Abstract
The study focuses on comparative analysis of accuracy for various methods for evaluation of the average causal effect (ATE). A range of methods includes the t-test, linear regression, and several data matching techniques such as nearest neighbor, full matching, and subclassification. The methods are tested using the generated synthetic samples of varying dimensions from 250 to 1000, with different proportions of data divided into experimental and control groups (50/50, 70/20, and 90/10), under the assumption of normal, exponential, and logarithmic distribution of predictors. The Mean Absolute Percentage Error (МАРЕ) error metric is used to evaluate the accuracy of methods. The numerical experiments reveal that using t-test and linear regression techniques leads to a significant under- or overestimation of the ATE values, while the full matching methods produce more accurate results regardless of covariate probability distributions. Also, the balanced and larger samples provide the ATE evaluations with higher accuracy. The obtained results can be used to justify the proper method for evaluation of casual effects under the direct influence of the binary nature and nonlinearity of one of the affecting factors.