This paper discusses past and current methods the IRS uses to determine which individual income tax returns to audit. The IRS currently uses the discriminant function to give all individual tax returns two scores; one based on whether it should be audited or not and one based on if the return is likely to have unreported income. The discriminant function is determined by the IRS’s National Research Program, which takes a sample of returns and ensures their accuracy. Previously, the function was determined by the IRS’s Taxpayer Compliance Measurement Program. However, this was too burdensome and time consuming for taxpayers. The data mining techniques of decision trees, regression, and neural networks were researched to determine if the IRS should change its method. Unfortunately IRS tax data were not obtainable due to their confidentiality; therefore credit data from a German bank was used to compare discriminant analysis results to the three new methods. All of the methods were run to predict creditworthiness and were compared based on misclassification rates. The neural network had the best classification rate closely followed by regression, the decision tree, and then discriminant analysis. Since this comparison is not based on IRS tax data, no conclusion can be made whether the IRS should change its method or not, but because all methods had very close classification rates, it would be worthwhile for the IRS to look into them.
Recommended CitationTorrey, Amber, "The Discriminant Analysis Used by the IRS to Predict Profitable Individual Tax Return Audits" (2008). Honors Projects in Mathematics. Paper 1.