Abstract
Policy makers and insurers promote the use of generic drugs because they can deliver large savings without sacrificing quality. But these efforts meet resistance from the public, who perceive generic drugs as inferior substitutes for brand name counterparts. Building on literature showing that negative emotions reduce risk-taking, the authors hypothesize that receiving bad medical news (i.e., negative information about one’s health) prompts patients to favor brand name over generic drugs as means to safeguard their health. The evidence exploits low-density lipoprotein cholesterol test results, where a discontinuity from clinical guidelines enables the authors to estimate the causal effect of bad medical news. Using data covering patients’ prescription drug choices across drug classes, the authors find that patients receiving bad medical news become 8% more likely to choose the brand name alternative. The findings are reinforced by a secondary analysis incorporating the similar context of hemoglobin A1c (blood sugar) testing. The authors also find that bad medical news reduces preferences for generics most strongly among drugs of direct clinical relevance for each test, but the effect also manifests among non–clinically relevant drugs.
The large and increasing health care expenditures of developed countries have been a challenge for policy makers and the general public. In the United States, the issue is best reflected by the sizable share of gross domestic product spent on health care, now close to 20% (Hartman et al. 2022). Critics argue that much of these expenditures reflect inefficiencies (Nunn, Parsons, and Shambaugh 2020), suggesting that costs could be contained without sacrificing the quality of care received by patients.
Regarding the prescription drug component of these expenditures, policy makers view policies fostering the substitution of brand name with generic drugs as options of particular interest. In addition to using the same active ingredients and dosages, the U.S. Food and Drug Administration (FDA) certifies that generic drugs have the same key pharmacological properties of brand name counterparts (within margins). Accordingly, many experts view generics as molecular replicas of brand name drugs and, thereby, as delivering the same objective therapeutic value. Given the much lower prices of generics, substituting brand name with generic consumption could therefore lower expenditures without sacrificing the quality of care received by patients. Estimates for the United States suggest that these savings could be large, about 10% of prescription drug expenditures (U.S. $36 billion a year) if patients always chose the generic option when available. 1
Prompted by these facts, public and private insurers have introduced a variety of incentives (e.g., coupons, free samples) aiming to encourage the use of generics. These efforts are nevertheless met with resistance from the public, who perceive generics as of inferior quality compared with brand name drugs (Dunne and Dunne 2015; Hassali et al. 2009). Prior research has rationalized such preferences on the basis of informational gaps, that is, the fact that patients lack information reassuring them of the therapeutic equivalency between the two types of drugs. For example, Bronnenberg et al. (2015) find that, compared with the general public, pharmacists—who know more about drugs’ properties—are more likely to prefer generic over brand name aspirin, while Carrera and Villas-Boas (2020) and Ching (2010a) provide evidence that bridging this informational gap increases generic choice.
We contribute to this literature by investigating how negative information shocks about the patients’ own health—“bad medical news”—impact the relative preferences between brand name and generic drugs. We are motivated by a series of findings linking similar psychological stimuli with subconscious effects on choice. For example, people make more risk-conservative gambling and job-selection decisions when experiencing anxiety (Raghunathan and Pham 1999). Similarly, decisions become biased toward the low-risk option in the presence of worry (Johnson and Tversky 1983), fear (Cohn et al. 2015; Lerner and Keltner 2000, 2001), trauma (Callen et al. 2014; Cameron and Shah 2015), and weather-induced bad mood (Bassi, Colacito, and Fulghieri 2013; Hirshleifer and Shumway 2003; Saunders 1993). By raising patients’ alarm about their own health condition, bad medical news may infuse some of these emotions, prompting patients to take action to safeguard their health (e.g., improving their diet or exercising more). Given that brand name drugs are perceived as having higher efficacy and safety than generics (thereby implying smaller health risks), such actions may also include favoring the brand name alternative whenever confronted with a drug choice. Accordingly, we hypothesize that bad medical news may increase patients’ propensities to choose brand name drugs over generics.
Our interest in how bad medical news may impact drug choices is also premised on the idea that these effects could operate at a large scale. For example, millions of women undergo testing each year for genetic markers of breast cancer, where a positive test outcome could act as a bad news event. Similarly, given the high prevalence of cardiovascular disease in the developed world, tens of millions of adults are tested regularly for low-density lipoprotein (LDL) cholesterol (aka “bad cholesterol”). For an individual who has routinely had test results in “optimal” ranges, a “borderline high” result may also deliver bad medical news. In all, given the high prevalence of bad medical news shocks arising organically as patients interact with the medical system, their total effects could be large.
Our main analysis focuses on the medical news implied by LDL testing results. Instead of inferring the presence of bad medical news from patients’ testing histories (as in the preceding example), we rely on a comparison across patients. We compare patients who receive 129 mg/dL and 130 mg/dL LDL results. We choose this narrow window because it marks the frontier between “near optimal” and “borderline high” ranges, as defined by clinical guidelines. Compared with individuals who test at 129 mg/dL, those who test at 130 mg/dL “cross” the frontier. We therefore posit that, compared with the former, the latter patients (130 mg/dL) receive a bad news treatment. Crucially, the differences resulting from this comparison can be given a causal interpretation because LDL results include a measurement error. Since these errors are due to factors such as the extent of fasting prior to the test or the patient's posture while the blood is drawn, they can be deemed as plausibly exogenous. We therefore assume that patients are locally randomized between the 129 and 130 mg/dL measurements. This assumption is supported by the tight balancing between the two patient populations. We make use of the difference-in-differences (DID) approach to estimate the differential effect that receiving the LDL results has on the drug choices of treated (130 mg/dL) versus control (129 mg/dL) patients. Because the perception that generics are of inferior quality than brand name drugs is general (i.e., it applies to all drugs), the bad LDL news may affect choice beyond drugs of direct clinical relevance to LDL results. Accordingly, our estimation utilizes comprehensive data including all prescription drug choices made by sample patients, which cover almost 500 drugs across six drug classes (e.g., anti-infective, cardiovascular, gastrointestinal). Using a similar DID design, we also investigate whether the bad news shock changes the quantity and bundle of drugs used by patients. However, we do not find evidence that the bad news shock influences these decisions.
In our main analysis we estimate the impact of the bad news treatment on the probability that a patient chooses the generic over the brand name alternative. Consistent with our hypothesis, we find that, compared with control patients, those treated with bad medical news reduce their generic choice probability by about .01 after the test. The estimate represents a 1.3% reduction in the average patient's propensity to choose the generic option. Given that brand name drugs have a smaller share of choices (14%) than generics (86%), this result can be equivalently expressed as an 8% increase in the propensity to choose the brand name option. Considering the average generic price discount relative to brand name drugs (80%–85%), this effect implies roughly a 3% increase in total prescription drug expenditures for the average patient.
Encouraged by this finding, we probe the effect by investigating the consequences of varying treatment intensity. In a series of analyses, we find that the effect on generic propensity changes in a way that is directionally consistent with the change in the intensity of the bad news treatment. For example, we find relatively larger effects on generic propensity among patients whose LDL test happens around a medical office visit (for whom the test result may be more salient) and among healthier patients (who may be more surprised by the bad news). Among others, we find evidence of a markedly front-loaded effect, concentrated in the immediate aftermath of the test (90 days). The main mechanism behind this result pertains to the adoption of new drugs; that is, the bad news shock is particularly influential for patients who are purchasing a drug for the first time. To investigate the generalizability of the effect, we next expand our analysis by incorporating the results of a different medical test (i.e., hemoglobin A1c, measuring blood sugar levels), wherein we exploit the 7% threshold that patients with diabetes use to manage the condition. The results of this analysis are broadly supportive of the idea that our main result may generalize outside the context of LDL tests.
As noted, the perceived quality difference between brand name and generic drugs acts as a precondition for the bad news effect on brand/generic choice. The effect's size should therefore increase with the difference in perceived quality. We examine this implication through two additional analyses. We first exploit brand/generic price differences. In a differentiated product market, price differences of two products should be positively correlated with their perceived quality differences. Consistent with this idea, we find that the bad news effect primarily manifests when the branded option is sufficiently more expensive than the generic option. We next leverage findings of prior literature highlighting that patients learn drugs’ “true” properties through usage. In the presence of this form of learning, perceived quality differences should progressively narrow as patients gain more consumption experience. Accordingly, we find that the effect decays with the patient's cumulative consumption experience.
Our final analysis studies how the bad news effect on generic choice varies between different types of drugs. Our main analysis compares the effects unfolding on clinically relevant drugs (cholesterol drugs for LDL testers, diabetes drugs for A1c testers) against those on drugs that are not clinically relevant (e.g., gastrointestinal drugs). While we expect the effect to operate on both types of drugs due to generics being perceived as of inferior quality in general, the bad test results can be particularly alarming for patients with a related condition. Consistent with this intuition, we find that the bad news effect is significantly stronger for clinically relevant drugs than for drugs that are not clinically relevant. We also provide evidence suggesting that accounting for consumption experience is important to correctly estimate the effects of bad news on generic propensity for clinically relevant and non–clinically relevant drugs.
We organize the rest of the article as follows. We start by describing the institutional and literature background, and then introduce our data set and research design. We proceed by investigating the potential impacts of bad LDL news on patients’ choices of prescription drugs. We then investigate the robustness of our results using bad A1c news. Our last set of analyses probes the importance of quality perceptions as a key driver of the documented effects. We conclude by discussing implications for practice and future research directions.
Institutional Background and Related Literature
Generic Drugs
Pharmaceutical drugs combine active and inactive ingredients. Active ingredients deliver the intended pharmacological effects, while inactive ingredients enable auxiliary features such as coloring and flavoring. Generics have the same active ingredients as their respective branded incumbents. Accordingly, public health agencies (e.g., the FDA, Health Canada) view them as providing the same objective quality as brand name drugs. Moreover, they do so at much lower costs: on average, at a discount of 80%–85% compared with brand name equivalents.
For most drugs, generic market entry is limited by patents protecting the use of active ingredients. Since the Hatch-Waxman Act defined modern entry requirements in 1984, generic penetration has increased steadily, from 36% in 1994 to 88% of all prescriptions in 2014 (IMS Institute for Healthcare Informatics 2015) and to 90% in 2019 (Woodcock 2019). This trend reflects the continued efforts of public and private payers to encourage generic substitution (Dunne and Dunne 2015). These efforts, which are primarily price-based, include tiered copayments (i.e., lower copayments for generic drugs), pay-for-performance schemes that target physicians and pharmacists, and reference price schemes, as well as the use of promotional tools such as coupons and free samples. Some insurers have even offered generic drugs at no cost to patients (Ching, Granlund, and Sundström 2022; O’Malley et al. 2006).
Despite these incentives, generic substitution faces some resistance from the public. This is mainly due to generics being perceived as of inferior quality compared with brand name drugs (Dunne and Dunne 2015; Hassali et al. 2009). These perceptions entail concerns about safety and efficacy, both of which associate generics with higher perceived health risks. The link between safety concerns and perceived health risks is straightforward: patients worry about the possibility of generics producing adverse side effects (e.g., migraine and vomiting). In turn, efficacy concerns may fuel perceived health risks by hindering patients’ ability to manage the drug's targeted condition. For example, less effective cholesterol drugs may increase the likelihood of future cardiovascular events, while less effective insulin products may increase the risk of diabetic seizures. Consistent with this view, Tootelian, Gaedeke, and Schlacter (1988) show that patients tend to disproportionally favor brand name drugs when they face larger health risks from the targeted condition. Similarly, Ganther and Kreling (2000) find that patients demand larger savings to purchase generic prescription drugs of higher perceived risk.
While many studies of the preference bias against generics survey perceptions of generics in general, others focus on drugs of specific domains, such as antipsychotics (Roman 2009), asthma drugs (Williams and Chrystyn 2007), and cardiovascular drugs (Kesselheim et al. 2008). These findings converge on the idea that much of the difference in brand/generic quality perceptions applies to all generics (i.e., across drug classes). Beyond patients, health care professionals (providers and pharmacists) can be negatively predisposed toward generics (Chua et al. 2010; Hassali et al. 2010). It is possible that professionals adopt these predispositions to avoid friction with patients who may exert pressure to receive the brand name option (Chua et al. 2010; Hassali et al. 2010; Williams and Chrystyn 2007).
Related Literature
Our main hypothesis builds on empirical findings that link negative emotional experiences (e.g., fear, worry, anxiety) to less risk-taking (e.g., Johnson and Tversky 1983; Kuhnen and Knutson 2011; Lerner and Keltner 2000, 2001). For example, in a lab experiment, Johnson and Tversky (1983) show that reading a newspaper homicide report designed to induce anxiety and worry leads one to have more pessimistic crime rate estimates. Interestingly, Johnson and Tversky (1983) find that such pessimistic bias also occurs for risks unrelated to the news article, such as the chance of dying in a fire or dying of leukemia. Similarly, experiencing anxiety has also been found to bias decisions in favor of lower-risk gambling and job-selection choices (Raghunathan and Pham 1999), while weather-induced bad mood reduces risk-taking in stock-market trading (Bassi, Colacito, and Fulghieri 2013; Hirshleifer and Shumway 2003; Saunders 1993). Like the effect of anxiety-provoking homicide reports on estimates for unrelated risks, weather-induced bad mood hinders financial risk-taking behavior, even though weather contains no information about the economy. Similarly, bad LDL news may affect drug choices beyond cardiovascular drugs (e.g., asthma drugs), even though cholesterol levels may contain little information about noncardiovascular health.
The choice between brand name and generic drugs has also been studied in the marketing and economics literatures (reviewed by Ching, Hermosilla, and Liu 2019). Ching (2010a) provides evidence of a preference bias against generics in the prescription drug market, while Bronnenberg et al. (2015) and Carrera and Villas-Boas (2020) do so for over-the-counter (OTC) drug choices. As noted in the introduction, these analyses focus on the role of information about generics, that is, patients lacking information that reassures them of the equivalency between brand name and generic drugs. By contrast, we examine how the psychological stimulus delivered by bad medical news impacts brand/generic choice, even though such news contains no information about products.
Our work also contributes to a growing literature focusing on behavioral hazard in health care decision making (Baicker, Mullainathan, and Schwartzstein 2015; Handel and Kolstad 2015; Mullainathan, Schwartzstein, and Congdon 2012). This literature highlights how behavioral biases may lead to treatment choices that do not improve consumer well-being. In our context, consumer well-being may be negatively affected through lost monetary savings implied by a strengthened preference for brand name drugs. As highlighted by Shrank et al. (2006), patients may also experience worse health outcomes given that patients who use generics are significantly less likely to skip doses. Finally, it is important to differentiate this research from prior work investigating the impact of news media coverage on drug demand (Ching et al. 2016). Rather than product-specific journalistic news, we consider patient-specific medical news.
Data
We utilize 2011 and 2012 MarketScan data, which track health care utilization (i.e., health care expenditures) for individuals living in the United States. Data are compiled from de-identified administrative claims, from a large array of employers and health plans, including government and public organizations. By 2012, MarketScan included claims covering almost 80 million employees (up to age 65) and their dependents.
A distinctive feature of the database is that it links individual records across different domains of health care in addition to drug utilization. Importantly, we have access to MarketScan's Lab files, which capture laboratory tests ordered in office-based practice settings. Our analysis leverages LDL cholesterol test results available from these files.
LDL Testing Results
LDL testing results are expressed in milligrams per deciliter (mg/dL). The distribution of all LDL test results observed in full the sample is shown in Figure W5 in the Web Appendix. The distribution has wide dispersion and is smooth around the 130 mg/dL threshold for “borderline high” levels, where our analysis takes place. We observe a total of 4,940 test results of 129 mg/dL or 130 mg/dL. However, a majority of these results correspond to individuals who have more than one test in the sample. Because multiple testing blurs our inference, for our analysis we primarily focus on the set of 2,282 individuals who have only one test in the data. (Multiple testers are analyzed separately.) Individuals with a single test in our main sample are about evenly split between 129 mg/dL and 130 mg/dL results (respectively, N = 1,169 and 1,113). Their average age (at testing) is close to age 49; about 57% of them are female and 43% are male.
Although MarketScan data do not record the specific reasons why patients take the LDL test, some statistics suggest that a majority of these tests may occur in the context of routine checkups. Specifically, claims for outpatient services indicate that about 70% of the tests happen within a week of a primary care office visit. Of these visits, over 50% are conducted by family practitioners and 20% by internists. Furthermore, the types of medical problems informed by these tests (as revealed by diagnostic codes, when available) concentrate on problems typically addressed in the context of routine checkups, such as general prevention or cardiovascular health (see Table W11 in the Web Appendix).
Drug Choices
The second pillar for our data set corresponds to claims for prescription drugs. To simplify our description of these data and results, we adopt a few terminology conventions. First, we use the terms “transaction,” “purchase,” and “prescription” interchangeably to refer to a single drug purchase. Second, we use the term “molecule” to refer to compounds or unique combinations of active ingredients, irrespective of branding. For example, we may say that the alprazolam molecule has both brand name (Xanax, Niravam) and generic (generic alprazolam) alternatives. This nomenclature is helpful partly because we seek to control for unobserved variation at the molecule level, for example, from assortment sizes or brand/generic price gaps. Our regressions will include a fixed effect for each molecule.
Because most health insurance plans do not cover OTC drugs, our data set has very little coverage of them (<1%). We remove OTC drugs altogether, and are thus left with a sample of prescription drug purchases only. Given our focus on brand/generic choice, we restrict our attention to molecules for which there is at least one generic and one brand name alternative in the market during the covered period (i.e., “multisource” drugs).
The resulting sample has claims for 35,080 prescription drug purchases, covering 484 different molecules. It is important to emphasize that this set of molecules covers the full set of drug needs of patients in our sample, not just those associated with high cholesterol. About 62% of individuals have at least one drug purchase. For these, the median number of purchases is 15 (interquartile range = [6, 33]). Panels A and B of Figure 1 show the distribution of total purchases per individual for individuals with test results of 130 mg/dL and 129 mg/dL, respectively. For each type of individual, Panels C and D of Figure 1 illustrate the distribution of purchases across the six drug classes (categories) listed in the data. Drugs targeting cardiovascular and central nervous system conditions command the chart. In line with independent data (IMS Institute for Healthcare Informatics 2015), most purchased drugs are generics (86% overall). However, there is significant generic share variation across drug classes. We will take advantage of this variation to investigate the generalizability of our results. Notice that, as further stressed subsequently, the distributions do not significantly differ between the two types of individuals.
Main Descriptives of the Drug Purchases (Claims) Sample: Treated (130 mg/dL) Versus Control (129 mg/dL) Individuals.
Quasi-Experimental Framework
The Role of LDL in Lipid Management
Cholesterol is a waxy substance that circulates in the blood stream and assists a series of biologic processes. Cholesterol can also enter the body through food, particularly meat, poultry, and dairy. There are two main kinds of cholesterol: high-density lipoprotein or “good” cholesterol, and LDL or “bad” cholesterol. The latter is problematic because it can narrow blood vessels and impede blood circulation. This problem can evolve into serious adverse health events, such as strokes. For this reason, LDL tends to be the main focus in clinical guidelines for lipid management.
Clinical guidelines for cholesterol management adopt a series of thresholds for LDL levels. In particular, 130 mg/dL marks the frontier between “near or above optimal” and “borderline high” level. For example, the formal National Cholesterol Education Program ATP-III lipid management guideline prompts providers to consider a cholesterol-management treatment for patients who have up to two risk factors and receive an LDL test result of 130 mg/dL or higher (National Institutes of Health 2001). Similarly, the Cleveland Clinic and Mayo Clinic both state in their patient-oriented websites that LDL should ideally be less than 130 mg/dL. The popular medical website WebMD provides similar guidance. 2
Local Randomization
We posit that, compared with a patient whose result is in the marginal category (129 mg/dL), a patient testing at 130 mg/dL is treated to a probabilistic shock of bad news about a health condition. For these “treated” patients (130 mg/dL), concepts such as “borderline high” and “abnormal” may be used with higher probability in connection to their health conditions, as compared with control patients (129 mg/dL). As a result, patients in the treated group may become disproportionally likely to experience the kind of negative emotions that tilt decisions toward lower-risk alternatives.
To identify the effect of such bad news on generic choice propensity, we rely on the assumption that the side of the 129 or 130 mg/dL frontier on which an individual lands is a random event. The main argument supporting this assumption pertains to the relatively large measurement error of LDL test results. The medical literature decomposes this error into “analytical” and “preanalytical” variability (Marcovina, Gaur, and Albers 1994). Sources of preanalytical variability include factors such as the individual's posture during sampling, the duration of tourniquet application, how strictly and for how long the individual adhered to fasting prior to testing, and various other factors (Narayanan 1996). In turn, analytical variability is determined by how the blood sample is handled and analyzed in the lab (e.g., whether and for how long it was frozen prior to analysis).
Marcovina, Gaur, and Albers (1994) estimate that analytical variability corresponds to 1.3% of the mean LDL result while preanalytical variability reaches 9.2%. In the context of our sample, these results would indicate that preanalytical and analytical factors add a standard deviation of about 11.5 mg/dL around the “true” LDL level. Given this large variability, we can assume local randomization between individuals testing at 129 mg/dL and 130 mg/dL.
Table 1 presents a series of preperiod (i.e., before testing) statistics showing that treated and control individuals have similar characteristics, operate in similar contexts, have similarly generous insurance plans, and behave in similar ways with respect to health care utilization. Panel A focuses on demographics: age and sex. As shown by the small standardized differences in Column 3, these variables balance tightly between groups. For example, the average age of treated and control patients differs by less than 1% of a standard deviation. This difference is much smaller than the usual 25% and 10% standardized difference thresholds used in the literature to declare covariate imbalance (Austin 2009). A similar result is observed for the female indicator and geographical location categories (regional indicators).
Balancing in the Quasi-Experimental Setup.
Notes: Parentheses show sample standard deviations.
Panel B of Table 1 considers a series of variables that track the amount of utilization measured through the number of filed claims. These variables are claims for drug purchases (total, generic, and cardiovascular-targeted), claims for inpatient admissions and outpatient services, and claims for medical tests (all measured as monthly averages). All variables balance tightly between groups. In Panel C we consider the number of (other) medical test claims filed within the same day as the LDL test under study. The average of around 12 tests suggests that patients take several tests at the same time, which would be consistent with routine yearly checkups. The small standardized difference again suggests that treated and control individuals have similar experiences on the day the LDL test is taken.
As mentioned, health care insurance plan characteristics are not observed in the data. Nevertheless, we can assess the implicit generosity of the available insurance through a “coverage ratio.” We compute this variable by dividing the patient's out-of-pocket payments (deductible plus copayment plus coinsurance) on the total payments made by the insurer to the pharmacy. As shown in Panel D, this ratio averages about 40%, without much of a difference between groups. Thus, unobserved insurance variation should not drive our results.
In Web Appendix A, we present a series of additional analyses that further document the experiment's tight balancing. In particular, we show the absence of systematic treatment/control differences in the following respects: (1) medical diagnosis codes associated with the LDL test, (2) kinds and amount of non-LDL testing carried out prior to the LDL test, (3) quantitative results of non-LDL tests taken prior to the LDL test, and (4) time when the tests (LDL and non-LDL) are administered. We interpret the summation of these results as strong evidence in favor of the validity of our research design.
Impacts on Drug Consumption Behavior
Although our main focus is on how bad medical news shocks may impact brand/generic choice, it is conceivable that these shocks may have broader impacts on drug consumption. For example, the bad news could lead to changes in the bundle of drugs used if patients are prompted to adopt preventive treatments or abandon those having relatively strong side effects. Similarly, the bad news could introduce changes in the quantity of drugs used, for example, if patients react by improving treatment adherence (e.g., becoming less likely to “miss a pill”). In this section we investigate the extent to which these “bundle” and “quantity” effects are observed in practice. Results do not support the presence of these effects.
Quantity Effect
To measure changes in the quantity of drugs that patients consume, we formulate an outcome variable that tracks the total number of prescription drug purchases by patient i during period t. Since most tests in the data (92%) are taken sometime during the month, we cannot define time periods as calendar months. (Testing dates are distributed relatively uniformly in time. See Figure W1, Panel A, in the Web Appendix.) Accordingly, we define time periods as 30-day windows relative to the testing date. For example, for an individual who took the test on September 18, 2012, this approach would give us t = −1 for August 20–September 17, 2012 (last period before the test); t = 0 for September 18–October 17, 2012 (period starting the day the test is taken); t = 1 for October 18–November 16, 2012; and so on, with t = 0 marking the beginning of the period after testing. This person enters the final data set through 23 observations, one for each of the fully covered 30-day periods, t = −20 (January 27–February 25, 2011) through t = 2 (November 17–December 16, 2012). The full data set contains 31,567 observations derived from the 1,408 individuals associated with at least one drug purchase. On average, patients purchase 1.03 prescriptions each time period (SD = 1.71). About 57% of the observations in the panel are zero. Using the data formatted in this way, we estimate the following DID specification:
Several aspects of Equation 1 are important to highlight. First, since we cannot qualify whether patients directly observe the test result (probabilistic treatment), β estimates are formally described as intention-to-treat results (Angrist, Imbens, and Rubin 1996). Second, given that the sample only includes individuals with test results of 129 or 130 mg/dL, the bad news shock corresponds to the differential information received by the latter compared with that received by the former individuals. That is, our framework does not allow potential treatment effect asymmetries rooted on whether the news is positive or negative. In other words, similar to interpreting a 130 mg/dL result as bad news relative to a 129 mg/dL result, we could interpret the latter as good news relative to the former.
Also notice that Equation 1 includes narrowly defined fixed effects. The individual-level fixed effects λ absorb the influence of all time-invariant characteristics of individuals (e.g., socioeconomic and demographic characteristics, overall health condition, and medical history). Considering that most people rarely change insurance plans (Handel 2013), λ effects also help control for unobserved insurance coverage differences. In turn, the time-period fixed effects δ control for effects operating in relation to the temporal proximity between the drug purchase and the medical test. For example, in anticipation of taking the test (e.g., during t = −1), patients could become less likely to skip or miss a dose of medication, thereby increasing total consumption. Equation 1 would flexibly capture such effects through a larger-than-average estimate for δ−1. Similarly, the interaction with the medical system implied by taking the test could lead to a temporary increase in drug consumption (e.g., through refilled prescriptions). These effects would be captured by δ parameters for the immediate aftermath of the test (e.g., δ0 and δ1). Lastly, note that Equation 1 omits the variables Treated and Post in stand-alone form because λ and δ effects make them redundant.
The estimated coefficient for β is presented in Column 1 of Table 2. While the coefficient's positive value associates the bad news shock with an increase in purchased quantities, the effect is statistically nonsignificant at conventional levels. We take this result as evidence that bad medical news does not impact the quantity of drugs consumed.
Main Results.
*p < .1. **p < .05. ***p < .01.
Notes: Linear probability specifications for the probability of choosing the generic option (Equation 2). Parentheses show standard errors. For estimates in Columns 1–2, errors are clustered at the individual level; for estimates in Columns 3–6, they are clustered at the individual/molecule level.
Bundle Effects
Recall that we conceptualize bundle effects as changes in the set of drugs used by the patients induced by the bad LDL news. We use the same procedures described previously to construct a variable that tracks the number of different molecules purchased by each patient during each time period. The resulting variable, BUNDLEit, equals the log of (one plus) the number of different molecules purchased by patient i during time period t. On average, individuals purchase .96 different molecules each time period (SD = 1.55). Column 2 of Table 2 presents the β estimate that we obtain by estimating Equation 1 using BUNDLE as dependent variable. We again obtain a positive and statistically nonsignificant estimate, which fails to support the hypothesis that bad news prompts changes in the set of drugs consumed.
Impacts on Generic Choice Propensity
Here we turn to our main objective, which is to estimate the impact of bad LDL news on individuals' generic choice propensity, that is, the probability that an individual chooses the generic over the brand name option conditional on purchasing a drug. Accordingly, we use the data in disaggregated form (i.e., a data set in which each observation corresponds to a drug purchase). Also recall that our data set includes drug purchases covering patients’ full set of drug needs, not just those associated with high cholesterol. Using this full data set, we estimate the following DID model, which minorly adapts Equation 1:
Main Effect Estimate
The estimate for the coefficient β in Equation 2 is presented in Column 3 of Table 2. The estimate is negative and marginally significant (i.e., significant with 90% confidence), supporting the notion that bad medical news increases the preference for brand name drugs. The estimate indicates that, compared with control individuals (129 mg/dL), treated ones (130 mg/dL) experience a reduced generic choice probability of an additional −.0109 after the test. Considering the pretesting generic propensity baseline, this result suggests that the bad news shock reduces the frequency of generic choice by about 1.3%. Equivalently, given the relatively modest share of brand name drugs (14% overall), this point estimate represents an 8% increase in the average patient's propensity to choose the brand name option. In addition to the plausibly random treatment assignment, recall that our model controls for unobserved individual- and molecule-level variation, as well as for potential anticipatory effects. Thus, it is difficult to attribute this result to reasons other than the LDL test outcome. We are further reassured by two additional sets of results. First, we implement two sets of placebo tests (falsified thresholds and testing dates), none of which falsifies the result. Second, a formal test rejects the presence of confounding pretrends. These analyses and their respective results are presented in Web Appendix B.
Treatment Intensity
Encouraged by the previous finding, we next probe the effect in relation to the intensity of the treatment experienced by different individuals. We consider a series of scenarios in which the intensity of the bad news treatment is implicitly altered. Across these contexts, we find directionally consistent changes in the estimated effects on generic propensity.
Co-occurrent medical appointments
We first consider a scenario in which treatment intensity may vary according to the salience of test results to patients. In particular, we leverage the idea that test results may be more salient for those patients who also have a primary care medical appointment (office visit) around the time of the test. As noted previously, information on the incidence of these visits is available from outpatient services claims, and about 70% of the sample had one such visit within a week of the test. The coefficient in Column 4 of Table 2 results from reproducing our estimation using only the drug claims data for this subset of patients. Consistent with increased treatment salience, the −.0174 estimate for β is larger in magnitude and more precisely estimated (statistically significant with 95% confidence) than its full-sample counterpart.
Health status
We next investigate how the patient's health status may moderate the bad news effect. Compared with sick patients, healthy patients may be less used to receiving information that unveils a health deficiency. Bad LDL news may therefore imply a larger shock for healthy patients than for sick patients.
The primary empirical hurdle to investigating this hypothesis stems from the fact that MarketScan data do not contain variables describing the overall health condition of individuals. Given this limitation, we use total health care expenditures as a proxy for health status, under the assumption that higher health care expenditures reflect poorer health condition. In particular, we create the indicator variable HighSpenderi =
We incorporate the HighSpender indicator into Equation 2, as illustrated in Column 5 of Table 2. Separate bad news effects are estimated for patients in each group (HighSpender = 0 and HighSpender =1). The estimated bad news effect parameter for individuals associated with HighSpender = 0 (i.e., good health) is −.0265, which is more than twice that estimated from the full sample, and statistically significant with 95% confidence. For individuals associated with HighSpender = 1 (i.e., poor health), the parameter is −.0085, which is much smaller as well as statistically nonsignificant. These results suggest that healthier patients are more vulnerable to the bad news effect on brand/generic choice.
Temporal effects
Research from multiple disciplines converges on the finding that emotional reactions tend to be short-lived (e.g., Card and Dahl 2011; Depetris-Chauvin, Durante, and Campante 2020; Ekman 1999; Verduyn et al. 2009; Verduyn, Van Mechelen, and Tuerlinckx 2011). From this finding, we conjecture that the bad news treatment may have higher intensity in the immediate aftermath of the test, decaying afterward (i.e., front-loading). To investigate this idea, we modify Equation 2 to allow for the estimation of separate bad news effects over three time periods after the test. We select the following cutoffs (which partition transactions after testing into approximate terciles): (1) first 90 days after testing (periods t = 0, 1, and 2), (2) days 91–210 after testing (periods t = 3 through 6), and (3) day 211 and after (periods t = 7 and higher). Compared with the full-sample estimate of Column 3 (−.0109), the −.0169 coefficient for the first of these periods (Table 2, Column 6) is larger and more precisely estimated (significant with 99% confidence). The signs of the estimates for the next two periods continue to be negative, but they are statistically nonsignificant. These results support the idea that the negative emotions infused by the bad news have front-loaded effects.
The effect's front-loading raises the question about the specific decisions through which the effect operates. There are two possible channels. First, the negative effect on generic propensity could unfold via switching decisions, that is, a combination of slowed-down brand-to-generic switching and accelerated generic-to-brand switching. However, with less than 2% of purchases representing switching between brand name and generic alternatives, this channel can at most play a minor role in explaining the effect in our data. The second channel pertains to brand/generic choice when patients adopt a new molecule (i.e., when they purchase it for the first time). It is possible that, in the context of adopting a new molecule, bad medical news tilts adopters’ preferences toward the brand name option. In Web Appendix C, we present evidence consistent with this hypothesis. We find that the bad news shock increases the propensity to choose the brand name option when patients adopt a new molecule. Consistent with the front-loading results, our estimates show that the effect on generic propensity for newly adopted drugs is also short-lived.
Multiple testing
We conclude by focusing on the issue of multiple LDL testing. People who test for LDL more than once may not only be more driven to thoroughly analyze the results, but also be better acquainted with the measurement error. Accordingly, we conjecture that multiple testing may be associated with reduced treatment intensity.
To analyze the problem, we construct a data set using the information of individuals (N = 2,341) who record at least one LDL test in addition to the test with a result at the frontier of 129 and 130 mg/dL. (Recall that these individuals were excluded from our main sample.) A series of empirical considerations is required to analyze these data (e.g., individuals vary in how many additional testing results they record), so we present our procedures and results in Web Appendix D. Consistent with the bad news effect, we also obtain a negative β estimate from this sample. However, the estimate is about one-third the magnitude of our main estimate in Column 1, as well as statistically nonsignificant. A second analysis of the same data suggests that the bad news effect may concentrate on individuals who did not previously receive a result of 130 mg/dL or higher. This result coincides with the intuition that individuals who have not previously received bad LDL news may be more surprised to receive it. In parallel, the 130 mg/dL result may represent good news to those individuals who previously obtained a borderline high result (≥130 mg/dL). Given several limitations and the lack of statistical significance, we interpret these results as merely suggestive of the idea that bad LDL news has a marginally decreasing impact on brand/generic choice.
Additional Evidence: Bad A1c News
The analysis presented in this section assesses the generalizability of the bad news effect on generic propensity. We do so by examining the impact of bad news generated by a different type of medical test (i.e., hemoglobin A1c tests, which are also contained in the MarketScan Lab files). These tests measure blood sugar levels and are used for diagnosing and managing diabetes.
Hemoglobin A1c Tests
Three features shared with LDL testing make A1c testing a good secondary candidate for our analysis. First, like LDL results, A1c results are also expressed on a continuous scale, that is, as the percentage of red blood cells with sugar-coated hemoglobin (typically 5%–8%). Second, A1c results also include a significant measurement error, around .5% (Phillipov and Phillips 2001). This error introduces the necessary local randomization around clinical thresholds. Third, like LDL tests, A1c tests are common, particularly among people who have been diagnosed with diabetes.
Despite these favorable features, two aspects of A1c testing introduce a measure of experimental noise. First, compared with LDL cholesterol, the interpretation guidelines for A1c tests are not as strict as those for LDL tests. For example, the American Diabetes Association writes that “providers might reasonably suggest even lower A1C goals than the general goal of <7% … conversely, less-stringent A1C goals than the general goal of <7% may be appropriate for patients with a history of …” (American Diabetes Association 2010). The website WebMD, which is highly popular among patients, includes a similar emphasis. This aspect suggests that the measurable impacts of bad A1c news on generic propensity may be muffled compared with those of bad LDL news.
Second, A1c results inform two distinct clinical decisions. The 7% threshold referenced previously is used by patients with a diabetes diagnosis to manage the condition. In addition, two other thresholds are used for diagnosing the condition, 5.7% and 6.5% (entry to prediabetic and diabetic ranges, respectively). Whereas the 7% threshold for diabetes management has been consistently adopted by official diabetes management guidelines (e.g., American Diabetes Association 2010, 2021), there has been an ongoing debate about whether the latter two thresholds should be relied on for diagnosis. 4 Consistent with this scenario, we only detect bad news effect on generic choice propensity around the 7% frontier. Estimates obtained using the 5.7% and 6.5% cutoffs are presented in Table W9 in the Web Appendix. In line with our results for the LDL sample, we fail to detect statistically significant impacts on the quantity or bundle of consumed drugs (see results in Table W8 in the Web Appendix).
Analysis
With these caveats in mind, we incorporate into our analysis the 143,165 drug claims associated with the 3,725 patients at 6.9% or 7% A1c frontier, with virtually no patient overlap with the LDL sample.
5
In our first analysis, we estimate Equation 2 using these data only. Consistent with our results from LDL testers, we also obtain a negative estimate,
We next reestimate Equation 2 on a data set that combines the drug purchases of LDL frontier (129 vs. 130 mg/dL) and A1c frontier (6.9 vs. 7%) individuals (N = 5,131). In this “pooled” regression, baseline differences between the two testing contexts are absorbed by the individual-level fixed effects included in the model. The resulting estimate,
To further compare the effects of bad LDL and A1c news on brand/generic choice, we analyze the heterogeneity of the bad news effect across molecule classes. In particular, we use a slight modification of Equation 2 to estimate bad news effects specific to each of the six drug classes listed in Figure 1. Results are summarized by a set of estimates
Effects of Bad LDL and Bad A1c News Across Drug Classes.
Two findings from Figure 2 support the generalizability of the bad news effect. First, most estimates are negative. That is, in both the LDL and A1c samples, estimates are directionally consistent with the bad news effect across the six drug classes covered by the data. Second, there is a positive correlation of .13 across the vector of six class-specific point estimates obtained from each test; the correlation increases to .82 when we omit the outlier (the hormones and synthetic substitutes class). We interpret this result as evidence that the bad news effect may generalize across different medical tests.
The Role of Perceived Quality Differences
As we have noted, a precondition for the bad news effect is the presence of perceived quality differences between brand name and generic drugs. Here we present two analyses that probe this idea. Our first analysis leverages the observation that, in a differentiated product market, perceived quality differences between brand name and generic options should be positively correlated with their corresponding price differences. Accordingly, we find that the bad news effect focuses on molecules where the brand name option is sufficiently more expensive than the generic counterpart. Our second analysis leverages a prediction from research on consumer learning, namely, that the accumulation of consumption experience helps patients grasp a drug's “true” therapeutic properties. Consistent with this prediction, we find that the bad news effect decays with experience, arising only for relatively inexperienced patients. 7 We use these results to inform our final analysis, which characterizes the bad news effect's heterogeneity with respect to molecules that are clinically relevant to each test and those that are not. We find that the bad news effect operates on both types of molecules. However, the effect is considerably larger for clinically relevant molecules (CRMs).
Evidence from Pricing Differentials
Given that generics are molecular replicas of their brand name counterparts, the bias against generics tends to be rationalized on the basis of their perceived quality differences. Accordingly, we should expect a stronger bad news effect when the perceived quality differences are larger. To test this implication, the main empirical hurdle is that we do not directly observe perceived quality differences. Here we circumvent this challenge by leveraging pricing differentials.
Our analysis builds on two strands of literature, both of which associate larger brand/generic price differences with larger differences in perceived quality. First, in structural models of drug choice, consumption utility is increasing in perceived quality and decreasing in price (e.g., Ching 2010a; Crawford and Shum 2005; Narayanan and Manchanda 2009). In this framework, profit-maximizing firms would charge higher prices for drugs of higher perceived quality (Anderson, De Palma, and Thisse 1992; Ching 2010b). Second, considering that drugs can be described as experience goods (Berndt 2002), patients may use prices to make inferences about quality (e.g., Erdem, Keane, and Sun 2008; Milgrom and Roberts 1986; Wathieu and Bertini 2007). According to these rationales, we posit that if the bad news effect stems from generics being perceived as having inferior quality than brand name drugs, then the effect should be stronger when the brand name option is relatively more expensive than the generic option.
To implement this test, we take advantage of the widespread brand/generic price differences observed in the market. For a molecule j, we operationalize the price differential as
We construct Δ differentials leveraging data from average wholesale prices (AWPs), which are the equivalent to sticker/list prices in traditional retailing (Alpert, Duggan, and Hellerstein 2013; Gencarelli 2002). Details on how we construct
In our econometric model, we account for the variation of Δ through a median split, as shown in Table 3. Whereas the two-way interaction Treated × Post captures a baseline bad news effect that applies to all molecules, the triple interaction Treated × Post × AboveMedianΔ captures an additional effect that would apply only to molecules for which the price differential is large enough in favor of the brand name option. (The specification is otherwise identical to Equation 2.) Columns 1 and 2 of Table 3 show the estimates obtained from the LDL and A1c samples, respectively; Column 3 shows the estimates from the pooled sample. (Estimation samples are somewhat smaller than in our previous analyses due to missing price information.) Consistent with our prediction, the bad news effect strengthens for molecules associated with above-median Δ values. This strengthening is clearer (and statistically significant) in the A1c and pooled samples. Moreover, for these two samples, it is possible to conclude that the bad news effect arises primarily for molecules associated with above-median Δ values.
Price as a Signal of Quality.
*p < .1.
**p < .05.
***p < .01.
Notes: Linear probability specifications for the probability of choosing the generic option (Equation 2). The pooled sample includes individuals from both the LDL and A1c samples. All models include fixed effects for molecules, individuals, and time periods. Parentheses show standard errors clustered at the level of patient/molecule pairs.
In Web Appendix F we present an analogous analysis that is based on approximated out-of-pocket prices instead of AWPs. We obtain consistent results that the bad news effect is primarily observed where the brand name alternative is sufficiently more expensive than the generic one.
Evidence from Consumption Experience
Several studies document how the accumulation of consumption experience leads to patients learning about drugs’ “true” therapeutic properties (e.g., Ching 2010a, b; Crawford and Shum 2005). Since generic drugs are molecular replicas of brand name counterparts (and hence have the same “true” properties), this form of learning should progressively level the perceived qualities of brand name and generic drugs. Building on this observation, we hypothesize that the bad news effect decays with consumption experience. The analysis presented in this section finds support for this hypothesis. The crucial input needed to implement the test is a measure of consumption experience. Since we cannot measure consumption experience for patients purchasing drugs early in the sample, we reserve the early portion of our data to assess patients’ experience levels. Accordingly, we use drug purchase data for 2011 (first year of our sample) to compute the experience measure. We then incorporate this measure into the sample of 2012 purchases, which we use to estimate the models. To counteract the sample size reduction, we rely on the sample that pools the data of LDL and A1c testers for estimation. 8
To facilitate the interpretation of our econometric estimates, we formulate a metric of inverse experience, or “inexperience.” This metric is defined as
Table 4 describes the variation of the inexperience score, presented separately for the six molecule classes codified in the data. The cardiovascular class is associated with the most experienced patients, with inexperience scores that average .45. As for other classes, these scores exhibit a significant amount of within-class variability (SD = .39). At the other extreme, the anti-infective class has the least experienced patients, with scores averaging .83 (SD = .30). This comparison between the cardiovascular class and the anti-infective class is intuitive in that, given that most cardiovascular conditions are chronic, the scope for experience accumulation is much larger. By contrast, most conditions treated with anti-infectives are acute (i.e., short-lived), thereby providing fewer opportunities for experience accumulation. This relationship generalizes to the full sample, where patients purchasing drugs for chronic conditions are associated with .2 lower inexperience scores than the average other patient (p < .01).
Consumption Inexperience Scores.
Notes: Inexperience scores are computed at the patient/molecule level, as per Equation 3. The scores summarize the amount of consumption experience that the patient has with respect to a given molecule, with lower scores reflecting more experience. Scores tabulated here are for the sample used to estimate the models of Table 5, which is composed of 2012 purchases by patients who tested for LDL or A1c during 2012.
Column 4 of Table 5, Panel A, presents estimation results for a model that incorporates the inexperience score as a moderator for the bad news effect. In addition to the Treated × Post interaction, the model includes the triple interaction Treated × Post × Inexperience and is otherwise identical to Equation 2. As in our previous analyses, the coefficient for Treated × Post captures a baseline bad news effect operating regardless of consumption experience. In turn, the coefficient for Treated × Post × Inexperience captures an additional bad news effect, which operates in direct proportion to inexperience. The coefficient estimate for the baseline bad news effect (Treated × Post) is positive, although quite small as well as statistically nonsignificant. By contrast, the estimate for the triple interaction parameter is negative, marginally significant (90% confidence), and large in magnitude. In Panel A, we include the bad news effect estimates from previous analyses to highlight that the bad news effect becomes contingent on sufficiently high levels of inexperience. Evaluated at maximum inexperience, the bad news effect amounts to a .0105 reduction of the generic choice probability. In addition, given the formulation of the inexperience score (Equation 3), the estimates in Column 4 imply that the bad news effect disappears after 2.7 purchases. All in all, we interpret these results as broadly supportive of the idea that, by reducing brand/generic differences in perceived quality, consumer learning reduces the scope of operation for the bad news effect.
Inexperience and Class Spillover Effects.
*p < .1.
**p < .05.
***p < .01.
Notes: Linear probability specifications for the probability of choosing the generic option. Columns 1–3 show results for Equation 2; Column 4, for a specification that enriches Equation 2 with the triple interaction displayed previously. The inexperience score used in this triple interaction is computed at the patient/molecule level (Equation 3). This summarizes the amount of consumption experience that the patient has with respect to a given molecule, with lower scores reflecting more experience. CRMs are molecules targeting cholesterol in the case of LDL testers and molecules targeting diabetes in the cases of A1c testers. We estimate the models of Column 4 on the sample of 2012 purchases by 2012 testers, as described in the text. All models include fixed effects for molecules, individuals, and time periods. Parentheses show standard errors clustered at the level of patient/molecule pairs.
Does Bad News Matter More for Clinically Relevant Drugs?
Recall that we have defined CRMs as those drugs targeting the medical condition that is managed based on the results of the medical tests considered for our analyses. As such, CRMs correspond to cholesterol drugs for LDL testers and to diabetes drugs for A1c testers. Non-CRMs correspond to all other molecules covered by the data. Here we study how the bad news effect varies between CRMs and non-CRMs.
In the background section we highlighted two points derived from prior literature that rationalize the bad news effect operating on all drugs, including non-CRMs. The first is the observation that generics are perceived as inferior to brand name drugs in general, across the spectrum of all drugs used by patients. Second, emotional stimuli like bad medical news appear to function by altering how individuals weigh alternatives rather than what they know about them, meaning that the effect does not require choice-relevant information (e.g., weather impacts on stock returns). Nevertheless, we may still expect a stronger effect on CRMs, for example, if the bad testing results are particularly alarming for patients with a related condition.
To implement our analysis, we begin by formally classifying CRMs. We do so by parsing through products’ approved usages, as described by their FDA labels. This codification reveals that the cardiovascular class and the hormones and synthetic substitutes class contain many molecules in addition to CRMs. For LDL testers, CRMs (primarily statins) account for only about a quarter of purchases in the cardiovascular class. In turn, for A1c testers, CRMs (mainly metformin and related products) represent about three-quarters of purchases in the hormones and synthetic substitutes class.
Another important consideration is that, as discussed in the previous subsection, high cholesterol and diabetes are chronic conditions and thus generate persistent drug needs. In other words, patients use these drugs over long periods of time, acquiring high levels of consumption experience and knowledge about them compared with drugs used sporadically. This element is evidenced in Table 4, where the cardiovascular class and the hormones and synthetic substitutes class are associated with the least inexperienced patients across all six drug classes. These statistics suggest that accounting for consumption experience may be important to correctly estimate how the bad news effect varies between CRMs and non-CRMs.
In Panel B of Table 5 we present a series of estimates for the bad news effect obtained from the sample of CRMs. Estimates in Columns 1–3 show results for our main specification (Equation 2), separately estimated on the LDL, A1c, and pooled samples. The estimate obtained from the LDL sample (Column 1) is positive, although estimated with significant error and ultimately statistically nonsignificant. This lack of precision may be attributable to the small size of the sample available to estimate the effect. On the contrary, from the A1c (Column 2) and pooled (Column 3) samples we obtain negative estimates, which align with the presence of the bad news effect among CRMs. However, both of these estimates are statistically nonsignificant. Column 4 presents the estimates for the specification that incorporates the experience moderator. The obtained estimates describe a statistically significant bad news effect, one that is driven by consumption (in)experience. The estimate indicates that, for fully inexperienced patients (no prior purchases), the bad news shock reduces the probability of generic choice by about .06 (7%).
We next turn to estimating the bad news effect that operates on non-CRMs. Results are presented in Panel C of Table 5. The presence of these effects is supported by the negative estimates obtained from the LDL, A1c, and pooled samples (Columns 1–3, respectively), of which the former (LDL) and latter (pooled) are statistically significant (with 95% confidence). From the specification that incorporates experience effects (Column 4), estimates again suggest that the bad news effect is driven by the lack of consumption experience. The most important aspect of these parameter estimates pertains to their magnitude. Holding consumption experience constant, the bad news effect (Column 4) for non-CRMs is about one-fourth the size of its counterpart for CRMs. This indicates that even though the bad news effect seems to operate on non-CRMs, it is considerably smaller than that operating on CRMs.
To conclude, we draw attention back to Figure 2. Recall that this figure provides class-specific estimates of the bad news effect, for each of the six drug classes in the data. A key aspect of these estimates is that they do not control for consumption experience. As a result of this omission, we estimate relatively bad news effects of relatively large magnitude for some non-CRM classes (e.g., gastrointestinal, central nervous system). Combined with the results of this section, these estimates illustrate that controlling for consumption experience may be important to correctly estimate the bad news effect.
Implications for Practice
Our findings have implications for several key stakeholders in the health care industry. First, health policy makers, generic drug manufacturers, and insurers (public and private) all share the common goal of encouraging patients to choose generics over brand name drugs. To design policies aimed at achieving this goal, insurers currently rely on two primary toolkits. The first corresponds to a set of demographic and socioeconomic predictors of generic-averse attitudes, which are leveraged for the targeting of interventions. The second toolkit corresponds to possible intervention tools, which in practice boils down to a set of price-based promotional activities (e.g., discounts, coupons, free samples). Our analysis delivers important new insights on the application of these frameworks.
Concerning the targeting decision, our findings suggest that relying solely on demographic and socioeconomic predictors may lead to neglect of an important observable: the arrival of bad medical news. Accordingly, enriching the targeting framework with variables for recency with respect to these events could improve the targeting campaigns’ allocative efficiency.
Our analysis of brand/generic price differentials also raises a potential concern about the common use of price-based incentives to encourage use of generics. In line with prior literature (Dunne et al. 2014; Lambert et al. 1980; Verger et al. 2003), our results are consistent with the idea that patients may use prices to draw inferences about brand/generic quality. If this behavior is pervasive in the field (which is not confirmed by our analysis), it would introduce a previously unrecognized trade-off. Namely, a price incentive that increases the generic option's share-of-wallet appeal (e.g., through a discount) may also deteriorate its perceived quality. In such scenario, the optimal design of price-based campaigns would need to strike a delicate balance between the direct, short-term share-of-wallet effects and the indirect, more slowly unfolding potential impacts operating through quality inferences.
For providers and administrators, it may be helpful to consider strategies to neutralize the impacts of bad medical news on brand/generic preferences. A first step in this direction consists of generating awareness that even routine medical tests can trigger behavioral responses such as the one we have documented. Although the medical profession places marked emphasis on adequately breaking bad news to patients (Baile et al. 2000; Buckman 1992; Faulkner 1998), the traditional focus has been on cases related to severe outcomes (e.g., death, cancer diagnoses). Expanding this focus to include the much more subtle type of medical news that we consider could have a positive impact on patients’ well-being as well as on the system's efficiency. A simpler approach would be to remind patients of the equivalency of generic drugs following bad medical news, for example, via text messages after test results are shared with the patient (Pop-Eleches et al. 2011).
Finally, regulators should take note of the implications for brand name direct-to-consumer advertising, which routinely encourages patients to get tested for a variety of symptoms. Our results suggest that such advice is also consistent with the goal of hindering generic adoption. If drug manufacturers understand these mechanics, they may promote medical testing above and beyond medically justifiable levels. Such distortion would imply excess health care spending in terms of both additional testing and forgone savings from generic use.
Conclusion and Directions for Future Research
To reduce inefficient health care spending, the substitution of brand name with generic drugs is one of the policies that attracts close attention from both policy makers and insurers. Given that generic drugs are molecular replicas of their brand name counterparts, these policies could deliver large savings without sacrificing patient health. Nevertheless, these policies are met with resistance from the public, who exhibit a preference bias against generics. In this article, we contribute to the literature by uncovering a new source of this bias—bad medical news. Our evidence supports the idea that patients may become more reluctant to use generics when they receive news that highlights a deficiency in their health. Since receiving bad medical news is an inherent component of patients’ interaction with the health system, the identified effects might operate at a large scale and be responsible for a large amount of overspending.
One question that remains open from our analysis pertains to attribution: does the bad news effect reflect the patient's decision or the doctor's decision? When considering this question, we first note that our results related to the importance of patients’ consumption experience suggest that patients play a role in the bad news effect. However, since several prior studies find that doctors mediate the brand/generic decision (Hellerstein 1998; Iizuka 2012), our analysis cannot rule them out, and future research should aim to characterize their role. For example, to what extent do doctors acquiesce to patients’ pressure in favor of the brand name option?
Second, although our analysis demonstrates the existence and basic properties of the bad news effect, it does not disentangle the specific psychological mechanisms at play. Related literature highlights two possible mechanisms: patients could shy away from generics because they become more pessimistic about their health status or because they become more averse to health risks. 10 It is also possible that, rather than being motivated to avoid adverse health events, patients may start to aspire to improve their overall health condition. Eliciting the specific channel(s) at play could be helpful to inform the design of remedy interventions and guide future literature. Highlighting the difficulties of addressing the question based on observational data, Bassi, Colacito, and Fulghieri (2013) propose an experimental approach that could also be deployed in the context of the bad news effect.
Note that, despite the signs of generalizability provided by the consistency of the effects of bad LDL and A1c news, the treatment effect of bad medical news could vary outside these contexts. This is because the LDL and A1c settings have two key commonalities that may not be shared by other types of medical news: (1) the patient is not directly confronted with outcomes of utmost severity (e.g., death, losing a limb), and (2) one therapeutic alternative (brand name) strictly dominates the other (generic) in terms of perceived efficacy and safety. The effects of bad medical news could be qualitatively different from our evidence when these conditions are not met. For example, Harmon (2010) describes the story of two patients who, after being diagnosed with end-stage skin cancer, made a dramatic plea to be treated with an experimental drug of formally unverified properties. That is, confronting a highly probable death outcome, and lacking therapeutic alternatives, bad medical news could lead to patients choosing options of high associated risk. Exploring these treatment effect differences across decision contexts would be an important avenue to improve our understanding about the bad medical news effect.
We conclude by highlighting two limitations of our analysis. First, as noted previously, available data do not contain variables designed to measure individuals’ overall health status. As a result, we have relied on a health status proxy constructed from total health care expenditures (see Web Appendix A). Since this proxy could include a reverse causality bias, readers should be cautious when interpreting our results. Second, our data set lacks information about the individuals’ income level or health insurance plan. Although the absence of this information does not introduce a bias into our estimates (because of the randomized treatment), it prevents us from examining whether the response to the bad news shock could vary with income or insurance coverage.
Supplemental Material
sj-pdf-1-jmx-10.1177_00222429231158360 - Supplemental material for Does Bad Medical News Reduce Preferences for Generic Drugs?
Supplemental material, sj-pdf-1-jmx-10.1177_00222429231158360 for Does Bad Medical News Reduce Preferences for Generic Drugs? by Manuel Hermosilla and Andrew T. Ching in Journal of Marketing
Footnotes
Acknowledgments
This article was previously circulated under the title “Bad Medical News and the Aversion of Generic Drugs.” The authors thank the JM review team for their excellent feedback and guidance. The authors also thank Dan Goetz, Simha Mummalaneni, Haiyang Yang, Meng Zhu, Alice Gao, and David Powell for providing detailed comments and feedback, as well as conference and seminar participants at Johns Hopkins University, University of Miami, University of Illinois at Chicago, the 2022 International Industrial Organization Conference, the 2022 Marketing Science Conference, and the 2022 Annual Health Econometrics Workshop. The authors gratefully acknowledge the support of a Pilot Award Grant from the Hopkins Business of Health Initiative.
Special Issue Editor
Harald van Heerde
Associate Editor
Jie Zhang
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Hopkins Business of Health Initiative (Pilot Award Grant).
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
