SMEP: Causality and Treatment Effects
Status: work in progress and planning
This refers to methods for data where observations refer to individuals or individual events, e.g. success of and individual participating in a treatment, and is different from causality in econometrics with aggregate data as for example in time series and system of simultaneous equations. Either an individual chooses A or B, either an individual participates or not.
The main concept is potential outcomes for unobserved alternatives or events.
<josef> I don't have a real overview yet. (Wooldridge, Angrist, Pischke, Imbens on econometrics side)
exogeneity
- no unobserved confounders, conditional independence
- unobserved, endogenous variables and instrumental variables
parametric
- fully parametric: heckman model, Greene articles
- semi-parametric identification and efficiency bounds, some averages can be estimated without parametric assumptions
population
What is the population for which we want to calculate an effect. (e.g. randomized sample is randomized from some population, stratification, survey sampling, effect on the treated, ...)
see also list in Agrist and Pischke (which I haven't read), and survey articles or chapters
- treatment effects
- inverse probability weighted and double robust estimators (e.g. teffects in Stata)
- propensity score matching and similar matching methods
- regression discontinuity design (no idea yet)
- difference in differences (is there anything to support in a stats package?)
- mediation analysis:
- parametric, nonlinear
- semiparametric ???
- instrumental variables and control variates
- fully parametric and fully specified system of equations (I guess Greene's survey articles are in here. I guess mixed models also belong here.)
graphical models (extra, tool)
The main characteristic that is not clear to me is in the relationship of using a parametrically specified model that is robust to misspecification, similar to GLM/LEF for the simpler case. If we use a parametric model to perform the calculations but the effect is semi-parameterically identified, then we might be robust to misspecification, or maybe not. (I haven't read Wooldridge often and carefully enough yet to understand this.)
i.e. What's all the fuss about the "modern" econometrics?
PRs, location, design ???