Trading Bankruptcy for Health: A Discrete Choice Experiment

Abstract

Background:

Although nearly two-third of bankruptcy in the United States is medical in origin, a common assumption is that individuals facing a potentially lethal disease opt for cure at any cost. This assumption has never been tested, and knowledge of how the American population values a trade-off between cure and bankruptcy is unknown.

Objectives:

To determine the relative importance among the general American population of improved health versus improved financial risk protection, and to determine the impact of demographics on these preferences.

Methods:

A discrete choice experiment was performed with 2359 members of the US population. Respondents were asked to value treatments with varying chances of cure and bankruptcy in the presence of a lethal disease. Latent class analysis with concomitant variables was performed, weighted for national representativeness. Sensitivity analyses were undertaken to test the robustness of the results.

Results:

It was found that 31.3% of the American population values cure at all costs. Nevertheless, for 8.5% of the US population, financial solvency dominates concerns for health in medical decision making.Individuals who value cure at all costs are more likely to have had experience with serious disease and to be women.No demographic characteristics significantly predicted individuals who value solvency over cure.

Conclusions:

Although the average American values cure more than financial solvency, a cure-at-all-costs rubric describes the preferences of a minority of the population, and 1 in 12 value financial protection over any chances of cure. This study provides empirical evidence for how the US population values a trade-off between avoiding adverse health outcomes and facing bankruptcy. These findings bring to the fore the decision making that individuals face in balancing the acute financial burden of health care access.

Introduction

That the American health care system is expensive is well known [1]. Nevertheless, discussing the out-of-pocket costs of care is often anathema, because any implications of care rationing are thought to defy a respect for health [2,3]. As Hall stated, “When we are ill, we desperately want our doctors to do everything within their power to heal us, regardless of the costs involved” [4]. This cure-at-all-costs presupposition has led to thorny ethical debates [4] but has rarely been tested [3].

In the United States, 62% of bankruptcy is medical, and, despite the fact that most medical costs are paid for by insurers, more than 75% of medically bankrupt patients were insured at the time of their catastrophic medical bill [5]. Although financial risk falls on all patients, medical bankruptcy is more frequent among the poor and patients with life-threatening conditions [6,7].

The World Health Organization [8], the United Nations [9], and the World Bank [10] have called for financial protection in health, but medical impoverishment persists, in part because individuals are willing to risk debt for medical care [11–13] and because health systems pay less attention to financial risk than to clinical risk. Although the high incidence of medical bankruptcy shows that some patients will face financial hardship to seek medical care [14–20], other potential patients choose noncompliance or to forgo care altogether because of high costs [21–23]. In patients with serious conditions, these decisions can be lethal [23,24].

Patients then face an implicit trade-off between financial protection and health protection, and health policies do not affect these two domains equally. The Oregon Medicaid Experiment, for example, provided coverage to previously uninsured Oregonians. After 2 years, improvements in health outcomes were limited, but significant improvement was seen for every reported measure of medical impoverishment [25]. Similarly, recent evidence from states that expanded Medicaid in 2014 also shows a short-lived increase in medical utilization [26] and improvements in financial risk protection [27], but no change in self-reported health [27,28].

The design of policy interventions would benefit from an understanding of how patients make this implicit trade-off, given that not all patients are willing to face the double burden of financial and medical toxicity [29].

How much bankruptcy risk individuals are willing to shoulder in seeking care is unknown, nor is it known how individual characteristics such as age, income, family composition, health status, and education influence this decision. This article explores these questions using a discrete choice experiment (DCE).

Methods

A DCE was performed with the goal of determining how much increased risk of bankruptcy an individual would be willing to face for an increased chance of cure. DCEs, described in detail elsewhere [30,31] and in the Appendix in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2017.07.006, are grounded in random utility theory. Formally, let Y represent the choice between two alternatives 0 and 1. Then:

$$Y=I\left(U_1>U_0\right),$$

where I(∙) represents an indicator function, taking the value of 1 if the expression in parentheses is true and 0 otherwise, and U₁ and U₀ represent the utilities of the two altesrnatives.

Because utility is unobservable, U_i; for each choice i is decomposed into a deterministic (observable) portion V_i, and a random (unobservable) portion η_i :

$$U_i=V_i+{\eta }_i$$

Given a set of observed choices among alternatives and an assumption about the underlying error distribution, V_i can be estimated.

This article’s hypothesis was that when cure from a lethal condition was possible, individuals would be willing to trade high risks of financial catastrophe to seek it—that is, in short, patients would value cure “regardless of the costs involved” [4]. Secondarily, we hypothesized that preferences would be influenced by age, sex, income, family structure, and experience with serious disease.

Examined Models

Three possible utility functions were evaluated in the study. The first (as well as the simplest and the most commonly used) is linear. In such a formulation, the utility for individual n of an alternative j is:

$$U_{nj}={\beta }_1{\text{SOLV}}_j+{\beta }_2{\text{CURE}}_j+{\eta }_{nj},$$ (1)

where CURE_j represents the probability of cure for the jth alternative and SOLV_j represents the probability of remaining solvent (i.e., 1 – the probability of bankruptcy).

Because each DCE question offered respondents two choices (j = 1 or 2), the respondent would select the first choice if U_n1 > U_n2. From the responses to the survey, population-level values for β₁ and β₂—and therefore a population utility function—can be estimated.

Nevertheless, the simple linear model is not intuitive: if, for example, an individual has a very low chance of survival, he or she might be more inclined to take larger financial risks than if he or she had a high chance of survival. A multiplicative formulation would allow this nuance:

$$U_{nj}={\mathrm{SOLV}}_j^{{\beta }_1}\times {\mathrm{CURE}}_j^{{\beta }_2}\times {\epsilon }_{nj}.$$ (2)

A utility function grounded in expected utility theory [32] would also allow the aforementioned intuitive interaction:

$$\begin{gathered} U_{nj}={\mathrm{SOLV}}_j{\mathrm{CURE}}_j{\beta }_0+\left(1-{\mathrm{SOLV}}_j\right){\mathrm{CURE}}_j{\beta }_1+{\mathrm{SOLV}}_j\left(1-{\mathrm{CURE}}_j\right){\beta }_2 \\ +\left(1-{\text{SOLV}}_j\right)\left(1-{\text{CURE}}_j\right){\beta }_3+{\epsilon }_{nj}. \end{gathered}$$ (3)

In this formulation, β₀ through β₃ represent an individual’s utility for a state of being after the choice has been made. β₀ represents his or her utility for being cured and remaining financially solvent. In β₁, the individual has been cured but has gone bankrupt as a result. Similarly, β₂ represents the utility for remaining solvent but succumbing to the lethal disease, whereas β₃ represents utility for bankruptcy and death.

Because utilities are unique up to positive affine transformations, two of the β values in Equation 3 can be set arbitrarily. An obvious choice is β₀ = 1 (for “Cured and Solvent”) and β₃ = 0 (for “Dead and Bankrupt”). These choices simplify Equation 3 to:

$$\begin{gathered} U_{nj}={\mathrm{SOLV}}_j{\mathrm{CURE}}_j+\left(1-{\mathrm{SOLV}}_j\right){\mathrm{CURE}}_j{\beta }_1 \\ +{\text{SOLV}}_j\left(1-{\text{CURE}}_j\right){\beta }_2+{\epsilon }_{nj}. \end{gathered}$$ (4)

Because β₁ and β₂ represent states that are worse than “Cured and Solvent” but, in theory, better than “Dead and Bankrupt,” it is expected that they take values between 0 and 1. Nevertheless, note that this formulation does not constrain β₁ and β₂ to any value. An individual who viewed “Dead and Not Bankrupt” as worse than “Dead and Bankrupt,” for example, would have β₂ less than 0.

Further discussions of the theoretical underpinning of these models can be found in the Appendix in Supplemental Materials.

Class and Model Selection

Latent class analysis allows for the possibility that there is more than one value across the population for each β utility (or “taste”) parameters [33]. Specifically, latent class analysis assumes that the population is made up of distinct segments (“classes”), each with their own values for β₁ and β₂ Moreover, the likelihood that an individual falls into one or another class can be predicted by that individual’s demographic characteristics. The probability π_cn, that individual n falls into class c can be calculated as a fractional logit model:

$${\pi }_{\mathit{\text{cn}}}=\frac{\exp \left({\theta }_{\text{c}}{z}_n\right)}{{\sum }_{a=1}^{C-1}\exp \left({\theta }_a{z}_n\right)},$$ (5)

where C is the total number of classes and z_n represents individual demographic characteristics. The best-fitting model, encompassing both class number and utility function formulation, can be selected using, in the case of this study, the Bayesian information criterion. Note that latent class analysis does not assign each individual to a particular class, but assigns to each individual a probability of membership in every class.

Survey Design

Paired-comparison surveys give respondents a choice between two discrete scenarios, differentiated along parameters of interest [31]. In this study, the survey instructed respondents to imagine that they had a hypothetical condition, lethal without treatment. They were asked to choose between two treatments, identical in every way except for their probability of a cure and their risk of driving the individual into bankruptcy. “Cure” and “bankruptcy” were explicitly defined. Respondents were told that the disease may or may not return after “cure” but that they could not know this at present, just as they could not know if some future event would drive them into bankruptcy.

Each probability of interest had five levels—10%, 25%, 50%, 75%, and 90%. Because the presence of certainty may introduce cognitive bias [34], and because no realistic medical intervention would have either a 0% or 100% chance of cure or bankruptcy, the ends of the scale were not included.

At the start of the survey, respondents were asked to rank the following possible outcomes of a choice, presented in random order: “Cured and Not Bankrupt,” “Cured and Bankrupt,” “Dead and Not Bankrupt,” and “Dead and Bankrupt.” Participants then received 12 paired-comparison questions (see example in Appendix Table S1 in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2017.07.006). To test consistency and understanding, the first two paired-comparison questions had one logically dominant alternative. If respondents picked the nondominant alternative in either question, they received a prompt alerting them to this fact. They received this prompt only once, even if they picked the non-dominated alternative both times. Irrespective of their answers on these validation questions, participants were allowed to complete the entire survey.

Data on age, sex, marital status, head-of-household status, number of children, employment status, education level, race, ethnicity, region, urban/rural location, self-reported health, tobacco use, and alcohol use were collected, and respondents were asked about experience with a lethal disease. A copy of the survey is in the Appendix in Supplemental Materials.

Sample and Validation

The survey was pilot-tested in two samples of respondents from Cambridge, MA. In the first pilot, a convenience sample of 10 respondents with either medical or health policy training was given a full-factorial version of the survey. Upon completion, comments were elicited. No respondent had difficulty with understanding but most complained of fatigue on answering 25 questions.

The second pilot group was composed of 50 respondents, matched to the sex, age, education level, race, and income of the population of Cambridge, MA, and recruited through the Harvard Decision Science Lab. This group took a mix-and-match fractional factorial blocked version [35] of the survey, with five blocks of five questions each. The purpose of this group was to establish a prior estimate for the relative importance of financial and health protection, which could then be used in sample size calculations. These respondents were also invited to make comments on the survey.

Responses from the second pilot group were analyzed using a linear conditional logit model (Equation 1). Respondents attached 1.8 times as much weight to cure as to avoiding bankruptcy, that is, β₂/β₁ = 1.8. Simulation was then performed to determine the sample size required to find a ratio between 1.5 and 2.0 with 80% power. A sample of approximately 1000 respondents was found to be necessary, consistent with formal sample size calculations performed for other valuation studies [30,31]. To allow for sub-group analysis and error in the initial estimate, we increased the desired sample size to 2200.

The final blocked survey was distributed to English-speaking Americans older than 18 years through GfK (formerly Knowledge Networks), which maintains a nationally representative [36] panel of respondents. Recruitment to this panel is initially done through a random sampling of addresses (as opposed to traditional random-digit dialing) [37]. Mailings and calls to address-matched phone numbers are then used to recruit respondents to the panel, and respondents are sampled with known probabilities of selection. This sampling frame covers 97% of all residential addresses in the United States [38], and the methodology ensures inclusion of respondents with only mobile phones (who are missed on random-digit dialing) and those without Internet (who are missed by Internet-only recruitment strategies). Furthermore, this sampling strategy explicitly avoids the problems encountered by Internet-based convenience samples for which participants can volunteer.

Once respondents are sampled for any survey, GfK provides response weights in accordance with the most recent March supplement of the Current Population Survey, using the following dimensions: sex (male/female), age (18–29, 30–44, 45–59, and > 60 years), race/Hispanic ethnicity (white/Non-Hispanic, black/non-Hispanic, other/non-Hispanic, 2+ races/non-Hispanic, and Hispanic), education (less than high school, high school, some college, and bachelor’s degree and higher), census region (Northeast, Midwest, South, West), household income (<$10,000; $10,000−<$25,000; $25,000− <$50,000; $50,000−<$75,000; $75,000− < $100,000; and > $100,000), home ownership status (own, rent/other), residence in a metropolitan area (yes/no), and Internet access (yes/no). Households without Internet access on initial enrollment in the panel are provided both access to the Internet and hardware, if needed.

GfK surveys, both with and without weighting for national representativeness, have been used in many previous health analyses [38–44]. No significant differences existed in the demographic characteristics of individuals receiving any of the five blocks (see Appendix Table S2 in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2017.07.006). From the completed questionnaires, a nationally representative sample was resampled on the basis of the probability-proportional-to-size-based weights provided by GfK. Results from both the nationally representative sample and the unweighted sample are presented.

Analysis

The underlying goal of this analysis was to determine a utility function describing the trade-off between the probabilities of cure and of solvency. The national average utility function was determined but because it is unlikely that the entire population has uniform preferences, responses were primarily analyzed using a latent class conditional logit model with concomitant variables [33,45]. From the resulting segment-specific preference weights, indifference curves for each class can be constructed. The ratio of weights is proportional to the ratio of importance respondents assign to the cure and financial solvency. The final number of classes into which the population is segmented is an exercise in model fitting. In this article, the selection of class number is performed using the Bayesian information criteria [33]. All collected demographic characteristics were evaluated as concomitant variables in Equation 5. The association between concomitant variables and class membership is evaluated as a fractional multinomial logit model, as discussed earlier [33].

The interpretation of indifference curves is relatively straight-forward. The probability of remaining financially solvent after an intervention is plotted along the vertical axis, whereas the probability of cure from an intervention is plotted along the horizontal axis. Individuals value any point on a single indifference curve equally, with curves toward the upper right preferred to curves toward the lower left. At the steeper portions of any one curve, an individual is willing to trade a greater risk of bankruptcy for a small increase in cure than on flatter portions of the curve. Nevertheless, the shape of the curves themselves gives information on the global trade-off—the more vertical the curves are overall, the more an underlying cure-at-all-costs preference structure would hold. Conversely, the more horizontal a curve, the more likely respondents would trade health protection to avoid bankruptcy.

To test stability of the results and to elucidate properties of the estimator itself, two simulation-based sensitivity analyses were performed. In the first analysis, a new set of responses was randomly generated. A well-functioning model run on this random data set should yield estimated coefficients on parameters of interest that are not significantly different from 0. The second tested the ability of the model selection procedure to select the correct underlying coefficients. To do so, responses were generated from a known underlying utility function and preference estimation was repeated. The postestimation predictive power of the model was also tested.

Statistical analyses were performed in R version 3.0 (R Foundation for Statistical Computing, Vienna, Austria), using the mlogit package [46], and in Stata (StataCorp, College Station, TX), using the lclogit package [33]. Demographic comparisons were performed using t tests, Mann-Whitney U tests, χ² tests, and analysis of variance, when appropriate. This study was deemed exempt by the Institutional Review Board of the Harvard TH Chan School of Public Health.

Results

Panel Characteristics, Validation, and Rank-Ordering Exercise

The survey was distributed to 4918 adult respondents, of whom 2975 completed it (response rate 60.5%). Six hundred sixteen surveys were discarded because they omitted more than one-third of the 12 discrete choice questions (597 answered no discrete choice questions; 19 answered at least one but fewer than four), leaving 2359 surveys for analysis (usable response rate 48.0%). Of note, including the 19 surveys with limited responses did not change the results.

Compared with nonrespondents, respondents with complete surveys averaged 1.6 years older, were 7% more likely to be male (52% vs. 45%), 2% more likely to identify as head of household, 5% more likely to be married or living with a partner, 10% more likely to have at least a bachelor’s degree, and 15% more likely to identify as white (Table 1). Median self-reported income was identical among groups, although the mean was statistically significantly higher for the group completing the survey. Employment outside of the home and Internet access at home were not different among the two groups.

Table 1 –

Respondents’ demographic characterstics.

Characteristic	Completes	SD	Others*	SD	P value
Total	2359		2559
Age	57.55	17.826	55.93	19.024	0.002
If >65 y	71†	65–94	71†	65–94
If <65 y	44†	18–64	41†	18–64
Age > 65 y	0.50	0.500	0.48	0.500	0.168
Sex, male	0.52	0.500	0.45	0.497	<0.001
Head of household	0.85	0.357	0.83	0.378	0.025
Married or living with partner	0.65	0.476	0.60	0.489	<0.001
Number of children	2.42	1.338	-	-
Any children	0.71	0.206	-	-
Education
Bachelor’s degree or higher	0.36	0.480	0.26	0.437	< 0.001
Employed outside the home	0.37	0.484	0.37	0.482	0.633
Race/ethnicity
White, non-Hispanic	0.80	0.399	0.65	0.477	< 0.001
Black, non-Hispanic	0.07	0.249	0.13	0.335	< 0.001
Other, non-Hispanic	0.05	0.225	0.08	0.270	< 0.001
Hispanic	0.08	0.270	0.14	0.350	< 0.001
Health
Self-reported health (10 = highest)	8†	0–10	-	-
Smoker	0.13	0.333	-	-
Alcohol consumption > 1 drink/mo	0.45	0.498	-	-
Experience with serious disease	0.20	0.404	-	-
Urban	0.85	0.358	0.85	0.361	0.827
Region
Northeast	0.18	0.387	0.17	0.379	0.397
Midwest	0.23	0.423	0.22	0.414	0.243
South	0.35	0.478	0.38	0.485	0.048
West	0.23	0.422	0.23	0.419	0.740
Employed outside the home	0.37	0.484	0.37	0.482	0.633
Income ($)	75,698	55,251	67,690	53,353	< 0.001

^*Incompletes and nonresponders. Health-related questions were not available from individuals without complete results.

^†Median.

The median self-reported health of the 2359 respondents was 8, on a scale where 10 represents perfect health and 0 represents death. The median age of respondents younger than 65 years was 44 years; for those older than 65 years, the median age was 71 years. The 2014 US census estimates report a median age of 41 years for individuals younger than 65 years and 73 years for those older than 65 years. Compared with the US population, the sample was more likely to be non-Hispanic white (81% vs. 63.7%), male (52% vs. 49%), and less likely to hold at least a bachelor’s degree (36% vs. 39.4%). Nevertheless, among the respondents in the nationally representative sample, 64% were white, 12% black, 16% Hispanic, and 7.6% of other or mixed race; 38.9% of the respondents had a bachelor’s degree or higher; and 50.8% were females.

In the two validation paired comparisons, 87.4% of respondents picked the dominant alternative when it was first presented; 91.3% did so after prompting. These proportions were nearly identical in both the unweighted and the nationally representative samples (see Appendix in Supplemental Materials for unweighted results).

If the central hypothesis—that the American population prefers cure at all costs—is true, then respondents should rank the possible outcomes of a choice in the following order: “Cured and Not Bankrupt,” “Cured and Bankrupt,” “Dead and Not Bankrupt,” and “Dead and Bankrupt.” This ordering was chosen by 68.0% of the respondents. The second most common ranking, chosen by 16.8% of respondents, reversed the two middle outcomes, contradicting the central hypothesis (see Appendix Table S3 in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2017.07.006). Individuals with the most common rank-ordering of outcomes had a median self-reported health of 8, higher than the median self-reported health of 7 among the remaining respondents (P < 0.001). There was no significant difference in the number of children at home. Income levels were no different between individuals with the most common rank-ordering and those with other orderings.

The Preferences and Compositions of Subgroups in the American Population

Of the three examined models (Equations 1, 2, and 4), the expected utility model (Equation 4) performed the best. In the overall population, health protection was found to be 5.10 times (95% confidence interval [CI] 4.48–5.84; unweighted 5.52, 95% CI 4.36–7.18) more important than financial risk protection (Table 2; see also Appendix Figure S1 in Supplemental Materials found at http://dx.doi.org/10.1016Zj.jval.2017.07.006). The data, however, suggested that the population is composed of distinct segments (Table 3; see also Appendix Figure S2 in Supplemental Materials found at http://dx.doi.org/10.1016/j.jval.2017.07.006), with heterogeneous preferences (Table 4). All examined models best fit the data when the population was segmented into either five or six classes. In the expected utility model, which is the best fit to the data in general, the lowest Bayesian information criterion value was achieved when the population was segmented into six classes (Table 3). The variance matrix, however, became highly singular at six or more classes, and so the five-class segmentation is presented instead.

Table 2 –

Unsegmented population preferences, using the expected utility model (Equation 4).

Alternative	Coefficient	Robust SE	95% CI	P for difference
(A) Weighted for national representativeness
Cure	0.612	0.013	0.586–0.638	<0.001
Solv	0.120	0.005	0.109–0.131
BIC	15,185
(B) Unweighted
Cure	0.618	0.040	0.540–0.696	<0.001
Solv	0.112	0.012	0.090–0.135
BIC	17,386

Note. Cure is the probability of cure, Solv is the probability of remaining solvent, and P for difference is the statistical significance of the difference between the coefficients on Solv and Cure. The coefficient on Solv is significantly smaller than the coefficient on Cure (P < 0.001), implying thatthe nation as a whole puts a premium on health protection; bankruptcy protection is important, but less so. The unsegmented model fits less well than the segmented models of utility seen in Table 5. Note that the coefficients represented here are scaled such that “Cured and Solvent” has a utility of 1.0

BIC, Bayesian information criterion; CI, confidence interval; SE, standard error.

Table 3 –

BIC for population segmentations of 2 through 7 classes, for the expected utility model, with responses weighted for national representativeness.

Number of classes	BIC
2	15,185.25
3	15,533.41
4	15,196.33
5	14,585.08
6	14,426.23
7	14,497.01

BIC, Bayesian information criterion.

Table 4 –

Preferences for the five segments of the population.

Segment	Coefficient	SE	95% CI	Class share	Relative importance of cure (95% CI)	P for difference
(A) Weighted for national representativeness
Class 1				0.085	0.03*	<0.001
Cure	−0.016	0.033	−0.081 to 0.049
Solv	0.577	0.043	0.493 to 0.662
Class 2				0.212	2.11 (1.20–4.11)	<0.001
Cure	0.166	0.018	0.131 to 0.200
Solv	0.079	0.015	0.049 to 0.109
Class 3				0.108	0.52† (0.04–1.63)	0.127
Cure	2.445	1.123	0.244 to 4.645
Solv	4.666	0.929	2.846 to 6.486
Class 4				0.313	157.9*	< 0.001
Cure	1.001	0.102	0.801 to 1.202
Solv	0.006	0.031	−0.054 to 0.067
Class 5				0.282	4.69 (3.41–6.62)	< 0.001
Cure	0.591	0.038	0.517 to 0.666
Solv	0.126	0.013	0.100 to 0.151
(B) Unweighted
Class 1				0.302	5.83(4.49–7.94)	< 0.001
Cure	0.561	0.038	0.487 to 0.635
Solv	0.096	0.012	0.073 to 0.119
Class 2				0.214	1.09(0.69–1.76)	0.704
Cure	0.174	0.027	0.121 to 0.227
Solv	0.160	0.027	0.108 to 0.212
Class 3				0.066	0.01*	< 0.001
Cure	0.008	0.059	−0.107 to 0.123
Solv	0.697	0.080	0.540 to 0.854
Class 4				0.118	0.93† (0.01–3.87)	0.904
Cure	7.846	3.837	0.326 to 15.365
Solv	8.450	3.243	2.094 to 14.807
Class 5				0.3	7.59 (4.50–22.60)	< 0.001
Cure	0.662	0.015	0.631 to 0.692
Solv	0.087	0.030	0.029 to 0.145

Note. Solv is the probability of remaining fiscally solvent, Cure is the probability of cure, Class share is the proportion of the population that falls into each segment, and P for difference is the statistical significance of the difference between the coefficients on Solv and Cure within each segment. The ratio of coefficients gives the relative weight that individuals place on cure vs. financial solvency. That is,individuals represented by class 1 find cure 5.8 times more important than solvency, where as individuals represented by class 3 find solvency 87.4 times more important than cure. Note that the coefficients represented here are scaled such that “Cured and Solvent” has a utility of 1.0.

CI, confidence interval; SE, standard error.

^*Because the 95% CI for individual preference weights crosses 0 in classes 1 and 4, the 95% CI for the importance ratio is undefined.

^†See text for a discussion of class 3 (Table 4A) and class 4 (Table 4B).

In the sample weighted for national representativeness, approximately one-third of the American population (Table 4A) prefers cure at all costs (class 4, class share 31.3%; β₂ = 0), whereas the national preference weighting in Table 2A reflects the preferences of another 28.2% of the population (class 5, relative importance of cure 4.69; 95% CI 3.41–6.62). Nevertheless, 21.2% of the population weight health protection and financial protection more equally (class 2, relative importance of cure 2.11; 95% CI 1.20–4.11) and, most strikingly, 8.5% of the population considers financial protection to dominate health outcomes in medical decisions. This latter group (class 1) is essentially willing to risk all health improvement to remain financially solvent. Indifference curves for the population segments are shown in Figure 1. (Note that the 95% CIs for classes 1 and 4 in Table 4A are undefined because the 95% CI for the preference weight on cure or solvency, respectively, crosses 0.)

Fig. 1 –

Indifference curves for segments of the American population, with their share of the population. See the main text for a discussion of the omitted classes. (A) Weighted for national representativeness. (B) Unweighted.

The preference weights for class 3 imply that individuals in this class would prefer death or bankruptcy to a state of being both cured and financially solvent. Compared with class 5, who represent the national average, having answered the validation questions incorrectly is significantly associated with membership in class 3 (P < 0.0001), as is worse self-reported health (P = 0.001), experience with serious disease (P = 0.03), and alcohol usage (P = 0.017). Respondents in class 3 are more likely to identify as Hispanic (P = 0.005) or of other or mixed race (P = 0.001) than do respondents in class 5 (Table 5).

Table 5 –

Concomitant variables.

Variable	Class 1	Class 2	Class 3	Class 4
(A) Weighted for national representativeness
Sex, female	0.142	0.031	0.408*	0.438†
Age	0.012	0.001	0.01	−0.002
Health (0–10)	−0.169	−0.067	−0.289†	−0.056
Previous serious disease	−0.092	−0.084	−0.24*	−0.306†
Serious disease or caretaker	0.152	0.136	−0.94	−0.217
Number of kids	0.018	−0.1‡	0.051	−0.074
Understood (0–10)	−0.042	−0.054	0.064	0.031
Education	−0.138	−0.096	0.119	0.02
Valid	0.298	0.212	−3.804†	0.543‡
Midwest	0.174	0.596§	0.152	0.375‡
South	0.027	0.624§	−0.175	0.539§
West	0.12	0.561§	−0.087	0.179
Tobacco (Y/N)	0.196	−0.781§	−0.529‡	−0.377
Alcohol (Y/N)	−0.187	−0.085	−0.446*	0.209
Married or living with partner	−0.161	−0.109	0.257	−0.241‡
Hispanic	−0.457	0.002	−0.832‡	−0.4‡
Non-Hispanic black	−0.277	−0.067	−0.276	0.066
Other race	0.313	0.162	1.168†	0.364
Income (logged)	0.02	0.026	−0.097	−0.005
Class share	0.085	0.212	0.108	0.313
(B) Unweighted
Sex, female	−0.246	−0.226	−0.258	−0.136
Age	−0.002	0	0.003	0.001
Health (0–10)	0.014	−0.051	−0.121	−0.153‡
Previous serious disease	0.141	0.151‡	0.15	0.171
Serious disease or caretaker	−0.08	0.131	0.067	0.121
Number of kids	0.021	0.022	0.032	0.042
Understood (0–10)	−0.014	−0.057	−0.034	0.007
Education	−0.142	−0.192	−0.437	0.042
Valid	−0.367	0.262	0.08	−3.674‡
Midwest	0.07	0.237	−0.216	−0.19
South	−0.132	0.062	−0.367	−0.381
West	0.038	0.262	−0.194	−0.117
Tobacco (Y/N)	0.17	−0.176	0.477	−0.465
Alcohol (Y/N)	−0.182	−0.123	−0.361	−0.321
Married or living with partner	0.163	0.185	0.054	0.378
Hispanic	0.016	0.054	−0.341	−0.515
Non-Hispanic black	0.058	−0.472	0.02	−0.18
Other race	0.082	−0.237	−0.213	0.757‡
Income (logged)	−0.079‡	−0.026	0.128	−0.106
Class share	0.302	0.214	0.066	0.118

Note. Class 5 is the reference class. Coefficients on these demographic variables can be interpreted as likelihoods of being in classes 1–4 when compared with the individuals in class 5, in which cure is valued most strongly. Understood is the self-reported understanding of the survey (0–10), Valid specifies that the validation questions were answered correctly, and Class share is the proportion of the population falling into each class.

^*Significant at <0.05.

^†Significant at <0.001.

^‡Significant at <0.10.

^§Significant at <0.01.

The female sex (P = 0.001), residence in the southern United States, and experience with serious disease (P < 0.0001) predicted cure-at-all-costs (class 4) preferences as opposed to preferences consistent with the national average, whereas no demographic characteristics predicted membership in class 1, whose members prefer financial solvency at all costs (Table 5).

In the unweighted sample (Table 4B), the national preference weighting from Table 2B reflects the preferences of only 30.2% of the population (class 1, relative importance of cure 5.83; 95% CI 4.49–7.94), whereas another 30% put slightly more weight on cure (class 5, relative importance of cure 7.59; 95% CI 4.50–22.60). Nevertheless, 21.4% of the population weight health protection and financial protection with equal importance (class 2, relative importance of cure 1.09; 95% CI 0.69–1.76) and 6.6% of the population considers financial protection to be nearly 90 times more important than health protection (class 3, relative importance of cure 0.01). In the unweighted sample, no demographic characteristics predict class membership significantly (Table 5B).

The robustness of the results to the inclusion of not-fully-complete surveys has been discussed earlier. Other sensitivity analyses are discussed in the Appendix in Supplemental Materials. They suggest that the estimator behaves as expected: when responses are generated from known underlying utility functions or at random, the estimation returns coefficients to match. Results from the unweighted sample are also given in the Appendix in Supplemental Materials, as are further tests of the model’s predictive capabilities (Appendix Table S4).

Discussion

The results of a nationally representative analysis suggest that a cure-at-all-costs rubric represents the preferences of only a subset of the American population. Although the population as a whole appears willing to trade financial risk for cure in the setting of lethal disease, this preference is not uniform. For one-fifth of respondents in this sample, health is only twice as important as financial solvency, and, most strikingly, nearly one-tenth of the population will trade all health to maintain financial protection.

Not having had experience with serious disease strongly predicts membership in the class of individuals who value cure at all costs. This may imply a potential change in preference when individuals face the impact of serious disease. Women and people living in the South are more likely to have these preferences as well. More interesting, however, is the fact that, barring respondents with invalid responses and those who value cure at all costs, demographic characteristics do not significantly predict individual preferences. Specifically, nothing significantly predicts membership in the class of respondents who value financial solvency more than any increase in cure, highlighting the importance of discussions of financial risk protection during the course of medical decision making.

As with any survey, ours has limitations. We cannot, for example, determine which behavioral model underpins these responses. It is possible that the experience of the rich is one of relatively good health care, and so a treatment with a 10% cure rate is not interpreted the same way as it would be by someone below the national poverty line. Similarly, a 90% risk of bankruptcy may be unimaginable to individuals who are largely well off. Although our survey did not specifically ask about the insurance status of respondents, employment outside the home was not associated with any additional explanatory power in our models.

The preference for financial risk protection over improved chances of cure is likely to be country- and context-specific. We have asked the general American public and have weighted the sample to be nationally representative, but evidence exists that individuals with lethal diseases are more willing than the general public to accept medically toxic treatments if they offer even minimal benefit [47,48]. Whether these patients would also be willing to accept financially toxic interventions is a matter for future research.

What respondents say they prefer (stated preference) may differ from how they act when actually faced with a decision (revealed preference). DCEs are stated-preference techniques; their direct applicability to patient behavior is unclear. Nevertheless, although decisions in the moment are subject to external pressures that may distort an individual’s preference, Gamble et al. [49] suggest that this distortion in medicine may be small. The value of the present study, however, is in its delineation of different preferences within the general population. To find how these preferences change when individuals are directly faced with the reality of lethal disease is beyond its scope.

Although the results presented are from a nationally weighted sample of the American population, generalizability from a few thousand respondents to more than 330 million must be done with caution, especially given that the survey was limited to respondents who spoke English. The size of this DCE, however, is consistent with others that have been performed to elicit national preferences [30,31]. It should be noted that although weighting for national representativeness did not change the number of classes into which the respondents segmented, it sharpened the estimated preferences. In the unweighted sample, no class preferred either cure or solvency at all costs, but these preferences became evident after weighting for national representativeness. National weighting, however, cannot fully over-come the limitations inherent in any sampling frame. Although GfK’s address-based sampling frame covers 97% of the US population and although computers and Internet access are provided free of charge to respondents who did not previously have it, no sampling frame likely fully captures the poorest, the sickest, the most disabled, and the most disadvantaged. Results must be interpreted with this in mind.

Finally, the questions in this DCE present medical treatment and financial catastrophe as the result of a single choice made by the individual for the individual. Although the third most common reason for medical bankruptcy in the United States is a single-event, acute injury [5], medical bankruptcy can accrue over time, and medical choices are not limited to those made by individuals for themselves. The results from this analysis cannot be generalized to decisions involving loved ones.

Getting medical care in the United States can be risky. Although emphasis has been placed on the risks of iatrogenic injury [50], patients face a risk to more than their health. Arguably, financial ruin is an equally far-reaching iatrogenic “complication”: a declaration of bankruptcy is associated with the loss of assets, including nonprimary homes and inheritances, as well as the degradation of an individual’s credit rating for 7 to 10 years.

The results of this survey highlight the fact that a large subset of the American population cares more about financial risk protection than may have previously been assumed. Although these results confirm that Americans value cure more than financial solvency, they highlight the fact that only a minority (30%) value cure to the exclusion of financial risk protection. Most are willing to trade some amount of cure for financial solvency, and 1 in 12 would not be willing to shoulder any risk of bankruptcy, even for improvements in health. As a result, discussions of the out-of-pocket costs of care with patients cannot be avoided, and providing financial risk protection that supersedes that commonly provided by insurance may be necessary in the design of a rational health system. These results should give policymakers useful information for policy development and should encourage physicians either to discuss the potential risk of financial toxicity with patients or to involve financial counselors early in health interventions.

Conclusions

Although Americans, as a whole, view health protection as more important than financial risk protection, a cure-at-all-costs mentality likely represents only a subset of the nation’s preferences. In fact, 1 in 12 appear to prefer financial solvency at all costs. These findings bring the financial burden of health care access in the United States to the fore and must be kept in mind as clinical and policy interventions are considered.