What public health interventions do people in Canada prefer to fund? A discrete choice experiment

Participants for our study were recruited online using paid advertising on Facebook and Instagram between May 13th, 2020 and June 15th, 2020, during which time Canadians were exiting the third wave of the COVID-19 pandemic and seeing breaking national headlines about mass burial sites at Indigenous residential schools [34, 35] (both of which are factors that we acknowledge may have considerable potential to influence perceptions on healthcare programming). Recruitment preceded for 1 month to support engagement of infrequent social media users. After clicking on a paid advertisement in either English or French, participants were screened for eligibility. Eligibility criteria restricted participation to individuals who (1) were 16 years of age or older, (2) reported living in Canada, (3) provided informed consent, and (4) were able to complete the questionnaire in English or French.

The survey questionnaire was hosted on the Qualtrics Platform. After being screened for eligibility and providing informed consent, participants participated in a discrete choice experiment (See Fig. 1). The experiment consisted of 8 exercises in which participants were asked to imagine that they were a decision maker choosing which of two public health programs should be funded. Each program was described by its (1) purpose, (2) expected increase in life expectancy, and its (3) target audience. JavaScript was used to randomly generate the descriptors for each program from a list of 12 program purposes (i.e., To prevent cancer; To prevent diabetes; To prevent drug overdoses; To prevent heart disease; To prevent mental health problems; To prevent sexually transmitted infections, including HIV; To treat cancer; To treat diabetes; To treat heart disease; To treat mental health problems; To treat sexually transmitted infections, including HIV; To treat substance use and addiction), 10 life expectancy estimates (i.e., 1–10 years), and 12 target populations (i.e., General Population; Gay, bisexual, and other men who have sex with men; People who use drugs; Indigenous people; African, Caribbean, and Black people; People engaged in sex work; Migrants and refugees; People living in or recently released from correctional facilities; Transgender people; People living with HIV; Youth and Young Adults; Seniors and Older Adults). The study was piloted with a group of 10 undergraduate and graduate students and a pilot study with 52 responses was used to identify issues or participant concerns/complaints prior to launching the study. Slight revisions were made based on these data and thus they were not included in the present study.

To adjust for sampling biases, we conducted iterative proportional fitting using the anesrake package in R (v. 4.0.3). Iterative proportional fitting is a statistical weighting approach that allows for the construction of a single weight based on several participant characteristics [36]. It is advantageous when the population distribution of each weighted variable is known, but when the intersections of these weights are uncertain [37]. The raking algorithm used to iteratively fit the population proportion assigns weights repeatedly to achieve convergence until all target variables are within 5% of the target weights. Using iterative proportional fitting allowed us to generate statistical weights that could be used in our descriptive and multivariable analyses. Target weights, reported in Table 1, were based off the Canadian Census Population (adjusting for age, gender, ethnicity, educational attainment, income, province of residence). As the data were collected via an online, opt-in survey, we also adjusted for the average time participants spent using social media (based on estimates from the Ryerson Social media Lab) [38] and political orientation (based on estimates from CBC News Poll Tracker Averages) [39]. Descriptive statistics were used to provide an overall description of the sample and assess how frequently programs with each characteristics were selected when participants were presented with an opportunity to choose programs with each given characteristic. A weighted generalized linear mixed-effects model was constructed using the lme4 package to identify program factors associated with program selection preferences. The outcome factor indicated whether the program was selected. A random effect term was used to account for within-person effects arising from the repeated observations (i.e., up to 8 per participant). Explanatory factors included measures of the program purpose, the target population, whether the intervention focused on a prevention or treatment approach, and the relative years of life gained for the selective (vs. unselected) program. An interaction term between the program purpose and target audience was included. Person-level characteristics were not analyzed in our multivariable model as all program characteristics were randomized to participants. Categorical data are regressed in the lme4 using the contrast schemes (also called dummy coding). As dummy variable coding yields the same estimates for variable attribute differences as the alternative approach (i.e., effects coding), we opted to retain the contrasts scheme as it is widely used in the public health field, currently implement in R, and offered a relatively easy interpretation given our research questions [40, 41]. Given our use of the dummy coding scheme, readers should be cautious to interpret categorical variables relative to the selected reference-levels (which were selected based on interpretability and consistency in illustrating our hypotheses). Results are reported consistent with the ESTIMATE checklist, which represents the ISPOR guidelines and recommendations for the conduction of discrete choice experiments [42, 43].

We conducted a discrete choice experiment asking participants to select between competing public health programs and aimed to identify program characteristics associated with the selected programs. Supporting previous findings and current practice in cost-benefit analyses [22, 23, 44], the results of our discrete choice experiment showed that participants were more likely to select interventions that contributed to greater gains in life expectancy [24, 26, 45]. We also observed that participants were more likely to select programs focused on treatment as opposed to prevention. The existing literature on this issue has been mixed, with data from the Swiss Household Panel Survey suggesting that the majority of citizens support preventative interventions over treatment-based ones [46] while other discrete choice experiments have shown a preference for treatments of more severe illness over less severe illness [33]. Our findings support the latter, showing that participants prioritize treatment over prevention.

Extending the existing literature, our study also examined the role of specific health conditions and target populations in shaping funding preferences. In doing so, we found that participants were less likely to select interventions targeting stigmatized health conditions (e.g., HIV and Other STIs) and more likely to select interventions targeting relatively more common chronic health conditions, such as heart disease and diabetes. This conforms with other research that highlight the role of stigma in shaping attitudes towards specific health conditions [47‐49]. Similarly, participants were less likely to select interventions tailored to key populations, favoring instead those that were tailored to the general population. These findings align with previous research by Skedgel, Wailoo, and Akehurst (2015) which showed that participants generally support interventions that maximize life expectancy gains, but that they are willing to prioritize smaller gains to preferred groups over larger gains to less preferred groups [24]. These findings therefore highlight potential mechanisms in which marginalized groups are disadvantaged by public opinion. Likewise, Norman et al. (2013) showed that participants tend to favor programs for individuals like themselves, which given our sample would contribute to the less support for marginalized individuals [50]. Optimistically, the present study highlights several notable exceptions to this general pattern: Participants were more likely to select interventions tailored to populations widely known or believed to be affected by specific health conditions. For example, interventions addressing HIV and other STIs were more likely to be selected when they were targeted to gbMSM, migrants and refugees, people engaged in sex work, or people who are or have been incarcerated but were less likely to be selected when tailored to seniors and older adults. This finding aligns with research by Purtle (2020) showing that the public generally believes that evidence should have “a lot of influence” on health policy decisions [51]. Likewise, these findings conform to work by Shmueli et al. (2017) and Olsen & Richardson (2013) which demonstrate the public’s support for equity and equality in healthcare priorities [52, 53]. Importantly, we note that participants support for these programs may arise from both accurate (i.e., awareness of epidemiological burden) or inaccurate (i.e., cultural assumptions about the “riskiness” of marginalized groups) perceptions of the need for care within these populations. Certainly, gbMSM and other key populations do experience a disproportionate burden of disease from HIV and other STIs, but these burdens are experienced regardless of whether they are recognized by the public. For example, substance use interventions targeted to Indigenous people were not more likely to be selected in our study, despite the fact that tailored and targeted programs are needed to support Indigenous people who use drugs [54‐56]. Likewise, tailored and targeted interventions addressing mental health and substance use are needed for gbMSM, even though this need may not be as widely recognized as their need for interventions addressing HIV and other STIs [57]. These findings suggest that the public may not be sufficiently knowledgeable about the inequities facing key populations and that if they were, they might be more supportive of tailored and targeted interventions for high-need populations.

The present study has limitations. First, this study leveraged data from a web-based survey advertised on social media platforms resulting in the under-representation of several key populations. While iterative proportional fitting was used to address this bias, it is impossible to know the extent to which hidden factors may have influenced recruitment of participants. Second, our discrete choice experiment was implemented without design restrictions – meaning that some of the combinations of life expectancy gains, target populations, and program areas are unlikely to be advanced or achieved. Our choice to keep all intersections was informed both by our technical capacity in implementing the study via the Qualtrics platform, as well as our desire to understand the full spectrum of effects. For example, comparing less commonly implemented interventions (e.g., Cardiovascular disease interventions for gay and bisexual men) helps us to understand when and when not the public might be supportive of an intervention (e.g., HIV/STBBI interventions for gay and bisexual men). Third, the present study does not allow us to understand how and why participants made the choices they did. Additional qualitative research or more refined quantitative methods are needed to explore the ways in which individuals think about the choices they made. This work is important to confirm our argument that participant’s preferences for some population-specific programming arises from the populations awareness or perceptions of needs for these marginalized groups. We acknowledge, however, that stigma and other factors could also play a role in shaping preferences. Fourth, we note that we did not conduct a formal power analyses for this study. Rather the number of participants recruited was based on budgetary restraints. Information on power calculation for Discrete Choice Experiments is available from Bekker-Grob (2015) [58]. In the present study, each intersection of target population and condition was presented approximately 330 times. Further work is needed to assess the relative contributions of these various factors.

As noted above, preferences for resource allocations are likely informed by participant’s characteristics. While we have controlled for participant demographic characteristics in the present analyses, the effect that participant identity on participant preferences was not explored. These analyses are feasible in the present dataset. However, we feel that a more nuanced, dedicated, and focused introduction and discussion are needed to contextualize the role of identity in shaping stated preferences as well as an exploration of potential mechanisms that might influence whether an individual is likely to support an intervention tailored for marginalized peoples. Such work will be explored in future analyses.