Seeking Shelter: Empirically Modelling Tax Shelters Using Financial Statement Information
Lisowsky, P. (2010). Seeking shelter: Empirically modeling tax shelters using financial statement information.The Accounting Review,85(5), 1693-1720.
This study uses confidential IRS data to examine firm determinants of Tax Shelter participation with the goal of creating a model of Tax shelter participation which can be followed by researchers without access to confidential information. In supplemental tests, the author looks at relationships between this variable on other tax avoidance variables found in the literature.
The model in the paper requires identification of firms that engaged in tax shelters. Traditionally, tax shelter use by corporate firms was identified through tax court cases or instances of media reporting. This results in a potential bias towards firms whose tax shelters (private firm tax planning) ultimately became public. Lisowsky (2010) uses IRS Office of Tax Shelter Analysis (OTSA) data to identify two types of tax shelters. Listed Transactions, which are illegal tax shelters, and Reportable Transactions, which although legal, require firms to file additional informational reporting forms with the IRS on their tax returns. OTSA gathers information on tax shelters through both reporting by the firm as well as through enforcement actions. The sample period for the tax shelter identification is 2000-2004 which results in a final sample of 267 firm-years with tax shelters and 744 shelter observations. The total sample for all firms (shelter and non-shelter) is 9,223 firm-year observations. IRS data is supplemented with data from Compustat. Due to the requirement of Compustat data, the sample includes only public companies.'
In the main analysis, Lisowsky uses a pooled, cross-sectional logistic regression model by firm and year. Robust standard errors are used, and all continuous variables are winsorized at the 1stand 99thpercentile.
First, Lisowsky compares the Wilson (2009) Tax shelter model which uses publically available tax shelter data with the same model run using OTSA tax shelter identification. Overall, results are consistent in sign, but in some cases different in magnitude and significance. The paper then adds in additional explanatory variables developed in the paper (including Tax Haven use, Lagged effective tax rate, NOL presence, and others). Ultimately in the full model, the psuedo-R2goes from 13% (Wilson, 2009) to 25.8%. Out of sample validation tests have Area under the ROC curve between 0.69-0.88, indicating that the test does have discriminatory power.
Using a similar approach to the out-of-sample validation tests, Lisowsky calculates various tax variables used in prior literature and using a univariate logistic regression calculates their discriminatory power to identify tax shelters. Using Discretionary Permanent Book-Tax Differences (Frank et al., 2009), Long-run Cash ETR (Dyreng et al., 2008), and Total Book-Tax differences the author finds that they have poor discriminatory power, with only Total Book-Tax differences having an Area under ROC Curve that is significant at 61%. Lisowsky also looks at Tax Cushion which is calculated using IRS data and finds that it has a higher (69%) Area under the ROC curve.
Finally, the paper looks at how Fin 48 regulations may have affected the model. Tax shelter scores are estimated using the main model results for a sample of firms in 2007 and are used to predict actual tax shelter usage in the Post-Fin 48 period. Although the result shows that there is still predictive power, the results are weak (p< 0.10), which may indicate that firm characteristics have changed since 2004, or that Fin 48 did change the regulatory and reporting environment.