTI [9,10]. Since both diseases have the potential for complications if the correct treatment plan is not followed, variations in antibiotic treatment regimes that arise due to health plan-level characteristics provide an avenue to test the health implications of utilization management.

2. Background

2.1. Prescription drug benefit

The Medicare prescription drug benefit, Medicare Part D, was enacted in 2003 and took effect in 2006 [11]. Individuals enrolled in Medicare Part D choose a plan in each year from a set of plans offered in the Part D region in which they live. There are 34 plan regions, which are collections of states, and the set of plans available to an individual vary over time and across regions. Plans are differentiated on salient features (premiums, deductibles, and cost-sharing tiers) and less salient features such as drug formularies that affect out-of-pocket costs. Drug formularies specify the drugs that each plan covers, the drug’s assigned “tier” which affects the out-of-pocket costs of a drug, and any forms of utilization management applicable to the drug. Part D regulations require that all plans cover a minimum of two drugs in each therapeutic class and designates six classes of drugs as “protected classes” for which plans are required to cover substantially all of the drugs in those classes.

The design of the standard Part D benefit includes an initial deductible, followed by an “initial coverage zone” in which the enrollee is responsible for 25% of drug spending (from $480 to $4430 in total costs in 2021), a coverage gap, or donut hole, (from $4430 to $10690 in total costs in 2021), and a reinsurance regime for which enrollees pay 5%. The donut hole has changed during our study period from requiring that enrollees pay all costs to the point where enrollees pay one-quarter of the cost, insurers pay 5%, and a mandatory 70% discount from manufacturers accounts for the rest of spending in the donut hole. To help low-income seniors afford prescription drug coverage, Medicare has implemented a low-income subsidy program that reduces premiums and out-of-pocket spending for individuals with income below the poverty threshold. These beneficiaries have reduced cost-sharing requirements but are restricted in the plans that they may choose.

The Medicare Part D program significantly increased prescription drug coverage for seniors and increased prescription drug utilization for this population [12]. In addition, there is a robust literature demonstrating substantial health benefits associated with the Medicare Part D program, particularly in areas such as cardiovascular diseases [13–15], chronic obstructive pulmonary disease [13,15], and diabetes and stroke [13], with some of these benefits arising from a reduction in beneficiaries’ cost-related drug nonadherence [16]. The introduction of Medicare Part D also increased antibiotics prescription [9], with prescribing rates and inappropriate prescriptions decreasing since 2011 [17].

2.2. Utilization management

Utilization management is implemented in three main forms: prior authorization; quantity limits; and step therapy. We focus on prior authorization and quantity limits since none of the antibiotics we studied were subject to step therapy during our study period.

Prior authorization policies require that an insurer approve a prescription before it is filled. This process requires justifying the choice of one drug over others, which may require assembling additional documentation for the plan provider [18]. This process burdens health service providers who must spend time communicating with plan providers regarding their prescription decisions, and physicians report that they regularly face delays of up to a day before their queries on a given course of treatment are resolved [19]. Nevertheless, prior authorization reduces pharmacy-related treatment costs and is a commonly used formulary restriction [2]. Because of the evaluative nature of prior authorization, we hypothesize that prior authorization policies may have ambiguous effects on beneficiary health. On the one hand, prior authorization programs reduce access to some drugs, on the other hand, the insurer may have an informational advantage that could assist physicians in making better and more informed prescribing choices.

Quantity limits are a second commonly used formulary restriction. Typically, these policies limit the number of units of a drug that may be dispensed at a time. These restrictions tend to be quite blunt since they often apply to all prescriptions for a given drug, rather than reflecting the needs of a drug under a particular set of circumstances. However, unlike prior authorization, which imposes costs on physicians, quantity limits primarily impose costs on patients who may need to make additional trips to the pharmacy to complete a course of treatment with a given drug. As a result, quantity limit policies are more like conventional financial barriers to care. Hence, we hypothesize that quantity limit policies will adversely affect beneficiary health.

3. Methods

3.1. Analysis sample

We used claims data for a 5% random sample of Medicare beneficiaries from 2010 to 2016. We extracted incident cases of CAP or UTI and collected adverse events data from a subsequent follow-up period.

We identified index CAP and UTI events based on the presence of a diagnosis code for CAP or UTI with no prior diagnosis for the preceding six months. In order to ensure that we always observe this washout period, we only include events after a beneficiary has been enrolled in traditional Medicare for six months or more. Some beneficiaries may have multiple index events either across the two samples (i.e., has CAP and UTI at different points in time) or in the same sample (i.e., has repeated CAP or UTI events). Following the AHRQ criteria for the UTI and CAP PQIs [7], we excluded CAP claims accompanied by a diagnosis of sickle cell disease or any secondary code indicating that a beneficiary was immunocompromised. For the UTI sample, we excluded claims accompanied by diagnoses for kidney disorders and those indicating that a beneficiary was immunocompromised. Our final sample includes 147,526 index events in the CAP sample and 632,407 in the UTI sample.

For the CAP (UTI) sample, we identified adverse events, defined as avoidable emergency room visits and hospitalizations, using the CAP (UTI) AHRQ Prevention Quality Indicators [7], over the subsequent six months (three months in a sensitivity analysis). These measures identify markers of low-quality treatment for ambulatory-care sensitive conditions which, in this case, are managed with outpatient antibiotics. We identified adverse events using Medicare outpatient claims files and the MedPAR file for inpatient hospitalizations. We counted hospitalizations originating in the emergency room as both a hospitalization event and an emergency room event and defined an omnibus indicator for having either an emergency room visit or a hospitalization associated with CAP or UTI..

We collected prescription drug utilization data from the Medicare Part D Event files. These files track prescription drug events that are paid under the Medicare Part D benefit. Therefore we do not observe drugs that individuals purchased for cash outside of the prescription drug benefit. We also identified the first antibiotic prescription after each index event that a beneficiary filled within two weeks of the index event date. In the case of multiple antibiotic fills, we chose the earliest and most expensive drug. We assigned the utilization management status of each drug based on the plan- and year-specific formulary covering the prescription.

We focus on two independent variables: exposure to utilization management; and incidence of utilization management prescriptions. We measured exposure to utilization management as the plan-year specific dollar-share of antibiotics that are subject to utilization management. For each antibiotic used within six months of either a CAP or UTI diagnosis we first calculate the relative proportion of each drug, separately for CAP and UTI. Finally, each of these relative proportions are weighted based on the share of each of these antibiotic’s contribution to total (antibiotics) drug cost. We measured the incidence of utilization management using flags in the prescription drug event file for prior authorization and quantity limits status.

We also constructed control variables for health status using Medicare claims files. We used the Carrier, Outpatient, and MedPAR files to identify diagnoses over the prior year for use as risk adjusters in our regression models. We aggregated the diagnosis codes using the Medicare Part D risk adjustment classification system (RxHCC) for data from 2015, which provides codes for both ICD-9 and ICD-10 and included an indicator for each RxHCC in the regression model. These variables provide controls for underlying health risk that may affect health outcomes and prescribing decisions and insurance enrollment decisions.

We defer a discussion of our instruments until we discuss identification below.

Basic information on the CAP and UTI samples are given in Table 1. Our CAP sample includes over 147,000 index events, of which 7% received a prescription subject to some form of utilization management. The UTI sample is considerably larger, consisting of 632,407 index events of which 7% received a prescription subject to some form of utilization management. Adverse events followed between 3.6 and 5.7% of the index events, with the most common adverse event being a hospitalization for CAP and an emergency room visit for UTI. Adverse event rates are also higher when we condition on the subset of index events that resulted in a prescription with some form of utilization management, which would indicate that managed drugs were more likely to be prescribed in more complex situations. There were only minor differences in the proportion of antibiotics subject to either prior authorization or a quantity limit in the various chosen plans. However, there were stark differences in the frequency with which drugs subject to prior authorization or quantity limits were prescribed. Drugs that were subject to utilization management also tended to be more expensive for both the CAP and UTI samples, which is consistent with insurers using utilization management to steer patients towards lower-cost alternatives.

Table 1:

Characteristics of each index event by sample

	CAP	CAP w/UM presc	UTI	UTI w/UM presc
Outcomes
ER visit as adverse outcome	0.012	0.012	0.022	0.031
IPH as adverse outcome	0.048	0.058	0.015	0.020
Any adverse outcome	0.057	0.066	0.036	0.051
Exposure
Proportion of PA drugs in chosen plan	0.202	0.224	0.209	0.205
	(0.129)	(0.127)	(0.118)	(0.097)
Proportion of QL drugs in chosen plan	0.113	0.436	0.097	0.177
	(0.171)	(0.203)	(0.132)	(0.175)
Prescribing patterns
PA drug prescribed	0.009	0.126	0.044	0.624
QL drug prescribed	0.062	0.885	0.027	0.387
Gross Drug Cost	45.15	124.80	27.50	72.39
	(160.99)	(458.16)	(106.21)	(290.59)
Demographics
Age	77.7	77.2	77.1	76.6
	(8.2)	(8.1)	(7.8)	(7.7)
Female	0.591	0.603	0.826	0.870
Not non-Hispanic white	0.040	0.042	0.040	0.039
Instruments
Proportion of PA drugs in plan market	0.200	0.196	0.207	0.219
	(0.063)	(0.071)	(0.062)	(0.055)
Proportion of QL drugs in plan market	0.115	0.141	0.098	0.104
	(0.045)	(0.056)	(0.033)	(0.035)
Gross Drug Cost	45.15	124.80	27.50	72.39
	(160.99)	(458.16)	(106.21)	(290.59)
Age	77.7	77.2	77.1	76.6
	(8.2)	(8.1)	(7.8)	(7.7)
Female	0.591	0.603	0.826	0.870
Not non-Hispanic white	0.040	0.042	0.040	0.039
Observations	147,526	10,392	632,407	44,195

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—Table presents means (standard deviations) for all index events in the dataset.

There are minor differences in demographic characteristics between the full sample for each acute condition and the subset with a prescription subject to utilization management. There are more notable differences across populations, with a greater share of women in the UTI sample. This difference is consistent with human biology.

3.2. Empirical model

Our regression of interest relates adverse outcomes to the prescription drug policies affecting drugs used to treat either CAP or UTI:

$$Advers{e}_{ist}={\beta }^S{X}_{it}+{\Gamma }^S{H}_{i,t-1}+{\pi }^{S}U{M}_{it}^S+stat{e}_s+tim{e}_t+{\epsilon }_{it},$$ (1)

where $i$ indexes each event (we use the term beneficiary interchangeably unless it creates an ambiguity) and $t$ is time, measured in calendar years. $X_{it}$ is a vector of demographic characteristics of the beneficiary, $H_{i,t-1}$ is a set of lagged clinical condition indicators, and $U{M}_{it}^S$ measures the utilization management policies that apply to setting $S$ (CAP or UTI) in the plan chosen by person $i$ in year $t.stat{e}_s$ and $tim{e}_t$ are sets of state and year fixed effects. $Advers{e}_{ist}$ is the adverse outcome of interest, which is one of an AHRQ avoidable ER visits, AHRQ avoidable hospitalization, or an indicator for a avoidable ER visit or hospitalization. The theoretical underpinning of our regression model is a production function of good health (i.e. lack of adverse events) that includes a vector of prescription drug inputs. These drug inputs are affected by the presence of utilization management so that our regression model estimates a reduced-form version of the production function.

The coefficients of interest in this model are the ${\pi }^S$ coefficients that estimate the relationship between utilization management policies and adverse health events. Note that utilization management indicates the dollar-weighted proportion of drugs in the beneficiary’s chosen health plan that is under utilization management. Hence, for each claim, $U{M}_{it}$ is a 2-vector, representing the weighted proportion of prior authorization and quantity limit within the chosen plan.

The key econometric challenge in estimating equation (1) arises because the utilization management policies of the plan chosen by person $i$ may be endogenous since one is likely to choose an insurance plan in light of one’s expected utilization or expected health outcomes. Such a situation could arise, for example, if someone chooses a plan that is amenable to their demand for drugs for foreseeable needs such as chronic conditions but is more restrictive towards drugs used in the ambulatory treatment of CAP and UTI. Likewise, if beneficiaries who lead healthy lifestyles and are unlikely to get sick choose more restrictive plans, the estimates of the effect of utilization management on adverse health outcomes will likely be biased downwards. We utilize an instrumental variables approach to tackle this issue.

3.3. Instrument variable

In order to identify the causal effect of utilization management on adverse health outcomes, we generate plausibly exogenous variation in exposure to utilization management from what is, effectively, the interaction of one’s state of residence, calendar year, and lagged health status.

We construct our instruments in four steps.

First, we collect information on all the drugs dispensed to each beneficiary. We then simulate 00P spending and utilization management exposure based on the characteristics of each Part-D plan available for the beneficiary. The program to run this simulation borrows from the “cost calculator” employed by [20]. It compiles the costs a beneficiary would face for the drugs they were dispensed under their chosen plan if they were enrolled in another Part D plan.

Second, we take these simulated expenses and estimate “ex-ante” 00P spending and spending on drugs subject to utilization management in each plan, including for the plan that was actually chosen, using a plan-by-year-by-region specific Poisson regression on demographic characteristics and lagged health status.

Third, we estimate the probability that a beneficiary chooses to enroll in a particular plan using a discrete choice model. We estimate a multinomial logit model of plan choice on plan fixed effects and the person-specific expected out-of-pocket spending and expected share of spending subject to prior authorization and quantity limits. We then compute the probability that a beneficiary enrolled in each plan available to her based on her demographics and lagged health status.

Fourth, we use the probabilities of plan enrollment as weights to calculate the weighted average of the plan-specific CAP and UTI utilization management indices. We report the values of our instruments in Table 1 in the rows labeled “Proportion of PA drugs in plan market” and “Proportion of QL drugs in plan market.”

The first stage regressions are of the form:

$${\widehat{UM}}_{it}^S={\tilde{\beta }}^S{X}_{it}+{\tilde{\Gamma }}^S{H}_{i,t-1}+{\tilde{\varphi }}^{S}I{V}_{it}^S+{st\tilde{a}te}_s+{ti\tilde{m}e}_t+{\tilde{\epsilon }}_{it},$$ (2)

where ${\widehat{UM}}_{it}^S$ is the predicted utilization management share for person $i$ at time $t$, and $I{V}_{it}^S$ is the constructed instrument. As in equation (1), $I{V}_{it}$ is a 2-vector, with components indicating the proportion of prior authorization and quantity limit drugs in all plans for time $t$ in beneficiary $i$’s market for plans. Note that if individuals do not know the probability of developing condition $S$, and if the utilization management status of drugs to treat condition $S$ is independent of person $i$’s other drugs, then there is no endogeneity problem since there is no private information that could bias estimates of equation (1). Results from our first stage regressions are presented in Table 5 and indicate that these instruments are strongly associated with the prior authorization and quantity limit exposure in an individual’s chosen plan.

The first stage results in Table 5 show that as the weighted average of drugs subject to utilization management in the choice set of all plans available to the beneficiary (the instrument) increases, the weighted average of drugs subject to utilization management in the beneficiary’s chosen plan also increases. The relationship is almost one-to-one and highly significant, suggesting that our instrument accurately predicts the chosen-plan index. In the CAP sample, we see that a one percent increase in the dollar-weighted proportion of prior authorization drugs in the universe of all plans leads to a 0.9 percent increase in the dollar-weighted proportion of prior authorization drugs in the beneficiary’s chosen plan and a 1-to-0.83 relationship in the case of quantity limits. Likewise, in the UTI sample, both prior authorization and quantity limits in the beneficiary’s chosen plan increase by roughly 0.9 percent with a one percent increase in the dollar-weighted proportion of drugs subject to utilization management in the universe of all plans. The F-statistics on the excluded instruments, which measures the relevance of our instruments for each endogenous variable, is consistently greater than ten, indicating that our instruments are highly predictive of the actual share of antibiotics in the plan that are subject to prior authorization or a quantity limit [21]. In the appendix we also report reduced form results that arise from substituting equation (2) into equation (1), with the coefficients of interest being the product.

4. Results

4.1. Effects of utilization management exposure on adverse events and prescribing behavior

Table 2 presents our main results. Panel A demonstrates that exposure to prior authorization is associated with a reduction in the incidence of adverse events for CAP, with a one percentage point (5.0%) increase in the share of spending subject to prior authorization reducing the likelihood of any adverse event by 1.0 percentage points, or 17.5% of the mean. The corresponding IV estimate is comparable in size, but the estimate is very imprecise. We also find that quantity limits have a causal effect on adverse events and emergency room visits. A one percentage point increase in quantity limit exposure (8.8%) increases the probability of an adverse event by 7.5 percentage points, or 132% of the mean. While most of these additional adverse events result in a hospitalization, our estimates are substantially more precise for emergency room-related adverse events. Panel B, which presents results for UTI, provides no evidence of a relationship between utilization management policies and adverse UTI-related events.

Table 2:

Association of the dollar share of drugs subject to utilization management with adverse health outcomes

	Any adverse		Emergency Room		Hospitalization
	OLS	IV	OLS	IV	OLS	IV
Panel A: CAP
Prop PA drugs in chosen plan	−0.010+	−0.013	−0.001	0.017	−0.005	−0.021
	(0.005)	(0.072)	(0.003)	(0.036)	(0.005)	(0.066)
Prop QL drugs in chosen plan	0.004	0.075+	0.001	0.039*	0.004	0.056
	(0.004)	(0.040)	(0.002)	(0.019)	(0.004)	(0.037)
Panel B: UTI
Prop PA drugs in chosen plan	−0.002	0.000	−0.001	0.007	−0.001	0.001
	(0.002)	(0.024)	(0.002)	(0.020)	(0.002)	(0.016)
Prop QL drugs in chosen plan	0.002	0.012	0.000	−0.003	0.001	0.015
	(0.002)	(0.016)	(0.001)	(0.013)	(0.001)	(0.011)

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—

^+,*,**,***indicate significance at 10%, 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for White, and for non-Hispanic, non-White; state; year; and RXHCC. Each column in each panel is from a separate regression. “Any adverse” includes both emergency room and hospitalization adverse outcomes. IV models are estimated using the weighted average exposure to drugs subject to prior authorization (“PA”) and quantity limits (“QL”) for each beneficiary. Standard errors clustered on beneficiary in parentheses.

Recall that we define adverse events as an emergency room visit or a hospitalization within six months of the index event when the individual is diagnosed with either CAP or UTI. To ensure that our results were not affected by our chosen definition of an adverse event, we run a second analysis where we define adverse events as ER visits or IPH events that occur within three months of the index event. These results are presented in Table 6, which yields qualitatively similar, although less precise, estimates as our main results. The notable exception is for the effect of exposure to prior authorization, which is challenging to identify because of the ambiguous direct effect of prior authorization status on outcomes. We discuss this issue in more detail below, when we present results using drug-specific utilization management status

4.2. Effect of utilization management exposure on prescribing behavior and adverse outcomes

Next, we consider one of the most plausible mechanisms linking exposure to utilization management with adverse health outcomes–prescribing behavior. To do so, we re-estimate our main models but use the prior authorization and quantity limits status of the prescribed drug as our dependent variable. The results of this exercise are presented in Table 3.

Table 3:

Association of dollar share exposure to utilization management and the utilization management status of filled drugs

	Prior authorization		Quantity limits
	OLS	IV	OLS	IV
Panel A: CAP
Prop PA drugs in chosen plan	−0.029***	0.064*	−0.035***	−0.094
	(0.002)	(0.028)	(0.004)	(0.063)
Prop QL drugs in chosen plan	0.004*	−0.028*	0.779***	0.824***
	(0.001)	(0.014)	(0.007)	(0.038)
Panel B: UTI
Prop PA drugs in chosen plan	−0.114***	0.567***	−0.064***	−0.025
	(0.002)	(0.028)	(0.002)	(0.019)
Prop QL drugs in chosen plan	−0.020***	−0.148***	0.365***	0.743***
	(0.002)	(0.017)	(0.003)	(0.017)

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—

^+,*,**,***indicate significance at 10%, 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for White, and for non-Hispanic, non-White; state; year; and RXHCC. Each column in each panel is from a separate regression. IV models are estimated using the weighted average exposure to drugs subject to prior authorization (“PA”) and quantity limits (“QL”) for each beneficiary. Standard errors clustered on beneficiary in parentheses.

There are two key results in the table. First, the IV estimates are consistently larger than the OLS estimates for the “own-type” comparisons. For example, the prior authorization index has an association of −0.029 in the OLS model for receiving a drug subject to prior authorization, but that relationship increases to 0.064 in the IV model. Results for the other “own-type” comparisons are comparable in direction and often larger in magnitude. This pattern is consistent with a form of deterrence in which utilization management policies reduce the likelihood of either being prescribed a drug that is subject to utilization management or filling a prescription for such a drug.

The second key result in the table is that the prior authorization and quantity limit indices are, with the exception of prior authorization for CAP patients, strongly associated with receiving a drug subject to prior authorization or a quantity limit. In the case of receiving a drug that is subject to quantity limits, a one percentage point increase in the spending share of drugs subject to quantity limits increases the probability of receiving a drug that has a quantity limit by 0.7 to 0.8 percentage points in our IV models for the CAP and UTI samples. The association between prior authorization exposure and the probability of receiving a drug that is subject to prior authorization is also increasing in the prior authorization index, but by less–only 0.6 percentage points per one percentage point increase in the exposure index.

One puzzle is why the relationship between prior authorization exposure and receiving a drug subject to prior authorization is so weak in the CAP sample. The difficulty arises from the reduced form regression (which is presented in Appendix Table 7). The instrument for prior authorization explains a very small fraction of the variation in the probability that the drug a beneficiary filled was, in fact, subject to prior authorization. Considering how prior authorization requirements work, it should not be that surprising that exposure to prior authorization is weakly correlated with receiving a drug subject to prior authorization. On the one hand, a prior authorization requirement discourages prescriptions for a given drug (or filling said prescriptions). On the other hand, if the share of drugs subject to prior authorization is high then it is more difficult to find a drug that is not subject to prior authorization.

We exclude prior authorization from the CAP sample because the instrument is not sufficiently strong in predicting the prior authorization status of prescriptions. This is demonstrated in the first column of Table 7, where we can see that the F-statistic from the reduced form regression, which predicts prior authorization prescriptions using the IV, is very low at only 5.77. The coefficients on the estimated regression equations also highlights the same issue as the coefficients on prior authorization are weak and not economically significant. We further illustrate this in Figure 1, which plots the linear bivariate relationships that correspond to the OLS and reduced-form regressions. The top right quadrant of panel A demonstrates that the estimates for CAP with respect to prior authorization is small in magnitude and greatly dispersed around the best-fit conditional expectation line, indicating poor fit.

Figure 1:

Binned Scatterplot for regression of utilization management status of prescribed drug on proportion of drugs subject to utilization management within health plan

The first stage regression excluding the prior authorization variable in the CAP sample is given in Table 8. Once we ignore prior authorization in the CAP sample, we are left with strong first stage results. Furthermore, the regression coefficients are also reasonably large: for example, in the CAP sample, a 10 percentage point increase in the proportion of drugs with a quantity limit in the universe of all plans is associated with a 6.14 percent increase in the probability of being prescribed a drug with a quantity limit. The results for the UTI sample are also in the same ballpark, with a 10 percentage point increase in instrument contributing to a 5 percent increase in prior authorization-prescriptions and a 6.6 percent increase in prescriptions with a quantity limit.

Finally, we turn to assessing if the utilization management status of the prescribed drug affects health outcomes. In doing so, however, we cannot estimate effects for receiving a drug subject to prior authorization in the CAP sample since our instruments are not powerful enough to identify a causal effect [22]. The model is virtually identical to equation (1), except that the utilization management variables now refer to the utilization management status of the drug the beneficiary filled. Table 4 presents the results of this analysis, which follows the same format and conventions as Table 2, except for the exclusion of prior authorization in the CAP sample.

Table 4:

Association of the UM status of the filled drug and adverse events

	Any adverse		Emergency room		Hospitalization
	OLS	IV	OLS	IV	OLS	IV
Panel A: CAP
PA drug presc	-	-	-	-	-	-
	-	-	-	-	-	-
QL drug presc	0.008*	0.091+	0.000	0.050*	0.009**	0.067
	(0.003)	(0.048)	(0.001)	(0.023)	(0.002)	(0.044)
Panel B: UTI
PA drug presc	0.019***	0.001	0.014***	0.012	0.006***	0.003
	(0.001)	(0.043)	(0.001)	(0.035)	(0.001)	(0.028)
QL drug presc	0.013***	0.016	0.006***	−0.002	0.008***	0.021
	(0.002)	(0.024)	(0.001)	(0.019)	(0.001)	(0.016)

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—

^+,*,**,***indicate significance at 10%, 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for White, and for non-Hispanic, non-White; state; year; and RXHCC. Each column in each panel is from a separate regression. “Any adverse” includes both emergency room and hospitalization adverse outcomes. IV models are estimated using the weighted average exposure to drugs subject to prior authorization (“PA”) and quantity limits (“QL”) for each beneficiary. Standard errors clustered on beneficiary in parentheses.

Table 4 estimates the probability of adverse events when a beneficiary is prescribed a drug subject to utilization management. This table corroborates the estimates from Table 2. Furthermore, while Table 2 echoes the effects of exposure to treatment, Table 4 provides closer estimates of treatment effect on the treated. As expected, the treatment effects from being treated with utilization management are larger in magnitude than the effects of being exposed to high proportion of utilization management.

For the CAP sample, we find that a prescription subject to quantity limits is associated with a 9.1 percentage point increase in the incidence of adverse events, driven by an increase in emergency room admissions. In the UTI sample, we find no statistically significant effects of the utilization management status of the prescribed drug in our IV models, although our OLS results indicate that receiving a drug that is subject to prior authorization or quantity limits is associated with a higher likelihood of adverse events—a drug with quantity limits, for example, is associated with a 1.3 percentage point increase in the probability of any adverse event. Given that the IV and OLS estimates are virtually indistinguishable for any adverse events, we interpret this as evidence of a causal effect of quantity limits on adverse events in the UTI sample, with a prescription subject to quantity limits increasing the probability of an adverse event by 1.3 percentage points. However, that logic does not extend to the specific types of adverse events, since the OLS and IV estimates differ considerably for those outcomes.

We repeat the analysis from Table 4 using 3-month windows for adverse outcomes and present the results in Table 10. Unlike Table 6 where we lose significance, the results are virtually unchanged here. For the CAP sample, where we ignore prior authorization, we find that receiving a prescription subject to quantity limits, as opposed to one without, was associated with a 1.1 percentage point increase in the probability of adverse events. Likewise, for the UTI sample, we find that a prescription subject to either a prior authorization or quantity limit requirement is associated with a 1.4 percentage point increase in the frequency of adverse events. The results are also similar to Table 4 in that for the CAP sample we find that hospitalizations contribute to most adverse events.

5. Discussion

Using two clinical settings—community-acquired pneumonia and urinary tract infections—we have assessed the effect of utilization management policies on adverse medical events (emergency room visits and hospitalizations). Utilization management in either form imposes a cost, albeit differently, for Medicare beneficiaries and physicians. In prior authorization, the physician faces a substantial time burden since they have to engage in additional communications with a patient’s insurer. On the other hand, a quantity limit increases costs for the beneficiary who may need to refill a prescription to complete a course of treatment. In addition, these policies may affect health by affecting a patient-physician dyad’s choice of prescription drug (we do not consider this issue explicitly in this paper, but note it as an opportunity for future research). In particular, by increasing the cost to the physician of prescribing some drugs, prior authorization policies may shift prescribing towards other drugs that are not subject to a prior authorization policy. However, doing so may lead a physician to prescribe a drug with which she is less familiar, which may adversely affect patient health [23,24] either because the drug is less appropriate or because the patient becomes less adherent. Similarly, a quantity limit may reduce treatment adherence, adversely affecting patient health [25]. These changes in prescribing patterns may be a fruitful avenue for future research, particularly as the prevalence of utilization management policies continues to increase.

In a mental model in which prior authorization provides targeted restrictions on use, but a quantity limit does not, one would expect a quantity limit to lead to adverse outcomes in settings in which more than one course of treatment may be required or for which the course of treatment is longer than the typical use for that drug. This hypothesis reflects, for example, the fact that a quantity limit is associated with increased non-adherence with treatment that may reduce treatment efficacy [25]. For CAP, unlike UTI, prescriptions are typically longer and, therefore, more likely to be affected by a quantity limit. These results suggest that payers should consider more flexible utilization management policies, potentially including indication-based quantity limits.

Ours’ is the first study to assess the effects of utilization management on individuals with CAP and UTI. These conditions are costly with each episode of CAP costing $8,000 (standard deviation=19,837) in 2012 dollars in the Medicare population [26], while UTIs accounted for 16 percent (SE 0.5 percent) of all infectious disease hospitalizations among older individuals [27]. Furthermore, in terms of studies on the health effects of utilization management, our study is one of the first to demonstrate a causal effect of utilization management, particularly quantity limits, on health service utilization in the Medicare population, as most studies in this domain present correlational analyses that describe the incidence of adverse ill effects among individuals receiving prescriptions subject to utilization management without estimating a causal effect [25].

For CAP claims, we found that enrolling in a plan that subjects a greater share of drugs to quantity limits increases the probably of an adverse event. We do not find any such relation for prior authorization in this sample. In the case of the UTI sample, while our results are in the same vein as the CAP results, they are much noisier and, as such, we cannot reject the null hypothesis that utilization management has no effect on the adverse event rate for UTI. Moving from utilization management exposure to utilization management treatment, in both CAP and UTI samples we find that increasing rate of prescriptions with a quantity limit affects adverse outcomes, and this result is at least significant at the 5 percent level.

The results for prior authorization in UTI are particularly interesting since they demonstrate that patients who are prescribed a drug that is subject to prior authorization have worse outcomes than patients who were not prescribed a drug subject to prior authorization. However, when we instrument for prior authorization status we find no significant differences–in other words the causal effect of receiving a prior authorization drug is 0. If, as we suspect, prior authorization policies allow insurers to target drugs better then we would expect to find no effect of prior authorization because the insurer would be denying prescriptions with a lower expected benefit. The resulting bias from the OLS estimator would, therefore, be positive due to an “adverse selection” style effect in which higher risk patients are more likely to receive drugs that are subject to prior authorization.

Given a health production framework, where the output is “fewer adverse health events,” our results may provide insight into the marginal rate of technical substitution, which measures the ease of switching between two different inputs (prescription drugs and other health services) to produce a fixed quantity of output. Assuming that utilization management increases the real price of prescription drugs, and given our findings of increased incidence of adverse health events with increased exposure and treatment to utilization management, we conclude that prescription drugs and other services are substitutes in the health production framework and suggest that regulators should consider how to optimally balance these incentives.

We also note that the effect of utilization management policies may depend on what drugs are prescribed in response to a utilization management policy. We did not model this process in this paper, but we view it as an important avenue for future research. We also did not model effects on adherence or use of appropriate treatment, both of which are important potential mechanisms by which utilization management may affect health outcomes [28–30].

Our results have several limitations. Since we only observe filled prescriptions, we cannot identify prescriptions that were sought by a patient, but rejected by the insurer. Similarly, we cannot identify prescriptions that were filled outside of the Medicare Part D benefit. We are also unable to distinguish antibiotics that were prescribed to treat the underlying UTI or CAP versus other antibiotics, but instances of multiple antibiotics in the period following an index event were rare. We also assumed that the effect of utilization management was homogeneous over time. Lastly, the prevalence of utilization management (7%) in the full sample was relatively low, however, the prevalence of prescriptions that were subject to utilization management may provide a misleading indication of the effects of utilization management since prescribing decisions may be affected by the presence of utilization management.

6. Conclusion

Understanding the health and other consequences of utilization management in the Medicare Part D program is vital given the increasing prevalence of utilization management in the Medicare Part D program [31]. We have demonstrated that some forms of utilization management may adversely affect beneficiary health, indicating that regulators and others should carefully scrutinize utilization management policies and their implementation.

Appendix

Table 5:

ICD-9 and ICD-10 codes for CAP, UTI, and exclusions

	ICD-9	ICD-10
CAP
Inclusion	481, 482.2–482.4, 482.9, 483, 485, 486	J13, J14, J1521x, J153, J154, J157, J159, J16x, J180, J181, J188, J189
Exclusion	282.41, 282.42, 282.6–282.64, 282.68, 282.69	D57x, except D573
UTI
Inclusion	590.1–590.3, 590.8, 590.9, 595.0, 595.9, 599.0	N10, N12, N151, N159, N16, N2884-N2886, N300x, N309x, N390
Exclusion	590.0, 593.7, 753	N11x, N130, N136, N137x, N139, Q600-Q639, Q641-Q6439, Q645-Q649
Universal exclusions
Diagnosis	042, 136.3, 199.2, 238.73, 238.76–238.79, 260–262, 279, 284.09, 284.1, 288.0, 288.00–288.03, 288.09, 288.1, 288.2, 288.4, 288.5, 289.53, 289.83, 403.01, 403.11, 403.91, 404.02, 404.03, 404.12, 404.13, 404.92, 404.93, 579.3, 585, 585.5, 585.6, 996.8, V42.0, V42.1, V42.6, V42.7, V42.8, V45.1, V45.11, V56.0-V56.2	B20, B59, C802, C888, C9440-C9442, C946, D4622, D470x, D471, D479, D47Zx, D6109, D6181x, D700-D702, D704-D709, D71, D720, D7281x, D7381, D7581, D76x, D80x, D810-D812, D814, D816, D817, D8189, D819, D82-D84, D893, D898x, D899, E40-E43, I120, I1311, I132, K912, N185, N186, T860x, T861x, T862-T865, T8681x, T8683x, T8685x, T8689x, T869x, Z482x, Z49x, Z940-Z944, Z948x, Z992
Procedure	0018, 335, 375, 3751, 410, 5051, 5059, 528, 5569	02YA0Z†, 0BYx†, 0FSGx, 0FYx†, 0TYx†, 0WY20Z0, 0XYJ0Z0, 0XYK0Z0, 30233AZ, 30233Gx, 30233Xx, 30233Yx, 30243AZ, 30243Gx, 30243Xx, 30243Yx, 3E03005, 3E0300M, 3E030U1, 3E030WL, 3E03305, 3E0330M, 3E033U1, 3E033WL, 3E04005, 3E0400M, 3E040WL † excludes codes with a final digit of 2

Sources—Extracted from the AHRQ PQI manuals available at https://qualityindicators.ahrq.gov/archive/pqi_techspec/icd9_v60 and https://qualityindicators.ahrq.gov/archive/pqi_techspec/icd10_v2019.

Table 6:

First stage regression for Table 2

	CAP		UTI
	Plan PA prop	Plan QL prop	Plan PA prop	Plan QL prop
Prop PA drugs in plan market	0.917***	0.104*	0.946***	−0.032*
	(0.030)	(0.048)	(0.012)	(0.015)
Prop QL drugs in plan market	−0.019	0.831***	−0.005	0.935***
	(0.014)	(0.031)	(0.007)	(0.013)
75 to 84	−0.004***	−0.008***	−0.005***	−0.006***
	(0.001)	(0.001)	(0.000)	(0.000)
85 above	−0.003**	−0.013***	−0.006***	−0.012***
	(0.001)	(0.001)	(0.000)	(0.001)
Female	0.001	−0.001	0.000	−0.001*
	(0.001)	(0.001)	(0.000)	(0.001)
Non-hisp non-wht	0.002	−0.001	0.002*	0.002*
	(0.002)	(0.002)	(0.001)	(0.001)
Observations	147,526	147,526	632,407	632,407
F-statistic on excluded instruments	481.833	59.886	2059.781	264.418

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—

^+,*,**,***indicate significance at 10%, 5%, 1%, 0.1% levels. All models include controls for: state; year; and RXHCC. Each column is from a separate regression. Standard errors clustered on beneficiary in parentheses.

Table 7:

Regression from Table 2 using 3 month window for adverse events

	Any adverse		Emergency room		Hospitalization
	OLS	IV	OLS	IV	OLS	IV
Panel A: CAP
Prop PA drugs in chosen plan	−0.007	0.087	−0.005	0.036	−0.002	0.052
	(0.007)	(0.093)	(0.003)	(0.040)	(0.006)	(0.079)
Prop QL drugs in chosen plan	0.009+	0.039	0.002	0.025	0.007	0.014
	(0.004)	(0.046)	(0.002)	(0.019)	(0.004)	(0.40)
Panel B: UTI
Prop PA drugs in chosen plan	0.001	−0.014	0.000	−0.017	0.001	0.003
	(0.003)	(0.001)	(0.001)	(0.001)	(0.002)	(0.016)
Prop QL drugs in chosen plan	0.002	0.014	0.001	−0.007	0.001	0.020
	(0.002)	(0.017)	(0.0004)	(0.012)	(0.001)	(0.012)

Sources—Authors’ analysis of Medicare claims files for 2010–2016.

Notes—

^+,*,**,***indicate significance at 10%, 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for White, and for non-Hispanic, non-White; state; year; and RXHCC. Each column in each panel is from a separate regression. “Any adverse” includes both emergency room and hospitalization adverse outcomes. IV models are estimated using the weighted average exposure to drugs subject to prior authorization (“PA”) and quantity limits (“QL”) for each beneficiary. Standard errors clustered on beneficiary in parentheses.

Table 8:

Reduced from for Table 3

	CAP		UTI
	Prescribed drug PA	Prescribed drug QL	Prescribed drug PA	Prescribed drug QL
Prop PA drugs in plan market	0.056*	0.000	0.541***	−0.047*
	(0.026)	(0.067)	(0.026)	(0.018)
Prop QL drugs in plan market	−0.024*	0.687***	−0.141***	0.695***
	(0.012)	(0.040)	(0.015)	(0.017)
75 to 84	0.002**	−0.009***	0.000	−0.003***
	(0.001)	(0.002)	(0.001)	(0.000)
85 above	0.004***	−0.016***	−0.003*	−0.005***
	(0.001)	(0.002)	(0.001)	(0.001)
Female	0.006***	−0.003*	0.023***	−0.002**
	(0.001)	(0.001)	(0.001)	(0.001)
Non-hisp non-wht	−0.001	−0.000	−0.003*	−0.002
	(0.001)	(0.003)	(0.001)	(0.001)
Observations	147,526	147,526	632,407	632,407
F	5.766	32.546	173.691	33.439

^*,**,***indicate significance at 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for white, and for non-Hispanic, non-white; state; year; and RXHCC). Each column indicates a separate regression. The dependent variables “Plan PA prop” indicates the dollar weighted proportion of PA drugs in plan. Likewise for QL. Standard errors clustered on beneficiary in parentheses.

Table 9:

First stage for adverse outcomes on prescription UM-status (Table 4)

	CAP	UTI
	Prescribed QL	Prescribed PA	Prescribed QL
Prop PA drugs in all plans	-	0.520***	−0.0668***
	-
Prop QL drugs in all plans	0.687***	−0.115***	0.659***
	(0.0382)
75 to 84	−0.00925***	0.000859	−0.00270***
	(0.00152)
85 above	−0.0155***	−0.00134	−0.00489***
	(0.00183)
Female	−0.00317*	0.0220***	−0.00198***
	(0.00143)
Non-hisp non-wht	−0.000201	−0.00332**	−0.00143
	(0.00337)
Observations	147,526	713,087	713,087
Adjusted R-squared	0.04	0.03	0.013
F-statistic	30.91	168.33	32.74

^*,**,***indicate significance at 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for white, and for non-Hispanic, non-white; state; year; and RXHCC. Each column indicates a separate regression.

Table 10:

Reduced form regression for Table 2

	CAP			UTI
	Any	ER	IPH	Any	ER	IPH
Prop PA drugs in plan market	−0.004	0.020	−0.013	−0.000	0.007	0.001
	(0.065)	(0.033)	(0.060)	(0.023)	(0.019)	(0.015)
Prop QL drugs in plan market	0.063+	0.032*	0.047	0.011	−0.003	0.014
	(0.033)	(0.016)	(0.031)	(0.015)	(0.012)	(0.010)
75 to 84	0.015***	0.001+	0.014***	0.010***	0.005***	0.006***
	(0.001)	(0.001)	(0.001)	(0.001)	(0.000)	(0.000)
85 above	0.032***	0.003**	0.032***	0.028***	0.012***	0.018***
	(0.002)	(0.001)	(0.002)	(0.001)	(0.001)	(0.001)
Female	−0.004*	−0.003***	−0.002*	−0.006***	−0.003***	−0.004***
	(0.001)	(0.001)	(0.001)	(0.001)	(0.001)	(0.000)
Non-hisp non-wht	−0.001	−0.001	−0.001	0.001	0.001	−0.000
	(0.003)	(0.001)	(0.003)	(0.001)	(0.001)	(0.001)
Observations	147,526	147,526	147,526	632,407	632,407	632,407

^*,**,***indicate significance at 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for white, and for non-hispanic, non-white; state; year; and RXHCC. Each column indicates a separate regression. “Any,” “ER,” and “IPH” indicate any avoidable adverse outcome, emergenecy room visit, or hospitalization, respectively. Numbers in parenthesis indicate standard errors clustered on the beneficiary.

Table 11:

Regression from Table 4 using 3 month window for adverse events

	Any adverse		Emergency Room		Hospitalization
	OLS	IV	OLS	IV	OLS	IV
Panel A: CAP
QL drug presc	0.011***	0.061	0.001	0.032	0.01***	0.028
	(0.003)	(0.058)	(0.001)	(0.024)	(0.003)	(0.050)
Panel B: UTI
PA drug prescr.	0.014***	−0.008	0.010***	−0.017	0.00324***	0.009
	(0.00138)	(0.0394)	(0.00108)	(0.0290)	(0.000818)	(0.0252)
QL drug presc	0.014***	0.018	0.007***	−0.011	0.007***	0.03
	(0.00191)	(0.0247)	(0.00145)	(0.0169)	(0.00121)	(0.0173)

^*,**,***indicate significance at 5%, 1%, 0.1% levels. All models include controls for: indicators for age range 65–74, 75–84, and 85 and above; female; race indicator for white, and for non-hispanic, non-white; state; year; and RXHCC(?). Each column indicates a separate regression. “Any” combines both ER and IPH as the dependent variable. OLS and IV indicate estimates from OLS and IV regressions respectively. Numbers in parenthesis indicate t-statistics. 137,856 index events in CAP sample, and 594,431 in UTI sample.