Econometrics Public Health

The long-term effectiveness and cost-effectiveness of public health interventions; how can we model behavior?

That is the title of an interesting review paper by Squires et al. 2023. The abstract is below:

The effectiveness and cost of a public health intervention is dependent on complex human behaviors, yet health economic models typically make simplified assumptions about behavior, based on little theory or evidence. This paper reviews existing methods across disciplines for incorporating behavior within simulation models, to explore what methods could be used within health economic models and to highlight areas for further research. This may lead to better-informed model predictions. The most promising methods identified which could be used to improve modeling of the causal pathways of behavior-change interventions include econometric analyses, structural equation models, data mining and agent-based modeling; the latter of which has the advantage of being able to incorporate the non-linear, dynamic influences on behavior, including social and spatial networks. Twenty-two studies were identified which quantify behavioral theories within simulation models. These studies highlight the importance of combining individual decision making and interactions with the environment and demonstrate the importance of social norms in determining behavior. However, there are many theoretical and practical limitations of quantifying behavioral theory. Further research is needed about the use of agent-based models for health economic modeling, and the potential use of behavior maintenance theories and data mining.

Here is more detail on the 5 primary types of models they analyze.

  • Econometric models: Models incorporate the relationship between price and consumption using regression analysis. The relationship is often measured as an elasticity. Applications of this approach include a model of alcohol consumption [Purshouse et al. (2014)] and obesity [Basu et al. (2014)]
  • Structural equation modeling (SEM). This approach relies on a set of methods that uses statistical models to assess the causal relationships between a set of unobservable (latent) and observable variables. SEM involves setting out the expected structural relationships between relevant variables using a path diagram, and then testing the relationships using statistical analyses. Bates (2021) considers various psychological mechanisms of action (i.e., dietary restraint, habit strength, autonomous diet self-regulation) on obesity in her SEM study.
  • Behavioral system dynamics modeling. “System dynamics models capture the ‘stocks’ (a quantity of a variable at a given point in time) and ‘flows’ (rates of change of the stocks), including positive and negative feedback loops within a system over time, to capture the behavior of the system…The relationships between the variables can be used to incorporate the fact that decisions within one part of the model will not be based upon full information of the entire system, thus incorporating bounded rationality.” These are generally cohort models. Luo et al (2018) uses a more complex differential equations model diffusion of food safety behavior.
  • Agent based modeling (ABM) and social network analysis. The authors claim that the key benefit of ABM is that individuals can interact with each other and with their environment and rational individual actions can lead to unexpected (but perhaps realistic) outcomes at the macro level. ABM have become more popular in recent years, including hundreds of infectious disease transmission models developed to predict interventions to fight the COVID-19 pandemic. Duan et al. 2015 provides a nice review of these models. 
  • Data mining. While data mining has been used for many years, large data sets and improved computing power have lead to the ability for more non-linear modelling such as neural networks and other forms of macine learning. However, the external validity of these exercises can be problematic if the data on which the mining occurs in not fully representative of the population of interest. Thus, combining data mining with theory may be a productive avenue for future research.

Behaviors modelled in the papers considered in the Squires et al. study included:

  1. Utility maximization
  2. Econometric analyses for modeling the relationship between price and consumption;
  3. Game theory
  4. Fixed behavioral patterns based on empirical data or schedules of the agent type;
  5. Behavior change is more likely, or occurs if some threshold is exceeded, depending on number of contacts, (perceived) behavior of contacts and/or distance to location, as well as other variables such as past experience and sociodemographic characteristics;
  6. Follow-the-average (i.e., , where behavior is adjusted to the average behavior of the social network or model population);
  7. Other heuristics (e.g., those based on price, distance, habits, preferences and/or neighbor behavior);
  8. Quantified behavioral theory;
  9. Using existing cognitive architectures which focus upon the inner workings of the brain.

The first 3 approaches are rooted in economics and assume people are rational, while the latter 6 approach use a bounded rationality approach.

The paper provides a nice review throughout and you can read the full study here.