Partially Specified Causal Simulations
Authors:
A. Zamanian,
L. Mareis,
N. Ahmidi
Abstract:
Simulation studies play a key role in the validation of causal inference methods. The simulation results are reliable only if the study is designed according to the promised operational conditions of the method-in-test. Still, many causal inference literature tend to design over-restricted or misspecified studies. In this paper, we elaborate on the problem of improper simulation design for causal…
▽ More
Simulation studies play a key role in the validation of causal inference methods. The simulation results are reliable only if the study is designed according to the promised operational conditions of the method-in-test. Still, many causal inference literature tend to design over-restricted or misspecified studies. In this paper, we elaborate on the problem of improper simulation design for causal methods and compile a list of desiderata for an effective simulation framework. We then introduce partially randomized causal simulation (PARCS), a simulation framework that meets those desiderata. PARCS synthesizes data based on graphical causal models and a wide range of adjustable parameters. There is a legible mapping from usual causal assumptions to the parameters, thus, users can identify and specify the subset of related parameters and randomize the remaining ones to generate a range of complying data-generating processes for their causal method. The result is a more comprehensive and inclusive empirical investigation for causal claims. Using PARCS, we reproduce and extend the simulation studies of two well-known causal discovery and missing data analysis papers to emphasize the necessity of a proper simulation design. Our results show that those papers would have improved and extended the findings, had they used PARCS for simulation. The framework is implemented as a Python package, too. By discussing the comprehensiveness and transparency of PARCS, we encourage causal inference researchers to utilize it as a standard tool for future works.
△ Less
Submitted 5 October, 2023; v1 submitted 19 September, 2023;
originally announced September 2023.
