Keeping it off the books: an empirical investigation of firms that engage in tax evasion

Abstract

This article uses a unique dataset that contains detailed information on firms from around the world to investigate factors that affect under-reporting behaviour. The empirical strategy employed exploits the nature of the dependent variable, which is interval coded, and uses interval regression which provides an asymptotically efficient estimator provided that the classical linear model assumptions hold. These assumptions are investigated using standard diagnostic tests that have been modified for the interval regression model. Evidence is presented that shows that the firms in all regions engage in under-reporting. Regression results indicate that government corruption has the single largest causal effect on under-reporting, resulting in the percentage of sales not reported to the tax authority being 51.3% higher. Taxes have the second single largest causal effect on under-reporting, resulting in the percentage of sales not reported to the tax authority being 18.0% higher, followed by access to financing at 8.9% higher and organized crime at 7.6% higher. Inflation, political instability, exchange rates and the fairness of the legal system were found to have no effect on under-reporting. It is also found that there is a significant correlation between under-reporting and the legal organization of the business, size, age, ownership, competition and audit controls.

Acknowledgements

I would like to thank Giorgio Valente, an anonymous referee, Tom Crossley, Mike Veall, David Bjerk, Wayne Simpson, David Giles and seminar participants at the University of Victoria, the 2006 Canadian Economics Association and the 2005 Tax Policy: Perspectives from Law and Accounting conference for their helpful comments and invaluable guidance. I am grateful for the diligent research support provided by Hassan Azziz and Adian McFarlane. I would also like to gratefully acknowledge the financial support from the Social Sciences and Humanities Research Council (No. 752-2004-1096 and No. 501-2002-0107), the University of Manitoba and the University of Victoria. I am solely responsible for any remaining errors and omissions.

Notes

¹ The focus of this article is on illegal tax evasion and not legal tax avoidance. Tax evasion or tax noncompliance refers to income tax that is legally owed but is not reported or paid whereas tax avoidance refers to legal actions taken to reduce tax liability. There exists a voluminous literature on corporate tax avoidance, notably work by James Hines at the University of Michigan, and interested readers are encouraged to consult this literature.

² Such methods include: the Currency-Ratio Approach (Gutmann, 1977); the Monetary-Transactions Method (Feige, 1979); Tanzi's Approach (Tanzi, 1980); National Accounts/Judgmental Methods and the Latent Variable/Multiple Indicator Multiple Indicator Cause (MIMIC) model (Frey and Weck-Hanneman, 1984).

³ As previously noted, an extensive literature exists on tax evasion by individuals and the seminal contribution was provided by Allingham and Sandmo (1972). This model has been extended in a number of dimensions over the last 30 years and this literature is nicely surveyed by Andreoni et al. (1998). There is a much smaller pool of literature that addresses tax noncompliance by businesses and Cowell (2004) provided an excellent review of this literature. The theoretical findings clearly suggest that the type of business, whether it be owner managed or otherwise, along with other factors such as tax rates, audit rates and penalties greatly affects the tax evasion decision of firms.

⁴ The TCMP features data from a random sample of individual and small corporate income tax returns filed in a given year that were subject to intensive audits by experienced examiners. For certain groups, such as nonfilers and proprietors who tend not to report a significant amount of their income, the results from special research studies are used to supplement TCMP data. Finally, data for large corporations are obtained from routine operational audits.

⁵ In the WBES, over 40% of the firms surveyed indicate that 100% of their sales are reported to the tax authority and between 6 and 14% of firms appear in one of the other categories. This means that the linear probability models estimated by Batra et al. will be either 0 or 1 inflated. This results in a large number of predicted probabilities falling outside the unit interval.

⁶ Unfortunately, the 1997 World Development Report dataset contains neither detailed information on characteristics of the firms nor the key variable of interest contained in the WBES dataset. As a result, it cannot be appended to the WBES dataset.

⁷ Unfortunately, the survey does not contain survey weights. The lack of survey weights means that individual indicators should not be used for precise country rankings in any particular dimension measured by the survey.

⁸ Similar to the data used by Johnson et al., the exact phrasing of the question was ‘What percentage of total sales would you estimate the typical firm in your area of activity reports for tax purposes?’.

⁹ Just over 18% of the firms in the sample did not respond to the question of interest. It is quite common in survey data to have missing data of this nature. If the data is missing for unknown reasons and the missing data is unrelated to the completeness of the other observations then it is reasonable to exclude the missing observations from the analysis, though efficiency is sacrificed. If, however, the missing data are systematically related to the phenomenon being modelled then there will be a sample selection problem. For example, if firms whose response is missing for the sales reporting question are more likely to be firms who under-report their income, then the missing observations represent more than just missing information. Unfortunately, there is insufficient evidence to definitively ascertain the reasons for the missing observations in this particular case. Rather, the relationship between these missing observations and firm characteristics that will be used as explanatory variables in our model were explored. If the missing observations are randomly distributed amongst these explanatory variables then this provides some evidence that the missing information may not cause systematic bias in the estimates. There appeared to be no significant systematic relationship between nonresponse to the sales reporting question and the explanatory variables, with one exception; no firm located in Albania responded to the question, indicating that survey was not carried out in a similar fashion to that in other countries rather than a response bias. As a result, Albania is excluded from the analysis.

¹⁰ Wik et al. (2004) extended the interval regression model to include random effects.

¹¹ Caudill and Jackson (1993) discussed the implications of heteroscedasticity in the interval regression model and apply the correction outlined here but they neither formally test for heteroscedasticity nor discuss the functional form and normality assumptions and associated tests. Reilly et al. (2004) discussed and applied the tests for functional form, heteroscedasticity and normality but, while they found that their model suffers from heteroscedasticity and nonnormality, they did not make any modifications to their model based on these results.

¹² When an intercept is estimated so that x always contains a unit element, e ² is redundant in the test for homoscedasticity.

¹³ In STATA, the heteroscedastic corrected interval regression model is estimated using the ‘het’ option on the ‘intreg’ command.

¹⁴ It should be noted that the sensitivity of the results and predicted values to the treatment of the end intervals is often overlooked in the literature, yet can be of clear importance to the economic validity of the resulting estimates.

¹⁵ The McElvey and Zavoina's (1975) R ² is computed in STATA with the ‘fitstat’ command.

¹⁶ The diagnostics tests for functional form, heteroscedasticity and normality that are outlined in this article are not performed on the heteroscedasticity corrected model. Not only are these tests much more difficult in the presence of heteroscedasticity (Gasser et al., 1986) but the form, distribution and power of these tests have not been confirmed for the heteroscedasticity corrected model. The author is unaware of any existing work that addresses these issues and these theoretical issues constitute an area for future exploration.