Summary of Rubin 2007 SIM article:

Key idea: Conduct the design prior to seeing the outcome data in such a way that any future model-based adjustments tend to give similar point estimates.

Design: Contemplate the collection, organization, and analysis of the data prior to seeing any outcomes (e.g., Analyses of covariate data to create matched treated-control samples or subclasses with similar covariate distributions for treated and control subsamples, as well as specification of primary analysis for outcome data).

The Rubin Causal Model can be a checklist in the statistic plan.

Rubin Causal Model (RCM) outline:

Step 1: Units, treatments, potential outcome (conceptual)

Define the units, treatments, and potential outcomes before seeing any data. This forces conceptualizations of causal questions in terms of real or hypothetical manipulations.

Step 2: Assignment mechanism

Formulate the assignment mechanism via a probability model.

• Create a list of covariate variables that define "similar" treatment and control units. Use the list for matching.
  • Matching.Example (Use Rubin)
    • Estimate propensity scores on all of the data via a logistic regression predicting treatment/control status from covariates.
    • Assess diagnostic measures of propensity scores (distributional differences in propensity scores, percent of covariates with specified variance ratio after propensity score adjustment).
    • Create initial matched groups based on the propensity score.
    • Re-estimate propensity scores in the matched samples as if they were the original sample.
      • Reassess balance via diagnostics measures.
    • Further subclassify on the propensity score (e.g., subclassify into 2,4,6,8, and 10 subclasses).
      • Compare treated/control units via diagnostics in each of the subclasses and combine by weighting each comparison by the number of treated units in each subclass. The final collection of subgroups then approximates a randomized block experiment with respect to the observed covariates.
    • Proposed analyses would take place within each of the subclasses and would be combined via weighting by the number of treated units in each subclass (e.g., the analysis could be the comparison of means for treated/control units, or it could be based on such a comparison adjusting for all covariates.).