Modelling
This session reviews modelling and analysis and how it impacts the results of studies.
Underline the impact of various modelling choices on the credibility of our results and discuss some modelling challenges we often face in applied work.
Summary
When building regression models, we make many choices: which variables to include in the model, which functional form, etc. These choices are most often not straightforward and, in this lecture, we discuss some of the challenges associated to the construction of our regression models. We first discuss the causal graph framework and Directed Acyclic Graphs (DAGs) in more details. We then review standard modelling assumptions and discuss what happens when they fail and further discuss a few aspects associated with the choice of controls. When the outcome is limited, eg binary, categorical or censored, usual linear models cannot properly represent the DGP and we need to turn limited outcome models. We discuss some of these models, emphasizing why they are helpful and how to implement and interpret them.
Session Outline
- Introduction: goals and definitions
- Causal Graph Framework
- Failures of Gauss-Markov conditions
- Controls
- Limited outcome models
- LPM, probit, logit
- Multinomial logit
- Poisson and negative binomial models
- Summaries
Materials
Specific resources for this lecture
DAGs
- Chapter 6 to 9 of Huntington-Klein (2022)
- Chapter 2 of Cunningham (2021)
Modelling
- A large share of the slides are derived from notes by Claire Palandri and related slides by Anna Papp
- Limited outcome models:
- Quick overview in Section 13.2.6 of Huntington-Klein (2022)
- Chapters 13 to 15 from Gelman, Hill, and Vehtari (2021)
- Bad controls: Chapter 3.2.3 of Angrist and Pischke (2009)