Overview of the Causal Treatment Personality
The Causal Treatment personality of the Fit Model platform enables you to use observational data to estimate functions of potential outcomes. Treatment and outcome models are fit using generalized linear mixed modeling (GLMM). The models can accommodate binary or continuous treatment types.
Binary Treatment
This personality provides the following techniques for estimating mean potential outcomes for a binary treatment:
• Inverse probability weighting with ratio adjustment (IPWR)
• Regression adjustment (REGADJ)
• Augmented inverse probability weighting (AIPW)
You can specify a treatment model, a response model, or both. The treatment model estimates the probability of receiving one level of treatment versus the other. The response model estimates the expected value of the outcome. The average treatment effect (ATE) is estimated using IPWR when you specify only a treatment model. The ATE is estimated using REGADJ when you specify only a response model. When both models are specified, the ATE is estimated using AIPW. This method is doubly robust, which means that as long as one of the models is correctly specified, the AIPW is a consistent estimator of the ATE.
Options for truncated weights and propensity score matching are also available when the treatment model is specified. The option to use stabilized weights is available when only the treatment model is specified and the estimation method is IPWR.
Continuous Treatment
This personality provides the following techniques for estimating mean potential outcomes for a continuous treatment and a continuous outcome:
• Marginal structural modeling (MSM) with IPWR (Hernán and Robins 2020)
• Generalized propensity score (GPS) modeling (Hirano and Imbens 2005)
Both methods are used to calculate potential outcome estimates at specified treatment levels. Specification of the treatment model is required, but specification of the response model is not supported when the treatment is continuous. The outcome of interest must be continuous when the treatment is continuous.
Report Components
The Causal Treatment personality reports include the following components:
• Potential outcome estimates in the whole population and within the treatment group
• The difference of potential outcome estimates in the whole population (ATE) and within the treatment group (the average treatment effect on the treated)
• Model fit results for the specified treatment and response models
• Visualizations to check the veracity of necessary causal assumptions