Model Assumptions
All quantitative causal analyses involve two major groups of assumptions:
- Causal modeling assumptions
- Statistical modeling assumptions
Details about each of them can be found in specialized literature (Ferrari, forthcoming).
Although some checks for statistical modeling assumptions are
provided, the causalinf module puts emphasis on the first type of
assumptions, namely, the causal modeling assumptions. The main
reasons is that the initial step in any causal inference analysis is
selecting among various causal inference models. This selection largely
hinges on assessments of the data generating process and which causal
modeling assumptions are plausible in the specific context of
application. Typically, users of causal inference methods know the
causal assumptions associated with each causal modeling approach and
then select the method whose assumptions are plausible in the available
data. Hence, the emphasis of causalinf on those assumptions.
Users can find functionalities to test statistical modeling assumptions
in various other existing modules outside causalinf. These external
functionalities can be used directly on the output of the estimation
object from causalinf (see Estimation). However,
differently from many statistical modeling assumptions, the validity
of causal modeling assumptions cannot be tested using typical
statistical test procedures. Instead, the judgment about their
validity depends on three things:
- Methodological expertise, which refers to the understanding of
what the assumptions mean (and do not mean) in terms of the
relations between the variables used and omitted in the model.
The literature on causal inference methods provides guidance in this
regard. Ferrari (forthcoming) is the reference text used for
the
causalinfimplementation. - Domain knowledge expertise, which refers to the understanding of how the data was generated and whether the assumptions plausibly hold for the particular intended application.
- Heuristic or suggestive evidence based on data. For some assumptions of specific methods, procedures exist to evaluate whether the causal modeling assumptions seem plausible given the data.
Whenever possible, the causalinf module facilitates that third aspect
of the assessment of causal modeling assumptions. The main function in
all submodules to get that assessment is check_assumptions(). For
instance, to check the causal assumptions for a DiD application,
use:
<args> vary across submodules depending on the method used. See
Methods for details in each case.
References
- Ferrari, D. (forthcoming). The Identification of Causal Effects. Cambridge University Press.