Socio-demographic correlates of wildlife consumption during early stages of the COVID-19 pandemic

We focused our research on countries/territories in Asia (specifically, Hong Kong SAR, Japan, Myanmar, Thailand and Vietnam) because COVID-19 had not spread much outside Asia at the time of data collection and the global effects were predominantly concentrated in East and Southeast Asia. Our five survey countries/territories were chosen because they all have relatively high levels of wildlife trade but also represent very different forms of trade (for example, the pet trade in Japan versus the wild-meat trade in Vietnam). Surveying respondents from markets with these different forms of trade thus allowed an examination of how the full variety of wildlife consumption types may be impacted by perceived disease risk. Budgetary constraints precluded the inclusion of further countries, although we believe those that were surveyed provide a valid snapshot of the main regional issues and patterns. The exception to this may be the exclusion of China, a key global player in the wildlife trade and the possible origin of the COVID-19 virus. Conducting research in China requires an extensive process to obtain permission that was not consistent with the opportunistic nature of our survey, which was mobilized quickly to target opinions from a snapshot view of an (at that time) emerging disease. Given the time-sensitive nature of the research, we were therefore unable to wait for the necessary permissions to include China in this survey.

Our online survey was conducted between March 3–11, 2020 and surveyed 1,000 respondents in each of the five target countries/territories. We designed and translated our questionnaires with local experts to ensure questions were culturally appropriate, understandable and relevant. The survey was a quantitative data collection instrument that comprised 32 questions, lasted on average 8 minutes, and respondents were offered an incentive for participating. Respondents aged 18+ were invited via email from an online panel of over 2.5 million people in the target countries/territories, and could answer on any internet-capable device (for example smartphone, tablet, laptop) at their convenience. Only respondents aged 18 and over were eligible to take the survey, which was entirely voluntary. Any respondents working in advertising, public relations, marketing, market research or media industries were screened out to prevent possible bias. The email invite that was sent to participants did not specify the exact nature of the survey to avoid skewing the participants towards those that believed they know about the topic. Instead, the invite indicated that the questions would be about ‘consumption and shopping habits’. The panel is maintained by Toluna (https://tolunacorporate.com/), an online data collection group focused on providing high-quality market research data to clients in various business and non-business sectors. Toluna builds and maintains large online consumer panels to collect these data while adhering to stringent global and local guidelines for panel management and data quality, and is a member of the European Society for Opinion and Market Research (https://www.esomar.org).

Toluna respects privacy and is committed to protecting personal data. Their privacy policy (https://tolunacorporate.com/legal/privacy-policy/) provides information on how Toluna collects and processes personal data, explains privacy rights and gives an overview of applicable legislation protecting the handling of personal information. Toluna only uses personal data when the law allows the data to be used.

Respondents were asked demographic questions, and quotas based on the most recent census data for each country/territory were used to ensure the final sample profile was nationally representative of age and gender, except in Myanmar where internet access skewed online panel members to a younger male demographic. Specifically, participants were excluded once quotas on age and gender were filled, and again, participants working in advertising/public relations, marketing research or media were excluded from the survey as we believed these jobs could influence responses. Respondents were asked about societal, economic and environmental concerns, their perception of COVID-19 and their attitudes towards wildlife and wildlife consumption (Supplementary Methods). We also excluded respondents who stated that they were unsure whether they or anyone in their social circle had recently purchased wildlife products (n = 421), as well as an additional n = 39 respondents who were unable to answer survey questions that were later included as covariates in our models.

Because of the potentially sensitive nature of wildlife consumption, we asked about past wildlife purchases indirectly, questioning respondents on whether anyone within their social circle, including themselves, had recently purchased wildlife products. Indirect questions can improve answer rates for questions that people may feel uncomfortable about answering honestly²⁷. During the pandemic, respondents may have felt uncomfortable about revealing wildlife purchases, given links between wildlife consumption and COVID-19. Additionally, although most wildlife consumption is legal (with restrictions) in the markets surveyed, some is not, and researchers can be perceived as having interests contrary to that of the respondent. For less-sensitive questions on future wildlife consumption and changes in consumption resulting from COVID-19, we asked respondents for their own response rather than that of their social group.

Previous studies have found a high correlation between an individual’s admission of using a wildlife product and their likelihood of being within a network of individuals who buy such products²⁸, and suggested that this is linked to homophily in social networks, especially in Southeast Asia. The homophily principle states that people’s personal networks are homogeneous with regard to many socio-demographic, behavioural and intrapersonal characteristics²⁹. Research on wildlife consumption in other Southeast Asian contexts suggests that social groups can be a motivator to begin or maintain consumption of wildlife products^28,30. Our own previous research supports this, indicating a strong correlation between one’s own tiger and ivory purchases and knowing someone within one’s social circle who has purchased such products. Additionally and recognizing the homophily principle, behaviour change campaigns targeted at social networks rather than individuals per se are likely to achieve better results than non-targeted campaigns. Changing perceptions of acceptability is a key aspect of social marketing and is used in the social mobilization domain of social and behaviour change communications, which has become a popular framework for reducing demand for illegally traded wildlife products³¹. Influencing people within a wildlife consumer’s social network may therefore have a higher rate of efficacy than attempting to influence the perceptions of individuals who do not know any consumers of wildlife.

We used hierarchical Bayesian regression models to assess relationships between socio-demographic explanators and our three response variables: (1) self-reported recent wildlife consumption, (2) change in wildlife consumption as a result of COVID-19 and (3) anticipated future wildlife consumption. Explanatory variables included 22 non-collinear variables in six categories: basic demographics, awareness and level of worry of COVID-19, COVID-19 personal impacts, support for and effectiveness of wildlife market closures, international travel habits and general attitudes towards global issues (Supplementary Table 1). Aside from household income (measured in US dollars per year), age (midpoint of year categories from the survey question) and education (ordinal, reflecting increasing level of schooling), all other variables were categorical; those with more than two categories were collapsed into dummy variables. Income, age and education were standardized and included to investigate whether a person’s general socio-economic status affects wildlife consumption. General attitudes towards global issues were expected to reflect aspects of respondents’ political tendencies, while travel habits were included to test the hypothesis that those who travel internationally more habitually are, and will be, more frequent consumers of wildlife. Questions regarding awareness and impacts of COVID-19, and concern about future disease epidemics, were asked to determine how the pandemic may be shaping wildlife consumption. Finally, support and perceived effectiveness of wildlife market closures were included as predictor variables since this measure has been suggested as a strong policy lever to reduce wildlife consumption.

The general structure of all three models was as follows:

This model allowed both coefficients and intercepts to vary across countries (that is, a ‘random-slope random-intercept’ model). In equation (1), y_ij is whether or not individual i in country j reported wildlife consumption, modelled as a Bernoulli trial with probability θ_ij. The logit transformation of θ (equation 2) is a linear function of parameters α and u₁ (the fixed intercept term and a vector of the country-specific intercept terms, respectively), as well as a vector of fixed regression coefficients β and a vector of country-specific regression coefficients u₂, with X and Z being the corresponding design matrices³². For α and β, we used an improper flat prior over the real numbers, while the group level parameters u₁ and u₂ were assumed to arise from a multivariate normal distribution with mean 0 and unknown covariance matrix. The covariance matrix was parameterized by a correlation matrix having a Lewandowski–Kurowicka–Joe prior, and a standard deviation with half-Student t prior with three degrees of freedom³².

For the three dependent variables, we evaluated the predictive power of a model containing all 22 variables, as well as six subset models, using Watanabe–Akaike Information Criterion and leave-one-out cross-validation³³. Each of these six subset models contained all explanatory variables except for those within one of the six categories described above (for example, all explanatory variables except those relating to international travel habits, all explanatory variables except those relating to support for wildlife market closures). We used this model-comparison approach to test whether any of these categories of explanatory variable were more or less important in explaining wildlife consumption; if particular categories of variable are stronger predictors of wildlife consumption, this could help inform where future conservation interventions should focus on. Watanabe–Akaike Information Criterion and leave-one-out cross-validation are both measures of model predictive accuracy (both use log predictive density as the utility function or comparison metric) and have been suggested as useful metrics for Bayesian model selection³³. We interpreted variable coefficients whose 95% Bayesian credible intervals did not contain 0 as providing strong evidence for the impact of that variable on the outcome in each of the three models for self-reported wildlife consumption (that is, recent, future and changes due to COVID-19). Models were estimated using the R statistical computing software³⁴, in particular the package brms³², with four chains of 1,000 iterations each, a 500-iteration warm-up period, and with successful convergence verified by confirming that R-hat statistical values were less than or equal to 1.01 (ref. ²²).

We used the Bayesian hierarchical model of anticipated future wildlife consumption and generated predicted probabilities of future consumption for our sample population (Fig. 2, grey bars). We then predicted future consumption probabilities for a hypothetical behaviour-change intervention (Fig. 2, coloured bars). This intervention was simulated by setting the ‘medical impact’ variable to zero for all individuals, and by assigning all individuals into the ‘aware lots’ and ‘support very likely’ categories for questions related to level of awareness of COVID-19 and level of support for government closure of domestic wildlife markets, respectively. All other variables for individuals were held at the levels recorded in the surveys. We considered the difference between these two predicted probabilities as the impact of the hypothetical behaviour-change intervention, which we examined at the level of the country/territory and within education, age, income and gender demographic classes. Strong evidence for the effectiveness of this hypothetical intervention among countries and demographic classes was suggested where Bayesian credible intervals around the mean predicted difference were less than zero (Supplementary Table 3).

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Source: Ecology - nature.com