Decomposing changes in the growth of U.S. prescription drug use and expenditures, 1999‐2016

Abstract

Objective

To analyze factors associated with changes in prescription drug use and expenditures in the United States from 1999 to 2016, a period of rapid growth, deceleration, and resumed above‐average growth.

Data Sources/Study Setting

The Medical Expenditure Panel Survey (MEPS), containing household and pharmacy information on over five million prescription drug fills.

Study Design

We use nonparametric decomposition to analyze drug use, average payment per fill, and per capita expenditure, tracking the contributions over time of socioeconomic characteristics, health status and treated conditions, insurance coverage, and market factors surrounding the patent cycle.

Data Collection/Extraction Methods

Medical Expenditure Panel Survey data were combined with information on drug approval dates and patent status.

Principal Findings

Per capita utilization increased by nearly half during 1999‐2016, with changes in health status and treated conditions accounting for four‐fifths of the increase. In contrast, per capita expenditures more than doubled, with individual characteristics only explaining one‐third of the change. Other drivers of spending during this period include the changing pipeline of new drugs, drugs losing exclusivity, and changes in generic competition.

Conclusions

Long‐term trends in treated conditions were the fundamental drivers of medication use, whereas factors involving the patent cycle accelerated and decelerated spending growth relative to trends in use.

1. INTRODUCTION

As of 2016, national health expenditures totaled $3.3 trillion or 17.9 percent of gross domestic product.1 Health care spending already places considerable strain on the budgets of governments, employers, and households, and growth is predicted to continue due to changing demographics, technological innovation, and rising health care prices.2, 3 For these reasons, health services researchers have shown considerable interest in whether the recent slowing of health care spending was a temporary by‐product of recession or the result of more sustainable trends.4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15

Prescription drug spending accounted for 11.6 percent of personal health care expenditures in 20161 and 21 percent of employer health benefits in 2014,16 and its growth rate has been more volatile than that of overall spending. The nominal annual growth rate for aggregate drug spending was 14.8 percent between 1999 and 2002—far outpacing that of overall spending.1 Prescription drug spending then decelerated far more rapidly than overall spending, slowing to an average growth rate of 0.8 percent between 2009 and 2012.1 More recently, drug spending has once again accelerated, outpacing growth in overall expenditures and increasing at an average annual rate of 6.1 percent between 2012 and 2016.1 Previous research has identified a long list of factors associated with drug use and expenditures over the past two decades: changes in the pipeline of new drugs, changes in pricing among existing single‐source drugs, loss of exclusivity, and large retailer pressures on generic pricing (aka “the Walmart effect”), as well as the aging of the population, deteriorating population health, changes in insurance including the introduction of Medicare Part D, the Great Recession, and the Affordable Care Act (ACA).10, 16, 17, 18, 19, 20

Our objective is to provide an integrated analysis of how each of these factors has affected trajectories for average per capita drug fills, average payments per fill, and average per capita spending. To do so, we apply a novel decomposition method to data from the Medical Expenditure Panel Survey (MEPS) that we enhance with information on drug approvals and loss of exclusivity. Our analysis covers 1999‐2016, a time span that captures initial rapid expenditure growth, the subsequent slowdown, and the recent reacceleration. We highlight the distinct contributions of trends in both single‐source drugs (those enjoying patent protection) and multisource drugs (branded drugs that have lost patent protection and their generic competitors), with particular attention on newer versus older drugs in each of these market segments.

2. METHOD

Our nonparametric analysis has two main components. First, we examine the extent to which changes in individual characteristics explain changes in average per capita fills, average payment per fill, and per capita spending. Second, we use patent cycle information to analyze the remaining changes in fills, payments per fill, and spending, holding individual characteristics constant.

A useful starting point is regression‐based Oaxaca‐Blinder decomposition.21, 22 Let Y _i be person i's outcome (eg, number of fills or expenditures). Let X _i be a vector of K explanatory variables. If Yi can be adequately modeled as a linear function $E(Y_i)=X_i^{\prime }\beta$, then the Oaxaca‐Blinder “contribution” of variable k to the average difference in Y between period 1 and period 2 is ${\widehat{\beta }}_k(\overline{X_k^2}-\overline{X_k^1})$, where ${\widehat{\beta }}_k$ is the ordinary least‐squares (OLS) regression coefficient on X _k, estimated using either period 1 data, period 2 data, or a mix, and $\overline{X_k^t}$ is the mean of X _k in time t.

Oaxaca‐Blinder decomposition relies on OLS, which can in principle produce inefficient and/or inconsistent estimates due to the non‐negativity, mass points at zero, and skewness of use and expenditure data. Alternatively, one could estimate better‐suited parametric models, such as a hurdle negative binomial model for fills or a two‐part log‐link gamma‐distributed generalized linear model for expenditures,23 and then apply Fairlie decomposition,24, 25, 26 which involves sequentially replacing one group's explanatory variables with values from matched cases from the other group and which can be adapted for applications involving sampling weights.27

Much of the expenditure decomposition literature avoids parametric modeling by implementing simplified cell‐by‐cell or arithmetic decompositions.5, 10, 11, 12, 15 For example, to identify the effect of changes in the age distribution on spending, each age group's average expenditures in the base year are multiplied by the group's change in frequency and then summed across groups. In this highly simplified example, cell‐by‐cell decomposition is the same as Oaxaca‐Blinder decomposition (because OLS regression on a set of age group dummies yields age‐specific means). Arithmetic decomposition has the advantage of avoiding parametric assumptions. Its weakness is that it becomes highly data‐intensive in more complex models, due to the need to estimate means of the outcome for every permutation of the explanatory variables. The nonparametric method we use is closely related to cell‐by‐cell decomposition, but easily accommodates a richer set of explanatory variables.

2.1. Nonparametric “raking” decomposition

Pylypchuk and Selden's27 nonparametric “raking” method implements detailed decomposition simply by adjusting sample weights. Consider the commonplace statistical practice of standardizing, as when a common age distribution is used when comparing mortality rates across two groups. The total (unadjusted) difference can be seen as the sum of the difference removed by the weights adjustment, which in the mortality example would be the difference due to age, plus the “adjusted difference” that remains, holding age constant. In this highly simplified example, decomposition by adjusting weights is mathematically equivalent to Oaxaca‐Blinder decomposition, which relies on regression coefficients, or arithmetic decomposition, which relies on conditional means. In more complex applications, however, there can be advantages to weights‐based decomposition.

DiNardo et al28 showed that decomposition by weights adjustment can be applied to a broad class of estimation problems (in their case changes in kernel density functions over time). Pylypchuk and Selden showed how estimates for each variable's contribution to the overall difference can be obtained by progressively adjusting the sample weights to align the marginal distributions of the explanatory variables using “raking” or “iterative proportional fitting.”

Raking is the sequential and iterative adjustment of sample weights to align a dataset with a set of marginal control totals.29 To obtain the contribution to the change in Y of the change in X ₁, holding (X ₂,…, X _K) constant, we adjust the weights in the period 1 sample so as to match the period 2 (marginal) distribution for X ₁ and the period 1 (marginal) distributions for (X ₂,…, X _K). Using these adjusted weights gives us a new estimate of the mean of Y, with the difference between this mean and the unadjusted period 1 mean reflecting change in the (marginal) distribution of X ₁ holding the (marginal) distributions of all other variables constant. To get the effect of X _2, we then rake the weights to match the period 2 (marginal) distributions of X ₁ and X ₂ and the period 1 distributions for (X ₃,…, X _K). The difference between the means of Y calculated with the new set of weights versus those from the preceding step measures the change in Y associated with the change in the (marginal) distribution of X ₂ holding the (marginal) distributions of all other variables constant. We repeat these steps for all K variables in the model, with the weights in the last step aligning marginal distributions across all dimensions of X. Whereas linear detailed Oaxaca‐Blinder decomposition results are not dependent on the sequence in which the K effects are computed, raking decomposition, like any detailed nonlinear decomposition, is path‐dependent. Our estimates are not, however, sensitive to computing the effects in reverse order.

Raking decomposition is easily implemented, can be applied to a broad range of estimation problems, and imposes few assumptions. In practice, however, Oaxaca‐Blinder decomposition estimates for average fills per person and average expenditure (Appendix S1) closely mirror our corresponding results from raking decomposition—perhaps not surprisingly given that all of the explanatory variables are categorical indicators (helping to reduce OLS misspecification).

2.2. Decomposing changes in average payment per fill

We seek to examine changes not only in the average number of fills per person and the average expenditure per person, but also in the average payment per fill, computed as the ratio of average expenditures divided by the average number of fills. Oaxaca‐Blinder decomposition is not directly applicable to analysis of ratios of conditional expectations, whereas this can be easily accomplished with raking decomposition.

2.3. Analysis of adjusted differences

Changes in individual characteristics will in general “explain” only a portion of unadjusted changes over time, whether one employs regression‐based Oaxaca‐Blinder decomposition or decomposition by reweighting. The analysis therefore results in residual or adjusted differences. The second portion of our decomposition uses the adjusted weights to separate these remaining differences in fills, payment per fill, and spending into effects associated with the following four drug categories: (a) newer single‐source drugs, (b) older single‐source drugs, (c) newer multisource drugs (drugs that recently lost exclusivity), and (d) older multisource drugs. We draw an essentially arbitrary line between newer single‐source drugs, in their first four years following U.S. Food and Drug Administration (FDA) approval, versus older single‐source drugs. This distinction enables us to separate use and expenditure changes due to the “pipeline” of new molecules entering the market versus all other changes affecting the single‐source market.30 We also make a similar distinction between newer and older multisource drugs based upon research showing that when drugs lose exclusivity, prices for branded pioneers and their generic copies typically adjust only gradually to longer‐term equilibrium levels.30, 31, 32 Based on estimated trajectories from that research, we allow three years for transition from single source to older multisource. This enables us to distinguish between compositional effects due to patent expiration (as high‐priced drugs enter the multisource market) versus changes in the more “mature” segment of the multisource market, such as occurred starting in 2006 when Walmart and some competitors offered selected generic drugs for $4 per month.33, 34 This distinction also helps disentangle the recent impacts of fewer blockbuster drugs coming off patent versus changes in competition among generic producers of older multisource drugs.32, 35, 36, 37

3. DATA

Data for our study come from the 1999‐2016 MEPS Household Component. MEPS is a survey of the civilian noninstitutionalized population, sponsored by the Agency for Healthcare Research and Quality, that offers several advantages over more commonly used insurance claims or pharmacy data. Unlike claims data, MEPS is an all‐payer database, and unlike pharmacy data, MEPS contains a rich array of socioeconomic, health coverage, and health status measures. Another advantage is that MEPS contains information on people whether or not they consume drugs. Although households can misreport detailed information regarding their prescription drug use, MEPS supplements household reports with pharmacy information including the National Drug Code (NDC) and payment by source.

Pooling across years, we have data on 609 073 person‐years and 5.4 million prescription fills. To highlight the periods of rapid versus slower growth, we examine adjacent pairs of years (1999/2000 to 2001/2002, 2001/2002 to 2003/2004, et cetera), with the resulting effects summing to the overall change from 1999/2000 to 2015/2016. We use CPI‐U to adjust all expenditures to 2016 dollars. To account for the clustered design of MEPS and any intrafamily correlation, standard errors were computed using balanced repeated replicates.38

3.1. Prescription drug data

We apply refined editing rules to MEPS public use prescription drug data, modestly increasing fill sizes for 2000‐2006 and removing caps, starting in 2011, on imputed payments for certain specialty drugs. MEPS fills were linked by NDC to data on active ingredients, route, form, and strength from Cerner Multum and single‐source/multisource data from the Master Drug Database.39 We also merged on FDA data regarding drug approval dates. To distinguish between new versus older single‐source drugs, we use the earliest FDA approval date for the drug's active ingredient or, for multi‐ingredient drugs, the earliest date of the most recently approved molecule. We do not count as new drugs any previously approved molecules with new patent protection due to changes in route or form (such as delayed release versions of a previously approved molecule). Finally, we standardized fill sizes (eg, pills per fill) by drug name, route, form, and strength to adjust for gradually rising average fill sizes over the study period. Doing so modestly increases growth in the number of (adjusted) fills per capita (see Appendix S2).

3.2. Conditions

Condition information in MEPS is recorded as verbatim text and then assigned 3‐digit ICD9 codes (in years 1999‐2015) or ICD10 codes (in 2016). To minimize possible impacts of questionnaire changes over time, we used data on treated conditions (ie, those for which individuals received any prescription drug or other medical care during the year). We include information on both acute and chronic conditions, constructing a set of 18 condition indicators based primarily on major chapters of the ICD9/10 code. To model patterns of comorbidity as flexibly as possible, we create indicator variables for all combinations of one or more conditions that occur in our sample with sufficient frequency to support reliable estimation (see Appendix S3).

3.3. Other explanatory variables

Our analysis includes a hierarchical measure of hospital/physician coverage: whether the person ever had private coverage during the year, else if they ever had public coverage, with the remainder having been uninsured the full year. The decomposition also includes indicators for whether the person ever held public or private prescription drug coverage during the year, as well as measures of age, sex, race‐ethnicity, income (as a percentage of the federal poverty line (FPL)), fair/poor health status, region, and urbanicity.

3.4. Limitations

Medical Expenditure Panel Survey drug data do not capture drugs administered in physician offices, clinics, or hospitals, nor drugs used by individuals in nursing homes and other institutions. MEPS also does not capture drug rebates paid by pharmaceutical companies to third‐party payers. MEPS is not likely to capture illegal use of prescription drugs or other drug use that might be sensitive for households to report. Nevertheless, Appendix S4 shows that MEPS growth rates for prescription drug spending mirror those in National Health Expenditure Accounts1 and IMS.19 Moreover, an analysis of linked Medicare Part D prescription claims data found that although MEPS respondents tended both to underreport the number of different medications taken and to over‐report the number of fills per reported medication, these errors had little impact on behavioral analyses (comparing administrative versus survey‐reported prescription data).40

4. RESULTS

Table 1 presents unadjusted means by two‐year periods, for per capita fills, average payment per fill, per capita expenditures, and the model's explanatory variables. The average number of (size‐adjusted) fills per capita rose by 48.5 percent during the study period, from 6.87 to 10.21. Average payment per fill, calculated as total expenditure divided by total (size‐adjusted) fills, increased from $73 to $124 (in 2016$), an increase of 69.4 percent. Per capita spending rose from $503 to $1265 (in 2016$), an increase of 151.3 percent. Expenditure growth was fastest between 1999/2000 and 2001/2002, rising at an average annual rate of 15.3 percent. It then slowed to an average annual rate of 1.6 percent between 2005/2006 and 2011/2012 and ended the study period with an average annual growth rate of 6.5 percent between 2011/2012 and 2015/2016.

Table 1.

Means of per capita drug fills, payment per fill, per capita expenditures, and explanatory variables, 1999/2000 to 2015/2016

Variables	1999/2000	2001/2002	2003/2004	2005/2006	2007/2008	2009/2010	2011/2012	2013/2014	2015/2016
Per capita number of size‐adjusted fills	6.87 (0.16)	8.20 (0.12)	8.86 (0.13)	9.28 (0.16)	9.14 (0.15)	9.71 (0.17)	10.01 (0.17)	10.28 (0.18)	10.21 (0.18)
Newer single source	0.59 (0.02)	0.50 (0.01)	0.52 (0.01)	0.41 (0.02)	0.21 (0.01)	0.12 (0.01)	0.12 (0.01)	0.08 (0.01)	0.11 (0.01)
Older single source	2.44 (0.06)	3.31 (0.05)	3.24 (0.05)	2.96 (0.05)	2.50 (0.05)	2.33 (0.05)	1.79 (0.04)	1.46 (0.03)	1.20 (0.03)
Newer multisource	0.24 (0.01)	0.59 (0.01)	1.01 (0.02)	1.28 (0.03)	1.29 (0.02)	1.17 (0.02)	0.90 (0.02)	0.88 (0.02)	0.50 (0.01)
Older multisource	3.60 (0.09)	3.80 (0.07)	4.09 (0.06)	4.63 (0.09)	5.14 (0.09)	6.09 (0.11)	7.20 (0.13)	7.86 (0.14)	8.40 (0.15)
Average payment per size‐adjusted fill (in 2016$)	73 (1)	82 (1)	92 (1)	97 (2)	98 (1)	98 (2)	98 (2)	107 (3)	124 (4)
Newer single source	151 (3)	163 (3)	163 (3)	187 (7)	296 (17)	265 (15)	306 (21)	957 (192)	1525 (196)
Older single source	97 (1)	113 (1)	138 (1)	167 (5)	217 (3)	261 (5)	329 (11)	399 (24)	500 (19)
Newer multisource	111 (3)	97 (1)	97 (1)	100 (1)	101 (1)	106 (2)	155 (3)	160 (5)	251 (6)
Older multisource	42 (0)	41 (0)	45 (0)	42 (0)	32 (0)	31 (0)	30 (0)	39 (1)	45 (1)
Per capita expenditure on prescription drugs (in 2016$)	503 (14)	669 (11)	812 (14)	896 (24)	900 (18)	953 (22)	983 (29)	1104 (38)	1265 (44)
Newer single source	89 (3)	81 (2)	85 (3)	77 (4)	62 (4)	33 (3)	37 (3)	77 (17)	165 (24)
Older single source	238 (7)	375 (7)	445 (8)	495 (19)	543 (12)	606 (18)	589 (23)	581 (36)	601 (28)
Newer multisource	27 (1)	57 (1)	99 (2)	127 (3)	130 (3)	123 (3)	139 (5)	140 (6)	124 (5)
Older multisource	150 (4)	155 (4)	183 (3)	196 (5)	165 (4)	191 (4)	218 (5)	305 (7)	374 (8)
Explanatory variables
Female (%)	51.2 (0.3)	51.2 (0.2)	51 (0.2)	51 (0.2)	51 (0.2)	50.9 (0.2)	51.1 (0.2)	51.1 (0.3)	51.1 (0.3)
Age 0‐20 (%)	30.5 (0.4)	29.7 (0.3)	29.2 (0.3)	28.9 (0.3)	28.7 (0.4)	28.6 (0.4)	27.9 (0.4)	27.5 (0.4)	27.1 (0.4)
Age 21‐64 (%)	57.0 (0.4)	57.7 (0.3)	58.1 (0.3)	58.4 (0.3)	58.4 (0.3)	58.2 (0.4)	57.9 (0.4)	57.6 (0.4)	57.2 (0.4)
Age 65+ (%)	12.5 (0.4)	12.6 (0.3)	12.6 (0.3)	12.7 (0.3)	12.9 (0.3)	13.2 (0.4)	14.2 (0.4)	14.9 (0.4)	15.7 (0.4)
White non‐Hispanic (%)	71.3 (1.3)	68.8 (0.8)	67.2 (0.8)	66.1 (1.0)	65.4 (0.8)	64.6 (1.0)	63.4 (1.1)	61.7 (1.1)	60.3 (1.1)
Black non‐Hispanic (%)	12.6 (1.0)	12.2 (0.5)	12.1 (0.5)	12.1 (0.6)	12.2 (0.6)	12.1 (0.7)	12 (0.6)	12.1 (0.7)	12.3 (0.6)
Hispanic (%)	12.1 (1.1)	13.5 (0.7)	14.2 (0.6)	14.9 (0.7)	15.6 (0.7)	16.2 (0.9)	17 (1.0)	17.4 (0.9)	17.9 (0.9)
Other non‐Hispanic (%)	4.0 (0.4)	5.5 (0.3)	6.5 (0.3)	6.9 (0.4)	6.8 (0.4)	7.1 (0.5)	7.5 (0.5)	8.8 (0.6)	9.6 (0.6)
<125% Federal poverty line (%)	16.1 (0.6)	16.5 (0.4)	17.2 (0.4)	17.2 (0.5)	17.5 (0.4)	19.4 (0.5)	20 (0.5)	19.5 (0.6)	17.5 (0.6)
125%‐199% Federal poverty line (%)	13.7 (0.4)	14 (0.3)	14.1 (0.3)	13.7 (0.3)	13.7 (0.3)	14 (0.3)	14.3 (0.3)	14.2 (0.3)	13.4 (0.3)
200%‐399% Federal poverty line (%)	31.8 (0.6)	31.6 (0.4)	31.4 (0.4)	31.3 (0.5)	31.1 (0.4)	30.4 (0.5)	30.2 (0.5)	29.3 (0.5)	28.8 (0.5)
400%+ Federal poverty line (%)	38.4 (0.9)	37.9 (0.7)	37.3 (0.6)	37.9 (0.7)	37.7 (0.6)	36.2 (0.7)	35.5 (0.8)	37.1 (0.7)	40.2 (0.7)
Any private insurance (%)	73.3 (0.8)	71.7 (0.6)	69.9 (0.5)	69.1 (0.6)	66.6 (0.6)	65.3 (0.7)	64.8 (0.8)	64.6 (0.8)	66.9 (0.7)
Only public insurance (%)	15.4 (0.6)	16.6 (0.4)	17.9 (0.4)	18.5 (0.4)	20 (0.4)	21.3 (0.5)	22.5 (0.6)	24.0 (0.6)	25.2 (0.6)
Uninsured (%)	11.3 (0.4)	11.7 (0.3)	12.2 (0.3)	12.5 (0.3)	13.4 (0.3)	13.3 (0.4)	12.6 (0.4)	11.3 (0.4)	7.9 (0.3)
Rx coverage, private (%)	66.1 (0.8)	65.5 (0.6)	64.5 (0.6)	64.3 (0.6)	62.2 (0.6)	60.7 (0.7)	60.4 (0.7)	60.6 (0.8)	63.0 (0.7)
Rx coverage, public (%)	12.3 (0.6)	13.4 (0.4)	14.5 (0.3)	16.2 (0.4)	18.3 (0.4)	19.9 (0.5)	21.1 (0.6)	22.6 (0.6)	23.7 (0.6)
No Rx coverage (%)	21.6 (0.6)	21.1 (0.4)	21 (0.4)	19.6 (0.4)	19.5 (0.4)	19.5 (0.5)	18.5 (0.4)	16.8 (0.4)	13.4 (0.4)
MSA (%)	81.3 (2.0)	81.5 (0.8)	82.1 (0.8)	83.5 (1.1)	84.0 (1.0)	84.2 (1.2)	85.2 (1.2)	85.4 (1.2)	86.1 (1.1)
Northeast (%)	18.8 (2.3)	18.8 (1.0)	18.6 (0.9)	18.4 (1.0)	18.0 (0.7)	18.0 (0.7)	17.8 (0.7)	17.8 (0.6)	17.5 (0.8)
Midwest (%)	23.2 (2.6)	22.7 (1.0)	22.4 (0.9)	22.1 (1.1)	21.9 (0.7)	21.7 (0.6)	21.4 (0.7)	21.4 (0.7)	21.0 (0.8)
South (%)	35.1 (3.2)	35.6 (1.3)	36.0 (1.2)	36.3 (1.4)	36.7 (0.9)	36.9 (0.9)	37.3 (0.9)	37.3 (0.9)	37.7 (1.0)
West (%)	22.9 (3.3)	22.8 (1.2)	23.0 (1.4)	23.2 (1.5)	23.3 (0.8)	23.4 (0.8)	23.5 (0.7)	23.6 (0.7)	23.8 (0.7)
Fair/poor health (%)	16.6 (0.4)	17.6 (0.3)	17.9 (0.3)	18.0 (0.4)	18.2 (0.3)	18.2 (0.3)	18.6 (0.3)	18.0 (0.3)	17.6 (0.3)
No treated ICD‐9 conditions (%)	32.7 (0.5)	30.7 (0.4)	31.2 (0.3)	31.6 (0.4)	32.2 (0.4)	31.9 (0.3)	31.4 (0.4)	32.2 (0.4)	32.4 (0.4)
1 treated ICD‐9 condition (%)	26.4 (0.3)	25.0 (0.2)	24.4 (0.2)	24.2 (0.2)	23.4 (0.2)	23.0 (0.2)	22.9 (0.3)	22.0 (0.3)	21.7 (0.2)
2 treated ICD‐9 conditions (%)	17.2 (0.3)	17.5 (0.2)	16.8 (0.2)	16.5 (0.2)	16.1 (0.2)	15.8 (0.2)	15.9 (0.2)	15.1 (0.2)	14.9 (0.2)
3+ treated ICD‐9 conditions (%)	23.7 (0.5)	26.8 (0.4)	27.6 (0.4)	27.7 (0.4)	28.3 (0.4)	29.4 (0.4)	29.9 (0.4)	30.7 (0.4)	31.0 (0.4)

Source: Medical Expenditure Panel Survey Household Component, 1999‐2016.

Table 1 also presents averages for fills, payment per fill, and spending for each of the four drug categories. These unadjusted trends presage some of the main results, and to save space, we focus attention on the adjusted results (presented below).

Among the notable trends in Table 1 is the rapid increase in percentage age 65 and over that began after 2009/2010, as the first baby boomers turned 65. Also notable are the increase in percentage Hispanic, from 12.1 percent to 17.9 percent, and the changing population share below 125 percent FPL, which rose from 16.1 percent to 19.5 percent in the aftermath of recession, before declining to 17.5 percent in 2015/2016. Over the study period, the share with any private insurance declined from 73.3 percent to 66.9 percent, versus an increase in the share with only public insurance from 15.4 percent to 25.2 percent. Uninsurance first rose and then, after 2009/2010, declined as the economy recovered from recession and as the ACA insurance provisions were implemented. The prevalence of public drug coverage rose steadily, from 12.3 percent to 23.7 percent, reflecting expanded public coverage for children, Medicare Part D, the shift toward Medicare Advantage, and expanded public coverage under the ACA. As a result, the percentage lacking any drug coverage dropped from 21.6 to 13.4 percent, despite modest reductions in private drug coverage.

Finally, Table 1 presents estimates of the number of treated conditions (rather than the more detailed condition groupings in our decomposition analysis). The frequency of having zero treated conditions remained steady over the study period and frequencies for having one or two treated conditions declined, whereas an important trend for our analysis is the monotonic increase in the prevalence of multiple (3 + ) treated conditions from 23.7 percent to 31.0 percent over the study period. This likely reflected a combination of the aging of the population, lifestyle factors such as increased obesity, increased diagnosis of existing conditions, and more treatment (with drugs or other medical care) of diagnosed conditions.41

4.1. Decomposition of average fills per person

Individual‐level characteristics explained 2.95 (or 88.6 percent) of the 3.33 increase in average fills per capita (Table 2). Health status and treated conditions accounted for about four‐fifths of the total increase, while changes in health insurance and drug coverage accounted for a little more than one‐tenth of the total increase. Socioeconomic factors did not have a statistically significant effect on fills.

Table 2.

Decomposition of changes in per capita number of fills, 1999/2000 to 2015/2016

	Initial per capita fills	Contribution of individual characteristics				Adjusted change by drug type					Total change
		Socioeconomic factorsa	Insurance and drug coverage	Health status and conditions	All individual characteristics	Newer single source	Older single source	Newer multisource	Older multisource	All drugs
All years	6.87 (0.15)	−0.08 (0.08)	0.35 (0.03)	2.68 (0.21)	2.95 (0.25)	−0.64 (0.02)	−2.12 (0.08)	−0.02 (0.03)	3.15 (0.15)	0.38 (0.24)	3.33 (0.24)
99/00 to 01/02	6.87 (0.15)	−0.01 (0.02)	0.02 (0.01)	0.80 (0.11)	0.81 (0.12)	−0.18 (0.02)	0.57 (0.05)	0.33 (0.01)	−0.21 (0.05)	0.51 (0.09)	1.32 (0.16)
01/02 to 03/04	8.20 (0.12)	−0.01 (0.01)	0.03 (0.01)	0.48 (0.10)	0.50 (0.11)	−0.01 (0.02)	−0.25 (0.05)	0.38 (0.02)	0.05 (0.06)	0.17 (0.11)	0.67 (0.15)
03/04 to 05/06	8.86 (0.12)	−0.02 (0.02)	0.03 (0.01)	0.29 (0.14)	0.30 (0.15)	−0.13 (0.01)	−0.39 (0.04)	0.22 (0.03)	0.42 (0.06)	0.12 (0.10)	0.42 (0.19)
05/06 to 07/08	9.28 (0.16)	0.00 (0.02)	0.05 (0.02)	0.22 (0.13)	0.27 (0.14)	−0.21 (0.01)	−0.52 (0.05)	−0.02 (0.03)	0.34 (0.09)	−0.41 (0.16)	−0.14 (0.20)
07/08 to 09/10	9.14 (0.14)	0.01 (0.07)	0.05 (0.02)	0.27 (0.12)	0.33 (0.15)	−0.09 (0.01)	−0.23 (0.05)	−0.16 (0.03)	0.72 (0.09)	0.24 (0.14)	0.57 (0.17)
09/10 to 11/12	9.71 (0.17)	0.00 (0.01)	0.05 (0.01)	0.33 (0.12)	0.38 (0.13)	−0.01 (0.01)	−0.62 (0.04)	−0.32 (0.03)	0.87 (0.09)	−0.08 (0.12)	0.30 (0.15)
11/12 to 13/14	10.01 (0.17)	−0.02 (0.01)	0.06 (0.01)	0.26 (0.14)	0.29 (0.15)	−0.04 (0.01)	−0.40 (0.04)	−0.06 (0.02)	0.48 (0.11)	−0.03 (0.14)	0.26 (0.19)
13/14 to 15/16	10.28 (0.18)	−0.03 (0.01)	0.08 (0.02)	0.03 (0.13)	0.07 (0.13)	0.03 (0.01)	−0.27 (0.03)	−0.40 (0.02)	0.49 (0.11)	−0.14 (0.13)	−0.07 (0.18)

^aSocioeconomic factors include age, sex, race‐ethnicity, family income, Census region, and MSA status.

Source: Medical Expenditure Panel Survey Household Component, 1999‐2016. Numbers of fills have been adjusted to hold fill sizes constant at their 2016 levels.

Examining period‐by‐period effects, insurance and drug coverage changes became somewhat more important over time, although these changes were associated with less than a tenth of a fill increase in every period. We do not observe substantial socioeconomic effects in any period, even due to income changes during the Great Recession and its aftermath. Perhaps the most noteworthy period‐by‐period pattern we observe is that the effects of health status and treated conditions diminished over time—a trend that may help dampen future expenditure growth.

Holding individual characteristics constant, the residual or adjusted change was 0.38 fills per capita. This effect, while small and not statistically significant at the 5 percent level, was the net effect of larger, offsetting changes in fill mix. Holding individual characteristics constant, the per capita consumption of newer single‐source drugs declined by 0.64 fills, and consumption of older single‐source drugs declined by 2.12 fills, reflecting a shrinking of the pipeline of new drugs combined with a large number of drugs losing exclusivity. Both declines were large relative to initial levels (see Table 1). Declines in older single‐source drugs were particularly large between 2005/2006 and 2007/2008 (a 0.52 fill decline) and between 2009/2010 and 2011/2012 (a 0.62 fill decline). These adjusted declines in single‐source fills were more than offset by adjusted increases in the use of multisource drugs, first (from 1999/2000 through 2005/2006) in the newer multisource category and subsequently in the older multisource category. Over the entire period, the adjusted increase in per capita use of older multisource drugs was 3.15 fills, relative to a base of 3.60 fills (Table 1). Not surprisingly, this shift in composition from single‐source to older multisource drugs had important consequences for average payment per fill and average expenditures.

4.2. Decomposition of average payment per fill

Table 3 presents decomposition results for average payments per fill by drug category. Changes in individual characteristics had little impact on average payments (there is little difference between the initial and adjusted estimates). Most noteworthy is the trend after 2011/2012 toward substantially higher average payments per fill for newer single‐source drugs. Driven by the introduction of a number of highly expensive drugs, including new treatments for Hepatitis C, cancer, and multiple sclerosis, average payment per fill for new drugs experienced nearly a 4‐fold increase between 2011/2012 and 2015/2016, from $306 to $1525. We also see accelerating average payments for older single‐source drugs, with the increase in average payments in the last period ($99) being approximately the same as the level of average payments in the first period ($97).

Table 3.

Changes in average payment per fill, by drug type (in 2016 dollars), 1999/2000 to 2015/2016

	Initial payment per fill	Adjusted payment per fill	End payment per fill	Adjusted difference
Newer single source
99/00 to 01/02	151 (3)	153 (3)	163 (3)	10 (4)
01/02 to 03/04	163 (3)	163 (2)	163 (3)	0 (4)
03/04 to 05/06	163 (3)	161 (3)	187 (7)	25 (7)
05/06 to 07/08	187 (7)	186 (7)	296 (17)	110 (18)
07/08 to 09/10	296 (17)	282 (23)	265 (15)	−17 (27)
09/10 to 11/12	265 (15)	269 (16)	306 (21)	37 (24)
11/12 to 13/14	306 (21)	308 (22)	957 (192)	649 (193)
13/14 to 15/16	957 (192)	958 (191)	1525 (196)	567 (220)
Older single source
99/00 to 01/02	97 (1)	99 (1)	113 (1)	14 (2)
01/02 to 03/04	113 (1)	114 (1)	138 (1)	23 (2)
03/04 to 05/06	138 (1)	138 (1)	167 (5)	29 (6)
05/06 to 07/08	167 (5)	168 (5)	217 (3)	49 (6)
07/08 to 09/10	217 (3)	219 (3)	261 (5)	41 (5)
09/10 to 11/12	261 (5)	261 (5)	329 (11)	67 (13)
11/12 to 13/14	329 (11)	328 (11)	399 (24)	72 (25)
13/14 to 15/16	399 (24)	401 (24)	500 (19)	99 (20)
Newer multisource
99/00 to 01/02	111 (3)	112 (3)	97 (1)	−14 (3)
01/02 to 03/04	97 (1)	98 (1)	97 (1)	−1 (2)
03/04 to 05/06	97 (1)	97 (1)	100 (1)	2 (2)
05/06 to 07/08	100 (1)	99 (1)	101 (1)	1 (2)
07/08 to 09/10	101 (1)	101 (2)	106 (2)	5 (2)
09/10 to 11/12	106 (2)	106 (2)	155 (3)	49 (4)
11/12 to 13/14	155 (3)	155 (3)	160 (5)	5 (5)
13/14 to 15/16	160 (5)	160 (5)	251 (6)	90 (8)
Older multisource
99/00 to 01/02	42 (0)	42 (0)	41 (0)	−1 (1)
01/02 to 03/04	41 (0)	41 (0)	45 (0)	4 (1)
03/04 to 05/06	45 (0)	45 (0)	42 (0)	−2 (0)
05/06 to 07/08	42 (0)	42 (0)	32 (0)	−10 (1)
07/08 to 09/10	32 (0)	33 (1)	31 (0)	−1 (1)
09/10 to 11/12	31 (0)	31 (0)	30 (0)	−1 (1)
11/12 to 13/14	30 (0)	30 (0)	39 (1)	8 (1)
13/14 to 15/16	39 (1)	39 (1)	45 (1)	5 (1)

Source: Medical Expenditure Panel Survey Household Component, 1999‐2016. Payments are per fill, where fills have been adjusted to hold fill sizes constant at their 2016 levels.

Table 3 also shows large differences in average payments between newer and older multisource drugs and very different trajectories over time. Average payments for newer multisource fills increased by 125.9 percent from $111 to $251, with most of the growth occurring after 2009/2010. Average payments for older multisource drugs were relatively stable through 2005/2006 and then dropped by approximately one‐quarter (from $42 to $32) between 2005/2006 and 2007/2008. This drop coincided with the advent of $4 generic prescriptions in Walmart and some of its competitors that began in late 2006.42 No other drug categories experienced any average payment decline during that period. Starting in 2011/2012, average payments for older multisource fills rose quite sharply in percentage terms, returning to levels in line with those at the start of the analysis.

To place these results in context, Table 4 presents the overall decomposition of average payments—combining the impact of average payment changes in Table 3, weighted by the mix of drug types, with smaller effects due to changes in the mix of drug types. Rapid increases in payment per fill for newer single‐source drugs had a muted effect on the overall average payment per fill, because new drug fills made up only a small (and declining) share of all fills. The much smaller average payment increases for older multisource drugs had a comparable overall effect, because these drugs comprised a much larger (and growing) share of the total. Together, newer single‐source and older multisource drugs accounted for a $28 increase in average payment per fill between 2011/2012 and 2015/2016 (relative to a base in 2011/2012 of $98).

Table 4.

Decomposition of changes in average payment per fill (in 2016 dollars), 1999/2000 to 2015/2016

	Initial expenditure per fill	Contribution of individual characteristics				Adjusted change by drug type					Total change
		Socioeconomic factorsa	Insurance and drug coverage	Health status and conditions	All individual characteristics	Newer single source	Older single source	Newer multisource	Older multisource	All drugs
All years	73 (1)	0 (1)	0 (0)	2 (1)	3 (1)	3 (2)	23 (3)	8 (1)	14 (1)	48 (4)	51 (4)
99/00 to 01/02	73 (1)	0 (0)	0 (0)	1 (0)	1 (0)	−4 (0)	11 (1)	3 (0)	−3 (0)	7 (1)	8 (1)
01/02 to 03/04	82 (1)	0 (0)	0 (0)	1 (0)	1 (0)	0 (0)	4 (1)	4 (0)	1 (0)	9 (1)	10 (1)
03/04 to 05/06	92 (1)	0 (0)	0 (0)	0 (0)	0 (0)	−1 (0)	3 (2)	2 (0)	1 (0)	5 (2)	5 (2)
05/06 to 07/08	97 (2)	0 (0)	0 (0)	0 (0)	0 (0)	−1 (1)	6 (2)	1 (0)	−3 (0)	2 (2)	2 (2)
07/08 to 09/10	98 (1)	0 (1)	0 (0)	0 (0)	0 (1)	−3 (1)	3 (2)	−2 (0)	1 (0)	0 (2)	0 (2)
09/10 to 11/12	98 (2)	0 (0)	0 (0)	0 (0)	0 (0)	0 (0)	−4 (3)	1 (0)	2 (0)	0 (3)	0 (3)
11/12 to 13/14	98 (2)	0 (0)	0 (0)	0 (0)	0 (1)	4 (2)	−3 (4)	0 (1)	8 (0)	9 (4)	9 (4)
13/14 to 15/16	107 (4)	1 (1)	1 (0)	−1 (1)	1 (1)	9 (2)	2 (3)	−2 (1)	7 (1)	16 (5)	16 (4)

^aSocioeconomic factors include age, sex, race‐ethnicity, family income, Census region, and MSA status.

Source: Medical Expenditure Panel Survey Household Component, 1999‐2016.

4.3. Decomposition of per capita expenditures

Table 5 presents decomposition results for per capita drug expenditures. Individual characteristics accounted for 37.5 percent ($285) of the $761 total change. As was the case with average fills, health status and treated conditions explained most of the expenditure change attributed to individual characteristics.

Table 5.

Decomposition of changes in per capita expenditure (in 2016 dollars), 1999/2000 to 2015/2016

	Initial per capita expenditure	Contribution of individual characteristics				Adjusted change by drug type					Total change
		Socioeconomic factorsa	Insurance and drug coverage	Health status and conditions	All individual characteristics	Newer Single Source	Older single source	Newer multisource	Older multisource	All drugs
All years	503 (13)	−3 (7)	35 (4)	253 (22)	285 (23)	51 (23)	205 (30)	65 (5)	155 (9)	476 (46)	761 (48)
99/00 to 01/02	503 (13)	0 (2)	2 (1)	66 (9)	67 (9)	−22 (4)	104 (8)	28 (2)	−13 (4)	98 (12)	165 (16)
01/02 to 03/04	669 (12)	−1 (1)	2 (1)	45 (9)	46 (9)	−1 (3)	45 (9)	37 (3)	16 (3)	98 (12)	144 (16)
03/04 to 05/06	812 (13)	−1 (2)	1 (1)	29 (13)	29 (15)	−10 (4)	33 (18)	24 (3)	8 (4)	55 (21)	84 (28)
05/06 to 07/08	896 (25)	−1 (2)	1 (2)	24 (13)	24 (15)	−17 (5)	34 (21)	0 (4)	−38 (5)	−20 (26)	4 (29)
07/08 to 09/10	900 (18)	−3 (2)	4 (1)	30 (11)	31 (12)	−28 (6)	46 (16)	−11 (4)	16 (5)	23 (19)	53 (23)
09/10 to 11/12	953 (22)	1 (2)	5 (1)	34 (12)	39 (13)	2 (4)	−42 (26)	11 (5)	20 (5)	−9 (30)	30 (29)
11/12 to 13/14	983 (27)	−1 (2)	7 (2)	28 (14)	34 (14)	39 (17)	−27 (40)	−5 (6)	80 (6)	87 (45)	121 (47)
13/14 to 15/16	1104 (45)	5 (5)	13 (2)	−2 (16)	16 (15)	87 (24)	12 (34)	−19 (6)	65 (8)	145 (50)	160 (53)

^aSocioeconomic factors include age, sex, race‐ethnicity, family income, Census region, and MSA status.

Source: Medical Expenditure Panel Survey Household Component, 1999‐2016.

Adjusted per capita spending on newer single‐source drugs increased by $51 over the study period, initially declining with the diminishing pipeline of new molecules and then rising in the last two periods due to sharply rising average payments per fill. Older single‐source drug spending increased by $205 per person, relative to a base in 1999/2000 of $238 (Table 1). This increase was driven by steadily accelerating payments per fill (Table 3) that more than offset the 2.12 decrease in per capita fills (Table 2).

Adjusted average spending on older multisource drugs increased by $155 relative to a base of $150 (Table 1). Once again, we may be seeing the effect of price cuts at major drug retailers, with older multisource spending declining $38 per capita between 2005/2006 and 2007/2008. Also noteworthy were older multisource spending increases starting in 2011/2012, stemming from both abruptly rising average payments and the ongoing, albeit slowing, upward trend in the average number of older multisource fills.

5. DISCUSSION

By combining the richness of MEPS with patent cycle information, our paper presents an integrated analysis of factors driving prescription drug spending growth. Increasing prevalence of treated conditions and declining self‐reported health status explained approximately four‐fifths of the per capita increase in fills and one‐third of the per capita increase in spending. Of less importance were the marked shift toward public coverage (and away from private coverage or uninsurance); strong increases in prescription drug coverage; shifts in the income distribution associated with macroeconomic conditions; and aging (except insofar as it affected health).

Our findings mirror those of Blavin et al,10 who found that between 2001 and 2009 chronic conditions accounted for 35.2 percent of the growth in real per capita drug expenditures, versus an additional 4 percent attributable to other individual characteristics. During a similar period, from 2001/2002 to 2009/2010, we find health status and all treated conditions (not just chronic) explained 45.0 percent of expenditure growth, with all other individual characteristics contributing less than one percent. In contrast, Dieleman et al43 found that age, sex and underlying or true prevalence of conditions together explained less than 10 percent of per capita prescription drug growth between 1996 and 2013.

The relationship we observe between health status and treated conditions and prescription drug spending is quite similar to corresponding results found in the literature decomposing total health care spending. Starr et al12 found that treated conditions explained 36.5 percent of per capita health care spending growth between 1996 and 2006, and Thorpe11 found that treated conditions explained 34.7 percent of per capita expenditure growth between 1987 and 2009. We find health status and treated conditions explained 35.6 percent of per capita spending growth for prescription drugs from 1999/2000 to 2005/2006. Roehrig and Rousseau5 found that treated conditions explained a somewhat lower share (26.3 percent) of overall per capita spending growth between 1996 and 2006.

Health status and treated conditions were the primary drivers of drug use not only over the entire time span, but also in our period‐by‐period analysis, with diminishing health status and treated condition effects helping to slow the growth in fills per capita over time. Roehrig and Daly40 show that treated prevalence increased more rapidly than underlying or true prevalence for diabetes, hypertension, and hyperlipidemia between 1999/2000 and 2011/2012. More research is needed to assess whether trends such as these explain why Dieleman et al43 found smaller impacts of underlying conditions compared to studies examining treated conditions—and whether convergence of treated and underlying conditions may now be helping to limit further condition‐driven increases in use to the growth in underlying condition prevalence. A better understanding of these forces may help us to better predict future prescription drug spending growth and to devise policies that would help control that growth.

Whereas individual characteristics explain most of the change in drug use, they explain almost none of the changes in average payment per fill and only a third of average expenditures. Moreover, because individual characteristics tended to exhibit relatively smooth patterns of change during the period of study, none was a strong predictor of fluctuations in spending growth. By incorporating information on the timing of FDA approvals and losses of exclusivity, we have been able to separate the remaining, adjusted changes into four patent cycle categories: newer and older single‐source drugs and newer and older multisource drugs. This simple two‐by‐two framework enables us to analyze adjusted spending growth in a manner that dovetails with key dimensions of the policy debate over prescription drug spending. Distinguishing newer versus older single‐source drugs highlights how the pipeline of new drugs became steadily smaller throughout our analysis when measured in terms of fills per capita, but nevertheless grew in dollar terms due to the rapid recent increases in average payment per fill. This trend raises the stakes surrounding the debate over how public and private payers control access to new drugs and negotiate prices paid. Distinguishing newer versus older multisource drugs enables us to observe more clearly how average payments for the latter were affected by the Walmart pricing initiatives and, more recently, declines in generic competition.31, 36 More research is needed into the role of monopsony buying power and into policies regarding generic competition.

The largest contributor to increased per capita spending over the study period was single‐source drugs more than 4 years removed from initial FDA approval, accounting for $205 of the $476 increase in per capita spending, holding individual characteristics constant. Much of this increase occurred in the early years of our study, and this category subsequently played an important role in the overall slowing of prescription drug spending. The slowing of spending growth in this drug category was due to declines in the average number of fills rather than average payment per fill, which rose at an accelerating rate throughout the study period. Renewed spending growth among newly approved drugs will likely help perpetuate rising average payments per fill in this older single‐source category, highlighting the importance of careful measurement of the newer and older single‐source categories to help inform policy debates on topics such as exclusivity extensions granted for innovations subsequent to initial FDA approval, barriers to FDA approval of biosimilars, pharmaceutical advertising, public and private price negotiation and formulary design, and more.

Supporting information

 

 

Selden TM, Abdus S, Miller GE. Decomposing changes in the growth of U.S. prescription drug use and expenditures, 1999‐2016. Health Serv Res. 2019;54:752–763. 10.1111/1475-6773.13164