Appendix A: Methods

Policing for Profit

This appendix provides further detailed information about the methods used in the statistical analysis of equitable sharing data.  Some parts of this discussion assume a working knowledge of quantitative research methods.

Explanatory Variables—Forfeiture Laws

Each of the state forfeiture laws can be distinguished by the degree of difficulty, or measure of burden, to forfeit property and the financial incentives for law enforcement to engage in forfeitures.[92]  Degree of difficulty is measured by two factors—standard of proof and innocent owner.  The first reflects the standard of proof the government is required to meet to determine property is subject to forfeiture.  This was coded as follows, where lower numbers equal less burden on the state:
1=prima facie/probable cause;
2=probable cause and preponderance of the evidence;
3=preponderance of the evidence;
4=preponderance and clear and convincing; 5=clear and convincing;
6=clear and convincing and beyond a reasonable doubt and
7=beyond a reasonable doubt.

As noted above, some states have two standards depending on the property.  We considered these potentially meaningful distinctions and coded them as falling between the different standards found in the law.

The innocent owner burden variable represents who has the burden (the state or the property owner) to establish whether the property owner qualifies as an innocent owner under state law.  This was coded as follows, where lower numbers equal less burden on the state:
1=the burden to establish innocence rests exclusively with the property owner/claimant;
2=the burden varies depending on the type of property being forfeited and
3=the burden rests exclusively with the government to establish that the claimant is not innocent.

Finally, the percent of proceeds to law enforcement variable was coded based upon the actual number (i.e., percent) reported within the state statute.  The forfeiture distribution language within some statutes was imprecise in terms of the actual percentage guaranteed to law enforcement.  In states with some vagueness, we coded the data based upon our reading of the statute and compared this coding with information reported in other sources[93] and with that conducted by the legal research staff at the Institute for Justice.  While we are confident that these procedures produced the most accurate assessment of proceeds to law enforcement, for any statutes with some remaining vagueness, our guiding principle was to code the percentages conservatively and in a manner that, if inaccurate, would bias results contrary to finding a significant relationship between the percentages and forfeiture revenue collected by agencies.

Since the outcome variable was based on multi-year averages, it was necessary to review all state statutes to determine if any statutory changes occurred to the primary variables of interest during this time.  A few states did change their forfeiture laws, and these changes were entered into the dataset where appropriate to reflect the changes in law affecting law enforcement agencies in those states.

Control Variables

The control variables included in the analyses included the number of full-time officers assigned to special or multi-agency drug enforcement units, the arrest rate (per 100,000 population) for drug manufacturing and selling, the violent crime rate (per 100,000 population), law enforcement agency type, whether the agency was primarily responsible for enforcing drug laws in their respective jurisdiction and region of the country.  Table A1 lists and provides a brief description of each variable used along with their means and standard deviations.

We controlled for the number of full-time officers assigned to special or multi-agency drug enforcement units to examine whether such units mediate the link between state asset forfeiture laws and forfeiture activity.  One might hypothesize, for example, that law enforcement agencies in less restrictive asset forfeiture states may be more inclined to assign a larger number of full-time officers to specialized drug units because doing so is likely to lead to more drug-related asset forfeiture activity and, eventually, additional revenue for the agency.  Data on the number of full-time officers assigned to specialized or multi-agency drug units were obtained from the LEMAS dataset.

We also included the arrest rate (per 100,000 population) for drug manufacturing and selling and the violent crime rate (per 100,000 population) in the analysis because both of these variables may be causally antecedent to both the type of civil asset forfeiture statutes put in place by state policy makers and the amount of drug asset forfeiture activity a jurisdiction can reasonably be expected to engage in based on the sheer number of drug-related transaction opportunities alone (i.e., more drug activity, more asset forfeiture).  If this were true, failing to control for these potentially causally antecedent variables would lead to spurious (or partly spurious) associations for the state asset forfeiture law variables.  Data on the number of persons arrested for selling and manufacturing drugs were obtained from the Federal Bureau of Investigation’s (FBI) Uniform Crime Reports (UCR).  Similar to the outcome variables, we used a multi-year average (2001 to 2003) to solve the problem that drug activity fluctuates widely from year-to-year.  Data on the number of violent crimes (i.e., homicide, forcible rape, robbery, aggravated assault) reported and recorded by law enforcement agencies for 2003 for each jurisdiction were also obtained from the FBI’s UCR.

Two of the controls were binary variables denoting law enforcement agency type (1=municipal agency, 0=sheriff’s department) and whether the agency was primarily responsible for enforcing drug laws in their respective jurisdiction (1=yes, 0=no).  Data for both variables were obtained from the 2003 LEMAS dataset.

The last set of control variables are binary dummy variables for each of the nine U.S. Census regions.  These variables are used to control for any well-established and unobserved (or unmeasured characteristics) of the jurisdictions served by law enforcement agencies that vary at the regional level and that could be expected to influence both state asset forfeiture laws and drug-related asset forfeiture activity.  Examples of such potential confounders would be regions of the United States where drug trafficking is more commonplace and regions in close proximity to a major port of entry for drugs.

Analytic Procedures

We used a censored regression model to determine the impact of three key components of state asset forfeiture laws on the per capita dollar value of forfeiture proceeds returned to law enforcement agencies through equitable sharing payments received from the DOJ’s AFF.  Censored regression models take into account potential biases that may be present when some observations on the dependent variable are not observable, as is the case here.  In the present study, both forfeiture proceeds variables (i.e., the dependent variables) are concentrated at the lower limit value of zero (denoting zero dollars received through asset forfeiture activity).  The appropriate censored regression model in this case is the Tobit model.[94]  Tobit regression estimates a linear regression model for a left-censored dependent variable, where the dependent variable is censored from below.

More specifically, slightly more than 11 percent (63 out of 563) of the agencies in the study sample received no equitable sharing payments from the DOJ between fiscal years 2000 and 2004.  If the probability of zero dollars related to drug-related forfeitures were the only phenomenon to explain, probit regression would provide a suitable model.  Of course, this would result in throwing away information on the value of proceeds returned when it is available.  That is the case here because if a law enforcement agency received forfeiture proceeds related to drug offenses, we have an estimate of the dollar value they received.

If there were no concentrations at a lower limit, and we only cared to explain the dollar amount of assets forfeited, multiple regression would be the appropriate statistical technique.  But, since there is a piling up of values of the dependent variable at a limit (in this case $0), ordinary least squares (OLS) estimates are biased because the dependent variables are not continuous and unbounded.  The solution to this problem is a hybrid of the two regression methods (probit and OLS) which economists refer to as Tobit models.  Similar to standard OLS regression, the Tobit regression model assumes the error terms are normally distributed, independent between observations and uncorrelated with the independent variables.  Model parameters are estimated via maximum likelihood.

Although it is useful to examine ordinary Tobit coefficients for sign and significance, they are not readily interpretable as effect sizes like their OLS counterparts.[95]  The reason interpretation of Tobit coefficients is more problematic than traditional regression coefficients is because the former must account for two distinct types of observations on the dependent variable.  The first set contains the observations, for which the dollar value of assets seized is zero.  For these observations, we know only the values of the independent variables and the fact that the dependent variable is less than or equal to zero.  The second set consists of all observations for which the value of both the independent and dependent variables are known.  Thus, two types of effects are modeled simultaneously in a Tobit regression model:  (1) the effect on the per capita dollar value of assets seized for cases with a nonzero value (uncensored) and (2) the effect on the probability of having a nonzero value for cases with the limit value of zero dollars (censored).  A problem arises, however, because only a single coefficient is provided in the output of a Tobit analysis for each of the two state asset forfeiture law variables.  Clearly, however, it is not possible for a single coefficient to capture both effects—one for cases at the lower limit value (zero dollars) and another for cases above the limit value (nonzero dollars).

Fortunately, a decomposition procedure[96] can be used to disentangle Tobit coefficients in such a way that both different effects are quantified:  (1) the effect of state asset forfeiture laws on the per capita dollar value of forfeiture proceeds returned to law enforcement agencies and (2) the effect of the laws on the probability of a law enforcement agency receiving forfeiture proceeds for those agencies failing to receive forfeiture proceeds associated to drug-related activity.  Decomposing the Tobit coefficients provides for a more complete understanding of the two separate effects state asset forfeiture laws can have on drug asset forfeiture activity.  The Tobit regression models were estimated using the Tobit command in Stata Release 9.0.[97]

Finally, one potential pitfall with using local law enforcement data in a study examining the effects of state laws is that the law variables do not vary across jurisdictions within a particular state.  As a result, errors in predicting asset forfeiture are likely to be correlated within clusters (i.e., states) and conventional estimates of standard errors for the state asset forfeiture law variables may be understated due to violations of the independence assumption.[98]  To address this problem, we used cluster-adjusted standard errors that adjust for the fact that observations within states may not be independent.[99 ]

[92] See generally, Edgeworth, 2008; Sorens et al, 2008; Worrall and Kovandzic, 2008. The coding strategy for the burden variables was similar to that utilized by Sorens et al., 2008.

[93]  E.g., Edgeworth, 2008.

[94] Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica, 26, 24-36.

[95] Wooldridge, J. (2005). Introductory econometrics: A modern approach, 3rd edition. Florence, KY:  South-Western College Publishing.

[96] McDonald, J. F. & Moffitt, R. A. (1980). The uses of tobit analysis. The Review of Economics and Statistics, 62, 318-321.

[97] Stata Statistical Software: Release 9. College Station, TX:  StataCorp LP.

98] Moulton, B. R. (1990). An illustration of a pitfall in estimating the effects of aggregate variables in micro units. Review of Economics and Statistics, 72, 334-338.

[99] Williams, R. (2000). A note on robust variance estimation for cluster-correlated data. Biometrics, 56, 645–646.