The first registered infection with the COVID-19 virus occurred in China at the end of 2019. The virus spread all over the world, turning a single infection into a pandemic. Most countries implemented policies to limit the spread of the virus. In general, these policies can be divided into pharmaceutical and non-pharmaceutical interventions. Vaccines are pharmaceutical interventions. Examples of nonpharmaceutical interventions are temporary closure of restaurants and bars or a limit to the number of people that are allowed to gather for social events. Although, these non-pharmaceutical interventions were effective in slowing down the spread of the COVID-19 virus, they also severely affected people’s social and economic life. For example, freelancers who worked in restaurants and bars lost their income and students experienced a decrease in mental health due to the lack of social engagement. These societal impacts of COVID-19 policy gradually reduced the public support and adherence to these COVID-19 measures over the course of different pandemic waves. Measuring how people weigh the societal impacts of COVID-19 policy during these waves, provides interesting insights that can improve the effectiveness of pandemic policies for a future pandemic. A way to measure how people weigh the societal impacts of COVID-19 policy during different waves of the pandemic is by conducting Discrete Choice Experiments (DCE). These DCEs are often analyzed with Multinomial Logit (MNL) models. The MNL model is able to quantify the relative importance of different societal impacts for the population as a whole. However, there also exist other models, such as the Mixed Logit (ML) model and Latent Class (LC) model, and different types of DCEs, such as the labeled and unlabeled DCE, that have their own advantages and disadvantages in eliciting the relative importance of societal
impacts during different phases of the pandemic. For instance, ML models are able to elicit how preference heterogeneity is distributed among individuals and the LC model is able to divide people into different classes with their own preferences. Also, the labeled DCE is able to measure how people weigh societal impacts when the COVID-19 measures that cause these impacts are explicitly mentioned. Unlabeled DCEs do not explicitly mention these measures, but only weigh the societal impacts.
Over the course of the end of 2022 and the beginning of 2023, the pandemic transitioned into an endemic. Therefore, this study will not only evaluate the advantages and disadvantages of using these models and DCEs during the pandemic, but also during the endemic, by asking the following question: What are the (dis)-advantages of using ML and LC models over MNL models to analyze (un)-labeled DCEs that weigh societal impacts of COVID-19 policy during the pandemic and endemic? The eventual goal of this study is to find out where the added value lies of ML and LC models compared to MNL models and labeled compared to unlabeled DCEs in informing future pandemic policy makers during different pandemic waves and the endemic. To answer the main research question, this study formulated three subresearch questions. The first subresearch question is defined as follows: What are the differences in using ML, LC and MNL models to analyze (un)-labeled DCEs that weigh societal impacts of COVID-19 policy during the pandemic according to the literature? To answer this question, the study conducted a literature review on the results obtained from labeled and unlabeled DCEs with MNL, ML and LC models in different waves of the pandemic. In the second part of the study, the following subresearch question is addressed: What are the differences between the results produced by ML, LC and MNL models obtained from (un)-labeled DCEs that weigh societal impacts of COVID-19 policy during the endemic? This question is answered by conducting a labeled and unlabeled Discrete Choice Experiment during the endemic. The third and last sub research question asks: What are the (dis)-advantages of using ML, LC and MNL models to analyze (un)-labeled DCEs that weigh societal impacts of COVID-19 policy during the pandemic and endemic according to experts? To answer this question, three expert interviews were conducted.
The results of this study show that the main advantage of the ML model is its ability to test for the existence of preference heterogeneity among individuals in a sample. This can help to check the reliability of the estimates produced by the MNL model. The reason for this is that the existence of preference heterogeneity means that the estimates of the MNL model do not adequately represent the majority of the sample, if the heterogeneity is high. Additionally, the existence of preference heterogeneity is a reason for further research into the origins of this heterogeneity. In this case, the LC model is able to estimate different classes with different preferences and characteristics that represent different subgroups in society. Providing such a classification is the main advantage of the LC model. The main disadvantage of the ML model is that it is time consuming to estimate the model.
For the LC model the main disadvantage is its sensitivity to changes in covariates and initial values. With regards to the DCEs, the study shows that the main advantage of the labeled DCE is its ability to measure the effect of COVID-19 measures on how respondents weigh societal impacts. For the unlabeled DCE, the main advantage is its ability to measure how people view different societal impacts without explicitly mentioning the COVID-19 measures that caused these impacts. With regards to the added value of DCEs for informing future pandemic policy in different waves of a pandemic and in an endemic, the study shows that unlabeled DCEs are most suitable to give a baseline estimation of the preference for societal impacts at the beginning of a pandemic. So that, these insights can be taken into consideration when the first pandemic policy package is created. Further, the study shows that labeled DCEs should be applied during and in between pandemic waves to evaluate and adjust implemented pandemic policy. Finally, a labeled DCE can be implemented during the endemic, that follows the pandemic, to evaluate the impact of COVID-19 measures. The insights obtained from this can help to inform a future pandemic. With regards to the added value of the ML and LC model in informing future pandemic policy in different waves of the pandemic and the endemic, the study shows that the ML model can be used to test the reliability of the mean coefficients of important societal impact attributes. This helps to verify if a specific societal impact attribute is a good target for mitigation with generic pandemic policy during pandemic waves. In between pandemic waves, the LC model can be used to elicit the origin of preference heterogeneity among people in the sample for different societal impact attributes. These insights can be used to create customized pandemic policy with increased public support and adherence for when a next pandemic wave hits.
The most important limitation of this research is the lack of unlabeled DCEs from both the pandemic and endemic and the lack of labeled DCEs from the endemic that are included in the literature review. A drawback of the included DCEs is that these experiments do not include exactly the same societal impact attributes and that these experiments are conducted during different waves of the pandemic in different countries. These factors make it difficult to compare the results of the studies. Also, this thesis recommends that future studies conduct further research on the societal impacts of COVID-19 measures with models that are extensions of the models that are used in this study, such as the logit or probit model and the LC model with distributed preference coefficients. Furthermore, it would be valuable to analyze the societal impacts of COVID-19 measures with models outside the domain of choice modeling, such as data driven models. Furthermore, this study emphasizes that a lot of research can be done into how and by whom the insights of this research should be implemented in the pandemic policy decision making process. For instance, questions for future research could be, should the insights of this study function as advice for the creation of pandemic policy or as directive? And does the government or the National Institute of Public Health decide upon this?