FormalPara Key Points

Methodological differences in value sets can lead to significant disparities when calculating quality-adjusted life-years.

This study provides economists with accurate health valuations for COVID-19 that can be used to analyze different health interventions.

Future research should focus on the preferences of hospitalized patients to fully understand the burden of COVID-19.

1 Introduction

Economic evaluations of medical interventions are an important tool for many healthcare systems around the world. Stakeholders allocate healthcare resources based on the information these evaluations provide. Quality-adjusted life-years (QALYs) serve as a unit of measurement for the burden of disease studies and are an integral part of the evaluative process. Quality-adjusted life-years can range from negative infinity to 1, depending on the study. Typically, 0 corresponds to “dying immediately” and 1 represents “1 year with no health problems then die.” Negative values would represent health outcomes that are worse than death. Health economists continue to strive for improved accuracy, applicability, and generalizability of QALY values.

Quality-adjusted life-year values can be calculated using value sets for the EQ-5D-5L instrument [1]. The EQ-5D-5L was introduced in 2005 and is an updated version of the EQ-5D-3L instrument allowing for greater sensitivity of measurement [2]. The EQ-5D-5L contains five dimensions: mobility, self-care, usual activities, anxiety/depression, and pain/discomfort; and 5 levels: no problems (level 1), slight problems, moderate problems, severe problems, and extreme problems (level 5). Respondents would indicate for each dimension which level corresponds to their current health state. The health state “11111” indicates no health problems in any domain and “55555” indicates extreme problems in all domains. In addition to these five items, the instrument includes the EQ-VAS, a visual analog scale (VAS) that captures self-reported health status from best (100) to worst (0) imaginable health.

To summarize EQ-5D-5L responses as QALY values, researchers typically use a published value set (also called tariffs). These value sets are based on health valuation studies and each study may use different elicitation techniques such as time trade-offs (TTOs) or discrete choice experiments. Apart from the elicitation technique, there are significant differences among value sets depending on the analysis and country of origin [3,4,5,6].

Obtaining accurate QALY values for specific diseases can also be a challenge if EQ-5D-5L responses for individuals with that disease are unknown. COVID-19 outcomes present an added challenge because of their recency and heterogeneous presentation. Figure 1 is a simplistic description of two risk groups who may contract COVID-19 and how moving from one stage to another could constitute losses or gains of QALYs. This cyclical model implies knowledge of the COVID-19 burden at each node; however, QALY values for these outcomes have yet to be published to the best of our knowledge.

Fig. 1
figure 1

COVID-19 outcomes

Full size image

The goal of our study is to compare the distribution of QALY values using a national survey and the two published value sets and to estimate the association between COVID-19 outcomes and QALY losses using a generalized linear model (GLM). Based on these associations, we will calculate the QALY values for the six groups in Fig. 1 for use in future economic evaluations and burden of disease analyses.

2 Methods

This study is a secondary analysis of a protocol that received an institutional review board exemption.

2.1 US EQ-5D-5L Value Sets

The first US EQ-5D-5L valuation study was conducted in 2016 with 8222 respondents from all 50 states of USA and the District of Columbia [7]. It comprised an online survey that included a discrete choice experiment, namely a paired comparison including the five EQ-5D-5L attributes and lifespan. During the analysis, the authors allowed for discounting of future health outcomes. The second study was conducted in 2017 with 1062 respondents from six metropolitan areas (Chicago, Philadelphia, Seattle, Birmingham, Phoenix, and Denver) [8]. It comprised an interview that included a composite TTO task, namely adaptive paired comparisons with the same six attributes trading off life-years. During the analysis, the authors did not allow for discounting. The first study also conducted a predictive model competition and differs from the standard EQ-VT protocol and the second study followed the EQ-VT protocol [7, 8]. Both studies were funded by the EuroQol Research Foundation; however, the second study also received financial support from the pharmaceutical industry, but the funding agreement ensured independence. We refer to each value set by its main methodological difference, i.e., discounted vs undiscounted.

The lack of discounting in health valuations causes QALY values of “poor” health to be lower (downward bias) [6]. For example, each TTO response equates 10 years with EQ-5D-5L attributes to less time in full health; therefore, “good” health trades long-term life-years and “poor” health trades near-term life-years. Under discounting, long-term life-years are typically worth less than near-term life-years. Likewise, the lack of discounting causes the unadjusted TTO analysis to produce biased estimates for poor health, known as its constant proportionality assumption [3]. These value sets can be found in Appendix 3 of the Electronic Supplementary Material.

2.1.1 Survey Instrument

The survey instrument comprised five sections: consent, screener, background, choice task, and follow-up (Screenshots) [9]. This paper is on information obtained from the screener, background, and follow-up portions of the survey, namely the EQ-5D-5L instrument. The screener included five questions that provided information on respondent demographics (Screenshots, p. 1–2). The background asked ten questions related to COVID-19 and whether the respondent was “a member of the following critical populations?” These included: person at increased risk of acquiring or transmitting COVID-19, healthcare personnel, person with limited access to routine vaccination services, person at increased risk for severe COVID-19 illness, and/or other essential worker (Screenshots, p. 3). Within the follow-up questions, the EQ-5D-5L instrument was included (five items in a fixed order followed by a VAS from best to worst imaginable health), as were other questions related to health of the respondent. Respondents were asked if they “had a fever in the past day,” “have a new or worsening cough,” and “any of these other symptoms” (Screenshots, p. 17–24).

On 5 November, 2020, the institutional review board at Advarra® determined that the study (Pro00047418) was exempt according to 45 CFR 46.104(d) [2]. The recruitment occurred from 9 to 11 November, 2020 by Dynata®, an online market research firm, rendering a national sample for US adults aged 18–87 years, including persons from the District of Columbia and all US states, except South Dakota. During recruitment, two exclusion criteria were applied: (1) residing outside the 50 US states or the District of Columbia and (2) being age 17 years or less (age 18 years in Nebraska or Alabama or age 20 years in Mississippi).

2.1.2 Analysis

Our analysis plan was as follows: (1) comparison of survey respondent characteristics to the 2019 American Community Survey; (2) association between VAS deciles and QALY values from the EQ-5D-5L responses and the two US value sets; (3) association between demographic variables and health/risk variables and losses in QALY values using a GLM; and (4) calculation of QALY values of COVID-19 groups adjusted for respondent demographics.

2.1.3 GLM of QALY Losses

Apart from the log link for the expected loss in QALYs (i.e., 1 minus QALY), the GLM family function was determined by completing a modified Park test [10]. This approach found that a gamma distribution was best to control for skewness and heteroskedasticity in unexplained losses [11,12,13].

For the base case, we selected 45 years of age; therefore, included adjusted age [i.e., (age in years − 45)/45] and its square in the GLM regression. Other demographic variables included indicators for female/other, Black or African American alone, Asian or other, and Hispanic or Latino. Thus, the constant term (0) represents a 45-year-old, non-Hispanic, white male US adult.

For the health variables, the GLM regression included indicators for whether a respondent was in a high-risk population, was considered to have had clinical COVID-19, and had a fever, cough, or at least one other symptom. A respondent was deemed high risk if they replied that they were at “increased risk for acquiring or transmitting COVID-19” or at “increased risk for severe COVID-19 illness.” Respondents were deemed to have had clinical COVID-19 if they selected either “I tested positive for COVID-19” or “a doctor ordered me to quarantine for possible COVID-19.” Respondents were categorized as non-clinical COVID if they did not test positive for COVID-19 or were not ordered to quarantine by their doctor because of COVID-19. Respondents had one or more symptoms if they indicated they had one of the following: congestion or runny nose, nausea or vomiting, new loss of taste or smell, headache, diarrhea, shortness of breath or difficulty breathing, muscle or body aches, fatigue, and sore throat.

3 Results

3.1 Respondent Characteristics

We first compared the respondent demographics to the 2019 American Community Survey using a chi-square test and found a few significant differences (Table 1) [14]. Respondents were more likely to be 35–54 years of age (38.68% vs 32.43%; p < 0.01), white (76.67% vs 73.61%; p < 0.01), and non-Hispanic (87.25% vs 83.60%; p < 0.01), and these difference may affect the interpretation of the COVID-19 QALY values.

Table 1 Demographic characteristics
Full size table

3.2 Comparison of US EQ-5D-5L Values

In Fig. 2, the y axis is the mean QALY values by EQ-VAS decile of the two US value sets. Frequency of respondent per decile ranged from 8.07 to 11.10%. The discounted value set produced a mean value that ranged from 0.86 to 0.98. The mean values of the undiscounted value set ranged lower, from 0.53 to 0.94 (p < 0.01) (Table 2).

Fig. 2
figure 2

Quality-adjusted life-year (QALY) values by EQ-VAS decile (mean and 95% confidence interval)

Full size image
Table 2 Generalized linear model analysis of ln QALY losses
Full size table

3.3 GLM Analysis of COVID-19 QALY Losses

The GLM analysis of COVID-19 QALY losses provides two models for each US EQ-5D-5L value set, (1) demographics only and (2) demographics with health/risk variables. The coefficients are reported as changes in ln QALY losses. For example, in the base case for model 3, a 45-year-old, non-Hispanic, white male US adult would face a ln QALY loss of − 1.65. We can calculate the QALY for the base case as 1 − \({e}^{-1.65}\), which equals a QALY of 0.81. When interpreting the coefficients, positive values would mean increases in QALY losses (i.e., lower QALYs) and negative values would mean decreases in QALY losses (i.e., higher QALYs). In terms of demographics, models 1 and 3 show significant differences in adjusted age (p < 0.01), but only model 3 shows slight significant differences by race (p < 0.10). After accounting for health/risk variables, models 2 and 4 find significant increases in QALY losses for high-risk individuals (0.55 and 0.66; p < 0.01), those with a fever in the past day (0.61 and 0.74; p < 0.01), and one or more other symptoms (0.79 and 0.84; p < 0.01).

3.4 QALY Values for COVID-19 Outcomes

Table 3 provides the marginal effects of different groups controlling for demographics based on the GLM analysis results. We then report the results as 1 minus the coefficient to convert into QALYs to make it easier for the reader. All QALY losses are statistically significant (p < 0.01). The discounted values range from 0.68 (95% confidence interval [CI] 0.49–0.87) to 0.98 (95% CI 0.97–0.98) after controlling for demographic variables. The undiscounted values are lower for all groups, from 0.10 (95% CI − 0.31 to 0.51) to 0.91 (95% CI 0.90–0.92). In summary, a symptomatic person with COVID-19 is in “fair” health (0.68) based on the discounted value set and close to death (0.10) based on the undiscounted value set.

Table 3 QALY values for COVID-19 outcomes
Full size table

4 Discussion

The differences in QALY values between the two US EQ-5D-5L value sets are apparent based on the VAS deciles or COVID-19 outcomes. When comparing mean QALY values in Fig. 2, the undiscounted values have several areas where they are non-monotonic, and the 95% CI is much wider than the discounted values. This is to be expected owing to the differences in discounting, representation, and sample size. There continues to be methodological advancement as economists strive for greater accuracy in health valuation. One area in particular is the assumption of linear time preferences and how this may violate basic human behavior. Recent studies have shown that allowing for nonlinear time preferences, through the use of discounting functions, provides a better statistical fit [6, 15]. These differences could lead to misallocation of resources and incorrect implementation of interventions against COVID-19 depending on which value set is chosen. Health economists will have to decide between the conventional approach of using undiscounted values, which would be biased, or choosing discounted values that are more accurate.

There were also dramatic losses of QALYs for those non-institutionalized individuals in the at-risk group that developed COVID-19 symptoms. When using the undiscounted QALY values, we see that these individuals are at a state close to zero (dead). COVID-19 is a unique disease, there are differences not only in infection rate and severity when compared with influenza, but some individuals experience long-term side effects [16, 17]. We can clearly see the enormous potential gains in QALYs by preventing symptomatic COVID-19 in individuals at high risk and those not at high risk.

4.1 Limitations

Our study is limited in the generalizability to the entire US population as there are significant differences between the survey respondents and the American Community Survey. This is because there are individuals that are more likely to take online surveys. We thought about using sampling weights but determining the proper sampling weights can be difficult and we questioned if this would actually make the data more representative of the population. We also do not have parameters for hospitalized patients with COVID-19 within this study, thus we are unable to predict the full burden of the disease. This study also only includes six different subgroups. This could be an issue if these do not align with future modeling studies. We also must be cognizant of when this survey was taken by respondents. This occurred right after the 2020 US election; therefore, the context of the survey is important when understanding a disease that has been highly politicized.

5 Conclusions

This study is one of the first to determine the QALY values for COVID-19 outcomes in the USA or elsewhere. Health economists will be able to use this information to better model the merits of health interventions against COVID-19. There are already studies taking place that use QALYs to determine the cost effectiveness of different COVID-19 suppression strategies. These measures may not be accurate as they are not using QALYs adjusted for COVID-19, instead using valuations adjusted for pneumonia, which is only one of the potential complications of this disease [18]. Future research should explore the preferences of hospitalized patients and caregivers to fully understand the burden of COVID-19.