When to be vaccinated? What to consider? Modelling decision-making and time preference for COVID-19 vaccine through a conjoint experiment approach

Modelling vaccine preference: Average marginal component effect

The statistical analysis is divided into two sections. In the first section we follow the approach proposed by Hainmueller et al. (2014) to estimate the average marginal component effects (AMCEs) of different attribute levels using their regression-based estimator. This can be achieved by running a linear regression of the choice outcome as the dependent variable on the sets of dummy variables for the attribute values. AMCEs can be interpreted as the effect sizes, which convey the expected change in the likelihood of selecting a particular profile when a given attribute level is compared with a chosen baseline attribute level.

The AMCE plot (Figure 1) yields the following results. Vaccine efficacy produces the largest spread of effect sizes. Vaccines with a 50% efficacy and those with a 70% efficacy are 30.7% and 16.9% less likely to be chosen than the baseline vaccine with a 95% efficacy. With respect to the probability of developing severe side-effects, a probability of 0.01%, 0.001% and 0.001% will make respondents 15.8%, 9.4% and 3.7% less likely to choose a vaccine compared with the baseline case of a 0.00001% probability. An unknown probability, meanwhile, will make respondents 22.6% less likely to choose that vaccine than the baseline. As for mild side-effects, a linear relationship similar to that of severe side-effects holds. The lower the likelihood of developing mild side-effects, the more preferable a vaccine will become.

• Download figure
• Open in new tab

Figure 1.

AMCE plot.

Note: The dots indicate AMCE estimates, while the lines show the 95% confidence intervals for the AMCEs. The dots without confidence intervals denote the baseline categories. For instance, the first line from the top shows that vaccines that come with a subsidy of HKD700 are on average 8% less likely to be chosen than a vaccine that is free of charge, which is used as a baseline category. The coefficients are displayed in Table S2.

As for pricing, the results indicate that making people pay for vaccines will reduce their likelihood of choosing them. A vaccine that costs HKD700 and HKD400 respectively are 1.1% and 1.8% less likely to be chosen than a free vaccine. Interestingly, subsidizing vaccines also reduces the desirability of vaccines. A vaccine that comes with a subsidy of HKD400 and HKD700 are 8.4% and 7.3% less likely to be chosen than a free vaccine. This means that the more subsidy the government provides, the less desirable a vaccine becomes. One explanation is that subsidies may generate a negative signaling effect that guides people to think about the negative consequences of receiving the vaccine. Overall, this suggests that a free vaccine is most preferred by respondents.

On the other hand, providing medical insurance against severe side-effects will make a vaccine 10.9% more likely to be chosen than providing no such insurance. Interestingly, providing an extra living subsidy does not make vaccines more desirable than simply by providing medical insurance. Yet, it also will not make it markedly less desirable. Finally, there is no consistent and significant effect regarding queuing time after registering for vaccination. A wait of 7 and 14 days respectively is neither increasing nor decreasing the desirability of a vaccine as compared with no waiting time. However, a wait of 30 days will make a vaccine 2.0% less likely to be chosen than the baseline. Interestingly, if the wait prolongs to 60 days, the negative effect becomes negligible again.

Time preference analysis

The second section focuses on vaccination time preference. While the conjoint experiment allows us to evaluate the relative importance of different vaccine attributes when making vaccination decisions, it does not tell us about the time at which respondents prefer to be vaccinated. The inclusion of the time preference question in each choice task provides an opportunity to examine this dimension. Our assumption is that people with different demographic backgrounds, government trust level and risk perception against COVID-19 would result in different vaccination time preference, based on earlier work on the determinants of vaccine uptake intention (Lazarus et al., 2020; Malik et al., 2020; Sherman et al., 2020; Paul et al., 2021). The summary descriptive statistics for these respondent-level characteristics are provided in Table S1.

We examine how vaccine attributes and respondent characteristics determine when they prefer to be vaccinated for the choice that they have made. Given that the time preference is a type of count data overdispersed with 10 (last 10% to be vaccinated) as the dominant time preference choice, we can model the data with zero-inflated negative binomial regression, which combines a negative binomial distribution and a logit distribution. Here, time preference, which is first recorded in reverse order, is specified as the dependent variable of the model. We include both the vaccine attributes in the conjoint experiment and a number of respondent-level characteristics as the independent variables in the negative binomial model. Meanwhile, we select two variables – 1) intention to vaccinate and 2) confidence in the vaccine – in the logit part of the model to predict the excess zeros. These variables are chosen because it is likely that respondents who do not want to be vaccinated and have low confidence in the vaccine will delay their vaccine preference as much as possible. We build the models hierarchically by adding blocks of variables on top of the others.

Figure 2 shows a histogram that shows the distribution of the time preferences for vaccination, which ranges from 1 to 10 (1 = first 10% to be vaccinated, 10 = last 10% to be vaccinated). Here time preference is also subsetted by respondents’ general intention to join the vaccination program. Majority (40.6%) of the choices are 10. The remaining responses are rather uniformly distributed along 1 to 9, but with a cluster toward the middle.

• Download figure
• Open in new tab

Figure 2.

Time preference for vaccination.

The regression models are presented in Table 2. To interpret the results, one has to apply the natural exponential function to the estimates. For instance, the estimate for being male is 0.164 in the full model and the exponential function of that is 1.178. This means the average preferred vaccination timing for male respondents is 17.8% later than that for female respondents in terms of the population quantile.

Both of our predictors for the zero-inflation logit model – 1) intention to vaccinate and 2) confidence in the vaccine – are both statistically significant. Their negative estimates suggest that those who do not have the intention to be vaccinated and those who have lower confidence in the vaccine are more likely to be the “excess zeros” in the model, i.e. those who indicated 10 for time preference. Second, for the count model, our quickest observation is that vaccine attributes do not matter as much as respondent characteristics in determining vaccination time preference. Among the attributes, only severe side-effects (two of its levels, as compared to the smallest probability) and the provision of medical insurance (vs. no insurance provided) are statistically significant. However, the impact is not consistent across levels.

In comparison, respondent characteristics seem to matter more. To begin, all the demographic variables are statistically significant. Male, older, less educated, and higher income respondents prefer to be vaccinated earlier. Respondents who work in high risk occupations, including medical workers, service workers, janitors and frequent travellers, also prefer to be vaccinated earlier. Meanwhile, respondents who have higher stress levels, measured by the composite index of PHQ-4, are more likely to prefer being vaccinated earlier. However, there is no statistically significant relationship between having chronic illnesses, and residing with vulnerable persons (infants, chronic patients, pregnant women and elderly), and preferring to vaccinate earlier or later.

Expectedly, intention to vaccinate and confidence in the vaccines are both statistically significant predictors in the count model. Intention to be vaccinated, and higher confidence in the vaccines, contribute to earlier time preferences. Meanwhile, risk perceptions generally do not have statistically significant relationships with time preferences. Respondents who perceive the COVID-19 pandemic to be more serious, and those who think they have greater risk of being infected, are not more likely to want to be vaccinated earlier. However, respondents who think they are taking sufficient preventive measures are more likely to delay vaccination. Financial impact of the pandemic has no statistically significant relationship with vaccination time preference.

Our models also seek to test how various types of trust correlate with vaccination time preference. Similar to previous studies that trust in government increases vaccination intention (Lazarus et al., 2020), we also find that it has a statistically significant relationship with time preference. Respondents who have higher trust in government are more likely to prefer to be vaccinated earlier. Meanwhile, there is no statistical relationship between trust in medical workers and vaccination time preference. Important to note is that trust in friends has a negative statistical relationship with time preference. Respondents who trust their friends more tend to delay their vaccination.

Finally, whether a respondent has been COVID-tested is statistically related to vaccination time preference. Generally, respondents who have been recently tested are more likely to prefer later vaccination (Model 7). This makes sense because receiving a negative test result may give people a sense of protection, which lowers their guard against infection. At the same time, we also find that respondents who live in neighbourhoods where there are recently confirmed cases are more likely to prefer later vaccination. To understand this counterintuitive finding, we include two interaction terms – first between testing and intention to be vaccinated, and second between testing and the presence of confirmed cases nearby – into Model 8. This is to take into account the fact that people get tested for different reasons. Some may have gotten the test voluntarily; others may have been “forced” to take the test because of compulsion or lockdowns. We believe that different motivations for testing will shape time preference differently. The first interaction term is a proxy for whether testing is voluntary (if one wants to be vaccinated and has been tested, it is likely that the test is taken voluntarily). The second interaction term is a proxy for whether testing is involuntary (if one has been tested and there are confirmed cases nearby, it is likely that the test is involuntary).

The result shows that both interaction terms are statistically significant. Respondents who have been tested and have the intention to be vaccinated are more likely to prefer to be vaccinated earlier than those who have been tested with no such intention. In contrast, respondents who have been tested and are close to confirmed cases are more likely to be vaccinated later than those who have been tested but with no nearby confirmed cases. Meanwhile, after adding the second interaction term, the statistically significant effect for confirmed cases becomes insignificant, showing that whether there are confirmed cases around on its own will not affect time preference.

Importance of efficacy and severe side-effect

With several COVID-19 vaccines being made available, public willingness to be vaccinated has become the next hurdle in fighting the pandemic. The results of our conjoint analysis showed that Hong Kong citizens are most concerned with vaccine efficacy and the likelihood of causing severe side effects when choosing COVID-19 vaccines. Although the importance of safety and efficacy have been identified previously in study in the US (Kreps et al., 2020; Motta, 2021), UK(McPhedran & Toombs, 2021) and China (Wang et al., 2020), our findings further contribute to these studies by showing that vaccines with 50% or 70% efficacy are 30.7% and 16.9% less likely to be chosen in comparison to a baseline vaccine with 95% efficacy, assuming all other attributes are comparable. This hypothetical situation means that if a country/region introduces vaccines with a great diversity of efficacy in the same mass vaccination program, citizens will be heavily inclined to choose the one with higher efficacy, potentially causing wastage of vaccines with a relatively lower efficacy.

We also found that the likelihood of causing severe side effects is a powerful contributor to preference for a vaccine. This result echoes previous research that indicates a more prominent role of adverse side-effect than mild side-effect in affecting public preference in other vaccines (de Bekker-Grob et al., 2010; Veldwijk et al., 2014; Guo et al., 2017). However, our results indicate that a vaccine with unknown probability of causing adverse side-effects is 8.1% less likely to be chosen than a vaccine with 0.01% chance of causing adverse effect. The finding suggests that information about the vaccines, especially their side-effects, should be made as transparent as possible, because uncertainty and low trust were prominent factors against timely vaccine uptake. Governments should enhance transparency of their vaccination program through effective communication.

Value of medical insurance and free vaccine

In countries with well-established social welfare systems, COVID-19 vaccination programs may be accompanied by subsidy, indemnity and extra living support. For example, Japan and Hong Kong have enacted laws to make COVID-19 vaccines free to residents. The United Kingdom has set up vaccine damage payment to those injured by the COVID-19 vaccine. Singapore has also agreed to provide support to those having serious side effects after the jabs through applicable healthcare plans.

However, our study suggests that subsidy is a double-edged sword: people want their vaccine to be subsidized, but they will develop hesitancy if too much subsidy is offered. Results from the conjoint analysis indicate that subsidized vaccines are more preferable to a vaccine that costs money. However, respondents are more likely to choose a free vaccine over a subsidized vaccine. One explanation is that subsidies may generate a negative signaling effect that guides people to think about the negative consequences of receiving the vaccine. As the valuation of vaccine changes, people who place a higher value on reducing risk generated by vaccination will be less likely to receive vaccination (Cook et al., 2009). Furthermore, having reimbursement from vaccination is not common in other vaccination programs. The deviation from established practice may invoke unnecessary risk perception. This explanation may be applied to indemnity programs and other healthcare plans related to COVID-19 vaccination.

Vaccination intention

Although up to 61% of respondents indicated a lack of intention to join the vaccine program, our findings show that their decision may change if they are presented with their preferred vaccine. Figure 2 suggests that while vaccination intention appears to correlate positively with time preference, there are more subtleties in the relationship. On the one hand, respondents who have the intention to be vaccinated may delay their time preference given a specific vaccine. On the other hand, although the majority of those who indicate no vaccination intention prefer to be vaccinated last regardless of the choice being presented, a significant proportion has chosen earlier time preferences. This implies that respondents who indicate no intention in being vaccinated may simply delay their timing. One possibility is that they are adopting a wait-and-see attitude to see whether the vaccines are safe enough. Another possibility is that rational people may want to avoid the potential cost of vaccination but benefit from herd immunity as public goods (Bauch et al., 2003; Bauch & Earn, 2004).

Ensuring the success of vaccination programs

Effectiveness of a vaccination program can be jeopardized when a significant number of citizens are reluctant to receive the jabs early on. Vaccination programs should be planned meticulously to ensure a head start and avoid contention that may diminish public vaccination intention. For instance, identifying volunteers to join the early phase of vaccination programs can help ensure their smooth rollout and wider acceptance. Our time preference regression model suggests that demographic attributes are crucial predictors in vaccination time preference. Male, older, less educated, and higher income groups have higher tendency to be vaccinated earlier. Those who work in high-risk environments are also more likely to prefer receiving the jabs earlier than others. High stress level, as measured by PHQ-4 in this study, is another predictor to willingness to receive early vaccination. Vaccination programs should target citizens with these backgrounds as the priority in the early promotion of vaccination campaigns.

Impact of trust

From the results, people who trust the government more tend to receive vaccination earlier, while people who trust their friends tend to delay their vaccination time and trust in medical workers have no relationship with vaccination time preference. Similar research has shown that individuals’ distrust or low trust against government is a significant predictor of vaccination hesitancy in other diseases (Quinn et al., 2013; Larson et al., 2018; Jamison et al., 2019). However, it remains largely unknown to us why individuals with higher trust in friends develop delay-vaccination sentiments. Extant literature shows that strong ‘bonding social capital’ (e.g., trust among friends) is associated positively with increasing vaccination intention and protective behaviour, which help create ‘public goods’ (e.g., herd immunity) under pandemic (Nagaoka et al., 2012; Rönnerstrand, 2013; Chuang et al., 2015). However, our results do not align with this finding. One possible explanation is that the distrust in government in Hong Kong undermines the benefit of state-led vaccination programs, leading to vaccination hesitancy and strong reliance on civil-society-led non-pharmaceutical interventions (Yuen et al., 2021). Another possibility is that people having better friendship networks may have some sorts of safety net to weather the uncertainties posed by the pandemic.

Impact of forced COVID testing

Model 7 and 8 of our time preference analysis suggest that forced testing around the outbreak area may delay vaccination time preference, possibly owing to the fact that people may feel ‘safer’ after being tested and thus reduce the urgency to take the vaccine. Many governments are relying on forced testing and quarantine to contain the risk of an outbreak. However, previous studies have identified that the majority of COVID-19 infection is attributable to presymptomatic and asymptomatic infection, making forced testing inadequate to interrupt a COVID-19 outbreak unless contact tracing is made rapid and thorough (Moghadas et al., 2020). In addition, stringent public health policy may not entail policy compliance with COVID-19 related measures (Yue et al., 2021). Thus, the stringency, scale and timing of forced testing should be scaled up or down, depending on the severity of the outbreak and planning of the vaccine program.