2. Survey design

We estimate values for symptomatic kidney disease, including specific impacts from disease progression, a more broadly applicable endpoint for valuing the impacts of chemical and biotoxin exposure that can cause kidney disease. Specifically, we value the risk of stages 3 and 4 chronic kidney disease, labeled as “serious kidney disease,” and the probability that it progresses to kidney failure (stage 5).

The survey presents an accurate but succinct description of chronic kidney disease, its impacts, and its causes. The survey explains the symptoms associated with serious kidney disease (e.g. lower back pain, nausea, difficulty concentrating) and how kidney disease can lead to other major health complications (e.g. high blood pressure, cardiovascular disease, fluid in lungs, weakened ability to fight infections). The survey description also includes the 35 % likelihood that serious kidney disease can lead to permanent kidney failure and provides a simplified, but detailed, description of what is involved in being on dialysis, the chances of finding a donor kidney, and lifestyle changes and medical care needed following a kidney transplant. The description of kidney disease concludes with information on the impact of serious kidney disease on life expectancy.

Early pretesting made it clear that a scenario based on treatment was not viable because it raised questions about why the treatment was not covered by existing public health programs in many countries, and how the elicited payments related to those programs. Consequently, the survey relies on a scenario where the mechanism for reducing risk is not specified, but the costs are clearly out-of-pocket expenses. This approach has been used successfully in prior stated preference surveys (Krupnick et al., Reference Krupnick, Alberini, Cropper, Simon, O’Brien, George and Heintzelman2002; Hoffmann et al., Reference Hoffmann, Krupnick and Qin2017). Respondents were informed that only the risk of kidney disease would be affected and that no other effects would occur.

We designed the survey instrument through a multi-step process followed by all surveys in the SWACHE project. This started with a description of the health endpoint and an evaluation of how to characterize the endpoint in the survey, followed by the development of the risk reduction mechanism, the payment vehicle, and the elicitation method. A steering group of experts, including academics, BCA practitioners, regulators, and health professionals, provided regular feedback throughout the process.

The draft survey was initially tested in 51 one-on-one interviews across Chile, Denmark, Italy, and the United States. The near-final version of the survey instrument was translated into the necessary languages for all countries, with the translations verified by native speakers, and subjected to 50-person pilot tests in each country prior to it being fielded.

The survey uses a double-bounded dichotomous choice (DBDC) question with an open-ended follow-up question. Risks are characterized as chances per thousand over 5 years, displayed numerically and graphically, together with a short tutorial on risk, risk changes, and how these are presented in the survey. Aggregating risk over 5 years was necessary for risk changes to be plausible and effectively communicated to respondents given baseline risk levels for kidney diseases, particularly for the youngest age group.

Respondents for each country were divided into three age groups (18–40, 40–60, and over 60) to establish the baseline risk of serious kidney disease (15, 25, and 60 in 1,000).Footnote ² Members of each age group were randomly assigned into one of two hypothetical risk reductions (2 in 1,000 and 5 in 1,000) and one of four possible starting bids. The follow-up bid for the second dichotomous choice question was the first bid multiplied by either two or one-half, following conventional approaches as described in Carson and Hanneman (Reference Carson, Hanneman, Mäler and Vincent2006).Footnote ³

The bid structure is shown in Table 1.

Table 1.

Starting bids (total 5-year additional expenditure in USD)

The payment is stated as both an annual figure and as a total over the 5 years. An example of the valuation question is shown in Figure 1.

Figure 1.

Sample WTP question.

After the risk tutorial, 91.5 % of respondents correctly answered a simple test employing the risk communication device in which they were asked to identify the scenario with the highest probability of occurring.

All surveys in the SWACHE project share a common set of debriefing questions to identify common issues of concern. Key questions include whether respondents understood that the risk reductions were to occur only over the 5 years described in the survey and were not permanent, and whether they considered benefits beyond those described in the scenario.

3. Data

The survey was administered between June 2021 and June 2022 to a sample drawn from a large panel of individuals maintained by Ipsos, a global marketing and survey research firm. Representative samples were drawn for each country based on quotas matching key country-specific demographic characteristics: gender, age group, level of education, and geographic region. Respondents were screened to exclude those with chronic kidney disease. A total of 14,641 individuals started the survey after passing the screening questions, and 12,614 finished the survey. Ipsos removed 614 respondents due to a low-quality response score based on survey speeding (completing the survey in less than 1/3rd of the median survey time), straight-lining, and the proportion of “don’t know” answers, leaving a sample of 12,000.

The data were then screened based on core principles for empirical analysis agreed upon by the SWACHE researchers.Footnote ⁴ Screening included removing individuals who failed the survey’s probability test (8.47 %) and “speeders” who completed the survey or valuation questions in less than 48 % of the country-specific median time of the sample (12.4 %).Footnote ⁵ This screening is in addition to the screening done by Ipsos, and is based on the recommendations of Survey Sampling International (2013) and Mitchell (Reference Mitchell2014), and reduced the sample to 9,709 observations. Additional factors that might affect WTP were also identified, including gender, monthly household income, level of education, health expenditure paid for out of pocket, higher or lower average perceived health, having a friend or relative who experienced kidney failure, being diagnosed with some other chronic disease, and being diagnosed with or having a friend or relative diagnosed with COVID-19. These additional factors were used as explanatory variables in the econometric analysis.

Respondents were asked to indicate their household’s monthly income after income taxes had been paid and were presented with 10 income ranges corresponding to income deciles in their respective countries. Income deciles correspond to equivalized income. Unequivalized income deciles are derived by multiplying equivalized income deciles in 2019 from OECD Income (IDD) database by the number of “equivalent adults” using data on family composition from OECD Family database.Footnote ⁶ Respondents who did not indicate their range for the income deciles were presented with bigger ranges corresponding to income quintiles in their respective countries. Most respondents (91 %) provided information about the total income of their household. Income ranges were then converted into a single amount to facilitate the use of income data in the empirical analysis. For the smallest income range between 0 and the first decile, the income is set equal to 0.5 times the first decile. For the largest income range above the last decile, the income is computed as equal to 1.5 times the top decile. For all the other income ranges, the computed income is the simple average between the two deciles. All income values were then converted into purchasing power parity adjusted U.S. dollars (USD PPP) using PPP for actual individual consumption data for 2019 from the PPPs and exchange rates OECD database.Footnote ⁷

We imputed missing income responses by performing country-level Ordinary Least Square (OLS) regression analyses of logged income as a function of indicators for age, couples, females, high education, number of people in the household, employment, part-time work, and whether the respondent was retired.

Table 2 presents the sampling target and the achieved sample for target characteristics of the screened sample. The table shows an over-representation of females in all countries except Denmark, Norway, and Türkiye. This over-representation was present in the full sample received by Ipsos but was larger after our screening. The achieved sample is close to the sample target for age groups. China and Türkiye show an under-representation and Demark an over-representation of individuals over 60. Low and medium levels of education were collapsed into a single group because of an under-representation of lower-educated respondents in all countries. Higher educated respondents are close to the target except for Norway and Türkiye.

Table 2.

Sample quota versus achieved sample for chronic kidney disease survey

A respondent’s monthly income is the average of the income range selected, converted to USD PPP. The median monthly income for Canada, Denmark, Italy, Norway, and the United States closely matches the population median. The achieved sample for Chile, Germany, and the United Kingdom show a slightly higher deviation, with an under-representation of low-income individuals. The two outliers are Türkiye and China, which show much higher incomes in the achieved samples than the population medians. The sample median income is 162 % higher than the target for Türkiye and 227 % higher for China.

Given that the Chinese sample has a median income of more than 225 % of the Chinese population median and that 28 % of the Chinese respondents completed the survey and valuation question exceptionally quickly, China was removed from the baseline analysis, but is included as a sensitivity analysis in the main results table and in the country-level parametric estimations of WTP.

To test for sensitivity to scope and scale, Tables 3 shows the response to the first dichotomous choice question is broken down by the risk reduction offered and the starting bid. For presentation purposes, the starting bid is presented as the 5-year cost for U.S. respondents, as described in Table 1. In practice, the starting bid is a continuous variable because the U.S. dollar values were rounded after being converted to the local currency. For both risk reduction values, the per cent of respondents who answered “Yes” to the first question declines with the starting bid, indicating that respondents are less likely to be willing to pay for a risk reduction as the bid increases. This is the expected sensitivity to scale. Additionally, for the three starting bids that are common to both risk reduction values, the per cent of respondents who answered “Yes” is higher for the larger risk reduction. This suggests that respondents are willing to pay more for a larger risk reduction, which is the expected sensitivity to scope.

Table 3.

Response to first dichotomous choice question by starting bid

Tables 4 and 5 display the summary statistics for the baseline sample of 8,905 respondents who remained after screening those who failed the probability test or were identified as speeders and after removing China. Table 4 provides the summary statistics for the important continuous variables used in the analysis. The first block of data (5 rows) shows the time to complete the survey and the time respondents spent on the two dichotomous choice questions. Respondents answered the second risk question twice as fast as the first one, probably due to increased understanding after answering the first question. The second block of data shows the bid values for the first and second dichotomous choice questions, with the second bid being either one-half or twice the first bid value. Twenty-five per cent of respondents chose the lower risk (responded “yes” to pay the cost) for both risk questions, and 40 % of respondents chose the current risk (responded “no” to pay the bid price) for both questions.

Table 4.

Summary statistics for continuous variables

Table 5.

Summary statistics for indicator variables

The last rows in Table 4 provide statistics for baseline risk and income, continuous variables considered to be important factors affecting willingness to pay. The baseline risk is 15, 25, or 60, depending on the age group. The PPP-adjusted reported median income ranges from $263 (in Chile) to $17,552 (in the U.S.). Almost 10 % of the baseline sample declined to report their income, but the income with the imputed values used for the missing responses produces a very similar income distribution and provides 8,905 observations for the estimation.

Table 5 provides the summary statistics for the indicator (0/1) variables used in the analysis. The table reports the number of observations for which the variable is equal to one, and the per cent of the total 8,905 observations. The first block on the left-hand side reports the observations for each country after screening for speeders and those who failed the probability test. China is not included in this list because it was removed from the baseline sample. Denmark, Türkiye, and the U.S. have a slightly lower per cent of the observations because of a larger number of speeders and respondents who failed the probability test. The second two blocks on the left-hand side of Table 5 show the age–gender distribution of the baseline sample. As noted previously, there is an over-representation of females in this sample.

The first block on the right-hand side of Table 5 shows the distribution of responses to both risk questions. Twenty-five per cent of respondents chose the lower risk (responded “yes” to pay the cost) for both risk questions, and 40 % of respondents chose the current risk (responded “no” to pay the bid price) for both questions. Approximately 15 % and 20 % of the respondents answered “yes-no” and “no-yes” to the first and second dichotomous choice questions, respectively.

The second block on the right-hand side provides statistics for the indicator variables considered important factors affecting willingness to pay. A large per cent of respondents (38.7 %) believe that their health is better than average, while 15.9 % self-report lower than average health. The majority of respondents (65.8 %) have never been diagnosed with a chronic disease, but over a third (34.8 %) had a relative or friend with kidney disease. Only 8.5 % of respondents report having been diagnosed with COVID-19, but almost 40 % have a close friend or relative who had been diagnosed with COVID-19.

The final block on the right-hand side of Table 5 contains the statistics for controls used to test the robustness of the screening analysis. 5.8 % of the respondents reported that they thought the risk reduction was permanent. 9.2 % considered changes in other health issues not described in the survey (co-benefits) when they made their choices. Eleven per cent of the respondents strongly agreed (on a 5-point Likert scale) with the statement that they would pay almost any amount to reduce risks to their health. Combining those who strongly agreed to pay almost anything with the 25 % of respondents who answered Yes-Yes to the two dichotomous choice questions produces 6.2 % of respondents classified as yea-sayers. 3.4 % of respondents reported that they did not answer as they would have in real life or said that the survey did not provide them enough information to make informed choices. These respondents were classified as protesters.

4. Estimation model

To derive mean and median WTP estimates for a reduction in the risk of serious kidney disease, we employ a maximum likelihood estimator using the binary outcomes of the double-bounded dichotomous choice.

We assume a Weibull distribution for the utility error as our baseline because it generally has a shorter right tail than the log-normal and, in its “spike” configuration, usually performs well (Kriström, Reference Kriström1997; Carson & Hanneman, Reference Carson, Hanneman, Mäler and Vincent2006).

We employ a spike model to account for the possibility that a small share of the population might still choose the status quo even if it costs them nothing. This spike could be significant in the case of serious kidney disease because the baseline risk is relatively small for younger respondents. Carson and Hanemann argue that failing to include a spike parameter can, in some cases, lead to overestimating WTP.

We measure the spike by using responses to the dichotomous question and responses to the open-ended question that followed the double-bounded dichotomous choice: “What would be the most you would be willing to pay, if anything, to reduce your chance of getting serious kidney disease within 5 years?”. The spike dummy variable equals one when respondent $ i $ chooses no to both valuation questions and responds with a value of zero to the open-ended questions.

A Weibull distribution $ \theta =\left\{k,\lambda \right\} $ is characterized by a shape parameter $ k $ and a scale parameter $ \lambda $ . All estimations assume a shape parameter equal to 1. The baseline specification of the scale parameter when $ b>0 $ is

(1) $$ {\lambda}_{ic}(b)={\alpha}_0+{\alpha}_1{\Delta R}_i+{\alpha}_2\mathrm{ln}\,{b}_i+\sum \limits_c{\delta}_c({d}_{ic}\times {\omega}_i) $$

where $ {\Delta R}_i $ is the risk reduction proposed to respondent $ i $ , $ \ln {b}_i $ is the logged cost or bid proposed to respondent $ i $ , $ {d}_{ic} $ is a country dummy equal to 1 when respondent $ i $ lives in country c, and $ {\omega}_i $ is the post-stratification weight of respondent $ i $ . Including post-stratification weights, $ {\omega}_i $ , as a control allows us to capture the fact that some categories of people were slightly under- or over-represented in the sample compared to the actual population.Footnote ⁸ The more respondent $ i $ is underrepresented in the sample, the higher his weight $ {\omega}_i $ . It is necessary to interact country dummies with the weights because the weights are defined at the country level.

The spike parameter when $ b=0 $ is

(2) $$ {\eta}_{ic}={\alpha}_0+{\alpha}_1{\Delta R}_i+\sum \limits_c{\delta}_c\left({d}_{ic}\times {\omega}_i\right). $$

The model is also estimated when the scale parameter includes additional explanatory variables as follows:

(3) $$ {\displaystyle \begin{array}{l}{\lambda}_{ic}(b)={\alpha}_0+{\alpha}_1{\Delta R}_i+{\alpha}_2\ln {b}_i+\sum \limits_c{\delta}_c\left({d}_{ic}\times {\omega}_i\right)+{\alpha}_3{\mathrm{Female}}_i+{\alpha}_4\ln {y}_i\\ {}\hskip2.4pc +{\alpha}_5{\mathrm{HighEduc}}_i+{\alpha}_6{\mathrm{Baseline}}_i\end{array}} $$

$ {\mathrm{Female}}_i $ is a dummy variable equal to 1 when respondent $ i $ identifies as a female, $ \ln {y}_i $ is the logged monthly income for the household of respondent $ i $ , $ {\mathrm{HighEduc}}_i $ is a dummy variable equal to 1 when respondent $ i $ achieved high education outcome, and $ {\mathrm{Baseline}}_i $ is the baseline risk presented to respondent $ i $ .

The model is also estimated when the scale parameter includes information on other important factors. These factors include whether respondents must pay for health expenditures out-of-pocket; whether they perceive their health as below or above the average of people of their gender and age; whether they know a relative who had kidney failure; whether they are diagnosed with any other chronic disease; and whether they or a relative was ever diagnosed with COVID-19.

The mean WTP for an average 3.5 in 1,000 risk reduction in serious kidney disease is computed as a simple average of the individual mean WTP follows:

(4) $$ \hat{WTP}=\frac{1}{n}\sum \limits_{i=1}^n{\hat{WTP}}_i $$

The individual mean WTP is computed by integrating the probability of responding yes to the valuation question over the interval from 0 to maximum bid with adjustment:

(5) $$ {\hat{WTP}}_i=\int_0^{b_{max}}\frac{f\left({\lambda}_{ic}(b),k\right)}{1-f\left({\lambda}_{ic}\left({b}_{max}\right),k\right)} db $$

where $ f $ is the density function of the Weibull distribution and k denotes the shape parameter. Truncation at maximum bid level $ {b}_{max} $ is necessary since the right tail is not null when the cost goes to infinity. The adjustment of the denominator compensates for the fact that the support of $ f\left({\lambda}_{ic}(b),k\right) $ does not stop at $ {b}_{max} $ . The median WTP is computed as a simple average of individual median WTP as follows:

(6) $$ {\overset{\sim }{WTP}}_i=\frac{\ln 2}{\left|{\alpha}_2\right|}{e}^{\eta_{ic}\left(\frac{1}{\left|{\alpha}_2\right|}\right)} $$

$ {\alpha}_2 $ is the parameter for the logged bid value as indicated in equation (1).

Mean and median WTP are also presented as the value of a statistical case of kidney disease avoided. VSC is calculated using the standard method of dividing annual WTP by the average annual risk reduction. Because this survey asks about risk changes over a 5-year period and payments over this same period, VSC can equivalently be calculated as 5 x annual WTP divided by the average 5-year risk reduction of 3.5 per 1,000.