Detecting tax evaders in Uganda : a comparison of logistic regression and finite mixture of logistic regression

Abstract

This paper presents data mining techniques to predict VAT evaders for inclusion in the audit plan. A dataset of 1,311 registered VAT taxpayers in Large Taxpayers’ Office and Medium Taxpayers’ Office for the period FY2017/18 across 11 features was used. An exploratory data analysis was used to establish the hidden patterns in the dataset and backward elimination method was used to identify the significant features for model development. Logistic regression (LR) and finite mixture logistic regression with and without concomitant variable were used to detect VAT evaders. BIC was used to select between the FMLR model with and without concomitant variable. Results of each technique were compared and the best technique was chosen based on accuracy, precision and recall were used to evaluate model performance. Findings of the study showed that number of no sales return, tax office, business sector and number of late payments were identified as significant features in VAT evasion detection. FMLR without concomitant variable had a lower BIC compared to FMLR with concomitant variable and was therefore considered. Model performance evaluation between FMLR without concomitant variable and LR was carried out and FMLR outperformed LR in accuracy, recall and precision. Though LR has been extensively used as a solution to tax evasion problems, the findings of the study suggest that FMLR provide better results compared to LR. The findings of the study can be utilized by URA with emphasis on the four (4) significant variables to detect VAT evaders for inclusion in the Audit plan. URA and future studies may employ other: evader attributes, data mining techniques and model performance evaluation metrics on similar dataset and compare the results.