Effectiveness of Behaviorally Informed Letters on Health Insurance Marketplace Enrollment

This randomized clinical trial investigates if low-cost behaviorally informed letters can increase health insurance enrollment across Affordable Care Act marketplaces.

INTERVENTIONS Study participants were randomized to either a no-letter control group or to 1 of 8 letter variants that drew on evidence from the behavioral sciences about what motivates individuals to take action.

Introduction
Through the expansion of Medicaid eligibility and the creation of health insurance marketplaces, the Affordable Care Act (ACA) has helped reduce the uninsured rate to record lows. 1 But every year during the open enrollment period, hundreds of thousands of individuals who initiate the enrollment process fail to complete it. Gaps in coverage or prolonged bouts of being uninsured cause disruptions in access to care and medication, increased financial strain, higher rates of medical debt, and lower levels of self-reported health. 2-4 Thus, identifying effective strategies to help individuals who have started the enrollment process obtain health insurance remains a priority for policy makers. 5 Barriers to health insurance take-up are well documented and include cost, application complexity, procrastination, a lack of awareness about available options, choice overload, and inertia. [6][7][8][9] A growing body of research seeks to understand how different forms of outreach can overcome these barriers to increase enrollment. A recent set of nonexperimental studies, for example, found an association between the volume of health insurance TV advertisements and reductions in the uninsured rate, as well as in ACA marketplace enrollment. 10,11 And randomized clinical trials (RCTs) have found that nudges using emails, letters, and telephone outreach increased health insurance take-up. [12][13][14] We build on this empirical evidence in 2 principal ways. First, in contrast with single-state RCTs, the present randomized intervention includes all 37 states that used the HealthCare.gov platform in 2015. The inclusion of multiple states is important within the context of the ACA, where states' policy decisions, such as Medicaid expansion, affect the cost of coverage and, in turn, whether individuals with low incomes can afford health insurance. Second, in lieu of nonexperimental studies that draw on self-reported survey data, we use administrative data paired with an RCT.
In the final weeks of the 2015 open enrollment period, we conducted an intent-to-treat RCT using behaviorally informed letters to increase health insurance enrollment among individuals who started the enrollment process but had yet to finish it. With 37 states and more than 744 500 individuals, this is, to our knowledge, one of the largest RCTs conducted on the ACA marketplaces to date, though there has been a larger RCT targeting tax filers who owed a positive penalty amount owing to the individual mandate. 14 Because letters are a low-cost option to reach a large number of uninsured individuals, they could represent a valuable tool for ACA marketplace administrators seeking to increase enrollment.

Study Design and Participants
This study used a parallel 9-arm design with 8 letter variants, each designed based on different insights from the behavioral science literature. The ninth arm was a hold-out control group that did not receive any letter, enabling us to measure the effect of receiving any letter as well as to tease apart the relative effect of the different behavioral features. The study followed the Consolidated Study participants were English-speaking individuals who, as of mid-January 2015, had visited HealthCare.gov and registered for a user account but not yet enrolled in an insurance plan. We chose mid-January as the cutoff to maximize the number of individuals eligible for the intervention while also leaving enough time to complete the requisite implementation steps so letters would arrive during the final 2 weeks of the open enrollment period.
Of the 811 795 individuals initially included, 18% were assigned to the no-letter control group, while the remaining 82% were assigned to 1 of 8 letter treatments ( Table 1; see eAppendix in Supplement 2 for copies of each of the letters used). The sample size and randomization scheme were chosen because HHS wanted to treat as many consumers as possible before the open enrollment period ended, while also learning about the effects of letter outreach.

Intervention
Individuals in the treatment arms were assigned to receive letters at the beginning of February 2015, giving them approximately 2 weeks to complete their enrollment. The 8 letters varied behavioral dynamics, including action language, an implementation intention prompt, a picture of then-chief executive officer of the marketplace Kevin Counihan, social norm messaging, a pledge, and loss aversion. These messages drew on evidence from prior randomized interventions that suggested these appeals would induce individuals to take action. 15

Randomization
Randomization was conducted by the first study author (D.Y.) based on user identification numbers using the sample function and a fixed seed in R, version 3.0.2 (R Foundation). The list with assignments was given to a contractor who mailed the letters.  because 67 285 individuals provided invalid mailing addresses, leaving them unable to receive letters or unable to enroll through the HealthCare.gov platform (Figure 1). In eTables 1 and 2 in Supplement 2, we show that the rate of invalid mailing addresses was approximately 8% across arms and was not correlated with treatment assignment. The administrative data also included pretreatment characteristics that we used to assess the validity of the random assignment and for stratification analyses, including self-reported race and ethnicity, state of residence, and age bracket.

Statistical Analysis
To estimate the effect of the letters overall and by subgroup, we used linear regression models with robust standard errors to account for heteroscedasticity. Data were analyzed using Stata, version 15 (StataCorp), and statistical significance was defined as a 2-sided P < .05.

Results
Of In exploratory analyses, we detected statistically significant differences across most subgroups, except for individuals younger than 30 years and those who did not provide a race and/or ethnicity when applying. The point estimate for Asian adults is substantively large (0.6 percentage points) but imprecisely estimated owing to a relatively small sample size ( Table 2). The largest enrollment  increase was among Hispanic adults, which was an increase of 0.7 percentage points (95% CI, 0.1-1.3 percentage points; P = .02), or 14%.
We additionally examined the effect of action letters by race and ethnicity and states' Medicaid expansion status (Figure 3). In expansion states, the effect of action letters was especially pronounced among racial and ethnic minorities, causing enrollment increases of 1.6 percentage points (95% CI, 0.6-2.7 percentage points; P = .003) among Black adults, 1.3 percentage points (95% CI, −0.3 to 2.8 percentage points; P = .11) among Asian adults, and 1.5 percentage points (95% CI, 0.0-3.0 percentage points; P = .046) among Hispanic adults, a pattern consistent with cost as an

Limitations
This research design is based on random assignment, which provides a strong basis for causal inference, but the study is not without limitations. First, owing to operational timelines, letters were only printed in English and sent to households with a written language preference of English; thus, we do not measure effects among harder-to-reach non-English-speaking households. 21 But because the intervention sought to address commonly cited barriers to enrollment, including procrastination and lack of awareness about the deadline or how to get help, we would expect the reminder letters to have comparable effects among Spanish-speaking individuals. Ultimately, though, this is an empirical question, and we encourage marketplace administrators to draw on promising experimental evidence-including in-language personalized telephone assistance, which has been found to considerably increase marketplace enrollment-during future open enrollment cycles. 22 Second, because race and ethnicity are optional questions on the ACA application, they are subject to missingness. In the present study, 62% of individuals answered these application questions. Third, letters were sent to individuals who took the initial steps of beginning the enrollment process.
have affected enrollment prior to the deadline, and it is possible that letters sent near the start of the open enrollment period could have had different effects.

Conclusions
In this RCT, we found that a low-cost letter, targeting individuals who took the first steps toward enrolling in ACA marketplace coverage but stopped short of selecting a plan, caused statistically significant and meaningful increases in ACA health insurance enrollment. From 2017 to 2020, enrollment in the ACA marketplaces declined from 12.2 million to 11.4 million in part because of a reduction in marketing and advertising. As the Biden administration seeks to expand coverage, particularly among racial and ethnic minorities hard hit by the COVID-19 pandemic, this study provides evidence that low-cost outreach-especially messages informed by the behavioral sciences-could help increase ACA marketplace enrollment.