The Econometric Model for Causal Policy Analysis
Abstract:
This presentation will focus on the econometric model of causal policy analysis and two alternative frameworks that are popular in statistics and computer science. By employing the alternative frameworks uncritically, economists ignore the substantial advantages of an econometric approach, and this results in less informative analyses of economic policy. I will show that the econometric approach to causality enables economists to characterize and analyze a wider range of policy problems than is allowed by alternative approaches.
Bio:
I am an Assistant Professor of Economics at UCLA and a research affiliate of the Human Capital and Economic Opportunity Global Working Group (HCEO) and the NBER. My research interests include policy evaluation, causality, and applied econometrics. A common theme in my work is the identification, estimation, and inference of causal effects. Recently, I have focused on using revealed preference analysis to improve causal inference in social experiments and on applying machine learning techniques to perform policy evaluations. I have examined various social experiments, including the Perry Preschool Intervention, the High/Scope Comparison Study, the Abecedarian Project, the Nurse-Family Partnership, the Jamaican Intervention, Programa Primeira Infância Melhor in Brazil, Oportunidades in Mexico, and Moving to Opportunity. Additionally, I have investigated observational data from several countries, including the US, Germany, and China. I have worked on approximately two dozen research articles, which have been cited over 9,400 times.
Summary:
Causality is studied differently in economics, statistics and computer science, which creates confusion
Fritsch perspective: causality is not something that can be proven about reality; it is a way of thinking/type of analysis
Haavelmo: causal effects are not empirical descriptions of real worlds but descriptions of hypothetical worlds (what if); obtained through models
Intuition of Ceteris Paribus (all else equal)
Have model of outcome as a function of inputs Y = f(a, b, c, x)
Fix some inputs inputs a,b, c to hypothetical values and vary remaining input x
Captures the causal impact of X on the output
Difference between conditional probability and counterfactya
Counterfactual: thought experiment about possible world
This is outside of statistical theory
Conditional probability: region of the event space
Identification of a statistical model is often conflated with causal estimation
Examples: diff-in-diff and granger causality are statistical, not causal
Causal Framework tasks:
Define Causal Models
Identifying Causal Parameters from known population distribution functions
Estimation of Parameters from Real Data
Causal Model:
Random Variables
Error Terms
Structural Equations
Causal Relationships: equations, directed acyclic model
Tools:
Local Markov Condition: each variable is independent of its non-descendants, conditional on its parents (all incoming causality is captured by parents variables)
Graphoid Axioms: independence conditions with respect to causal graph neighborhood
Fixing Operator: fix some variable to a constant
Causal exercise that hypothetically assigns a value to an input in the equation
Determines counterfactual outcomes
Defines the Average Causal Effect of X on Y when x is fixed at a value
Fixing ≠ Conditioning
Causal Inference: directional causality
Potential Outcomes:
Focus on independent relationships of counterfactuals
Emphasis on which variables need to be controlled for to estimate without bias
Needs to know these confounding variables
Foundation for matching, diff-in-diff, two-stage regression, etc.
Instrumental Variables:
Focuses on scenario where confounding variables are not known
Emphasis on the causal sources that are random and can be used to estimate without bias
Marginal Treatment Effect: key concept from which most other causal parameters can be inferred: ATS, TT, TUT, IV, OLS, LATE
Proposing approach to model do actions in Pearl’s do-calculus as modification of causal graph where the intervention variable is cut off from its causal parents
Separation captures the idea of “causal impact of just this variable”
New, sparser causal structure focuses statistical analysis on causal propagation emanating from the intervened-upon variable