Hofman2021IntegrativeModeling

Jake M.Hofman, Duncan J. Watts, Susan Athey et al. "Integrating explanation and prediction in computational social science"

Bibliographic info

Hofman, J.M., Watts, D.J., Athey, S. et al. Integrating explanation and prediction in computational social science. Nature 595, 181–188 (2021). https://doi-org.proxy.library.uu.nl/10.1038/s41586-021-03659-0

Commentary

The most interesting part of this article is the integration of two opposite modelling styles. On the one side there is explanatory modelling, which is a modeling style commonly used by social sciences that prioritize interpretatively satisfying explanations. The social sciences are often criticized for failing to generalize, failing to predict the outcome of interest and failing to offer solutions for real world problems. On the other side is the computer science field, which prioritizes developing predictive models, but their weakness is that they do not create models that correspond to causal mechanisms or are interpretable. The main contribution of this paper is the integration of these two sides under the name 'integrative modelling'. This type of modelling uses the predictive power of explanatory models to prioritize the causal effects and quantify how much they actually explain. These explanations can then help to make more robust models that generalize better with changes. In general, integrative modelling focuses on making predictions on as-yet unseen outcomes in terms of causal relationships.

Excerpts & Key Quotes

Benefits: explanatory approach & predictive approach

Comment:

This quote nicely summarizes the benefits of both approaches. The computer science approach gets us closer to the real-world environment, as the predictive power can take in a lot of 'real' (training) data for the study of the phenomenon. On the other hand, the explanatory approach from social sciences can zoom out of the training data and take a more general view of the phenomenon under study, thereby contributing to more robust models of the phenomenon.

Relationship explanation & prediction

Comment:

The difference can be found in the balance between explanability and predictive powers, whereby the explanatory side logically prioritizes explainability
and the computer science side prioritizes the predictive power of the model. To be more concrete: the goal of explanatory modelling is to identify and estimate causal effects, but does not focus on predicting outcomes whereas the goal of predictive modelling is to predict the outcome of interest but not explicitly explain or identify the causal effects beneath it.

(Replace this heading text for this passage)

Comment:

The authors of this paper propose to integrate two sides into one, from explanatory modelling and predictive modelling to integrative modelling. However, this quote contains the word 'ideally', which is a good indicator of why there is a meager application of integrative modelling so far: often, it is infeasible to apply this approach in practice. For instance, if certain phenomena can only be explained by extremely complex machine learning models, the explanatory power is limited. Nonetheless, the stated definitions of different modelling approaches by this research can result in more transparent and open research once the field adds these approach-labels to their research, thereby indicating what goals they follow.