Appendix C: New Mexico Crime Analysis Methods
Sample and Data
The study used two analytical models—difference-in-differences and interrupted time series—to compare crime rates in New Mexico to those in neighboring Colorado and Texas to determine whether New Mexico’s forfeiture reform had an effect on crime. Arizona enacted forfeiture reforms during the study period, so we could not use it as a control state.
The unit of analysis was the county. Law enforcement agency-level data for each county were summed to create county totals. Agencies included in the analyses were sheriffs, police and others assigned a county FIPS code.
To have data from enough time periods to run these models, we collected data at the monthly level between 2010 and 2017. Colorado enacted forfeiture reforms in July 2017, so we were unable to use it as a control after that date. However, some analyses used only Texas, which enacted no forfeiture reforms during the period, as a control. Those analyses go through December 2017. To generate a balanced panel, we dropped data for 2010 to 2012 from the analysis due to inconsistent agency reporting. This gave us 53 months in the pre-period and 24–30 months in the post-period.
The literature suggests a one- to two-year delay may be necessary before a policy change’s effect on crime rates, if any, becomes detectable. For this reason, we examined two and a half years of post-reform monthly crime rate data. 1
Data sources are indicated in the table below.
Table C.1: Data Sources
|Crime rates||Overall monthly offenses per capita||Federal Bureau of Investigation Uniform Crime Reporting Program 2|
|Monthly arrests per capita, by type of offense:
• All arrests
• Drug possession
• Drug sales
|Population||Annual county populations, linear interpolation used to generate monthly figures||U.S. Census Bureau|
|Unemployment||Annual county unemployment, linear interpolation used to generate monthly figures||Bureau of Labor Statistics|
|Police||Annual county number of sworn law enforcement officers, linear interpolation used to generate monthly figures||FBI UCR Program; some 2017 figures collected directly from counties|
The offense data provided by the FBI include some imputed figures due to agency non-reporting, which may impact the data’s reliability. 3 As reporting compliance has improved in recent years, the need for imputation has decreased and become less common. 4 In fact, the data used here are very recent and have only small amounts of imputation, thus significantly increasing reliability.
The arrest data were not imputed by the FBI at the agency level, so we performed two different methods of imputation to balance the panel and account for agencies with inconsistent reporting. In the first method, we dropped agencies with fewer than 48 months of data. For remaining agencies with missing data, we interpolated the months with missing crime data using a linear interpolation method drawing on the nearest months with non-missing crime data for each agency. In the second method, we ran all models using arrest data on which we conducted multiple imputation but did not drop agencies with poor reporting. The two imputation methods produced consistent results.
We also interpolated monthly estimates of law enforcement officers and population. Those data are available only at the annual level, so we used a linear interpolation method to estimate the monthly numbers.
Table C.2: Descriptive Statistics
|Crimes Per 1,000 Population||Covariates|
|Offenses||All Arrests||DUI Arrests||Drug Possession Arrests||Drug Sales Arrests||Police Officers||Population||Unemployment|
|NM Pre-St. Dev.||1.58||2.49||0.23||0.19||0.09||275||133,229||3.0%|
|NM Post-St. Dev.||12.27||2.61||0.27||0.26||0.07||288||133,904||2.2%|
|CO Pre-St. Dev.||1.29||1.98||0.30||0.13||0.09||495||188,005||2.2%|
|CO Post-St. Dev.||1.36||2.30||0.30||0.20||0.05||506||195,546||1.1%|
|TX Pre-St. Dev.||1.75||3.11||0.23||0.84||0.23||946||422,110||2.0%|
|TX Post-St. Dev.||1.52||3.34||0.22||0.56||0.76||993||438,730||1.8%|
Note: Differences between the covariates in the offense and arrest models were trivial. We present estimates as they appear in the arrest models.
We ran models on five different dependent variables, all measured monthly and transformed into natural logs: overall offenses, overall arrests, DUI arrests, drug possession arrests and drug sales arrests. Offenses are the number of crimes that come to the attention of law enforcement, while arrests represent the number of offenses that are cleared by arrest. All models used robust, clustered standard errors. Variables included:
- Y = natural log of per capita crime rates
- NM = 1 if a county is in New Mexico, 0 otherwise
- Timecount = linear count of months in the study period
- Timecount2 = Timecount squared
- Post = 1 if the month was in July 2015 or later, 0 otherwise
- Months_post_change = 0 if the time is pre-July 2015, a linear time count of months after
- NM*Timecount = Interaction of NM and Timecount
- NM*Timecount2 = Interaction of NM and Timecount2
- Post*NM= Interaction of Post and NM
- Months_post_change*NM = Interaction of Months_post_change and NM
- Θ = A vector of time-varying covariates: monthly population, monthly unemployment, monthly number of sworn law enforcement officers
- Ω = Month fixed effects
- Φ = County fixed effects
Analysis 1: Difference-in-Differences
a. Comparing New Mexico to Colorado and Texas as controls, using data through June 2017.
Model 1: Y = β0+ β1Post + β2NM + β3post*NM + θ+ e
Model 2: Y = β0+ β1Post*NM + Ω + Φ + θ + e
b. Comparing New Mexico to Texas as a control, using all available data (through December 2017), running models 1 and 2.
c. Running a. and b., limiting the sample to border counties only.
Analysis 2: Interrupted Time Series
d. Comparing New Mexico to Colorado and Texas as controls, using data through June 2017.
Model 3: Y = β0+ β1Post + β2NM + β3Post*NM + β4Timecount + β5Months_post_change + β6Months_post_change*NM + θ+ e
Model 4: Y = β0+ β1Post + β2Months_post_change + β3Post*NM + β4NM*Months + β5Months_post_change*NM + β6Timecount + Ω + Φ + θ + e
Model 5: Y = β0+ β1Timecount + β2Timecount2+ β3NM*Timecount+ β4NM*Timecount2 + Ω + Φ + θ + e
e. Comparing New Mexico to Texas as a control, using all available data (through December 2017), running models 3, 4 and 5.
f. Running d. and e., limiting the sample to border counties only.
The tables below present regression results from Model 5, which estimates the relationship between forfeiture laws and crime rates as quadratic. We present the quadratic results because they appeared to best fit trends in the data. Results from all models are available upon request.
As explained above, the variable Months is a simple chronological count of the months in the sample period, and Months2 is the square of that variable. NM*Months multiplies the month count with a variable that = 1 if a county is in New Mexico and 0 otherwise, and NM*Months2 multiplies NM and Months2. NM*Months2 enables us to detect if there is a deflection in crime rates and in what year and month it occurred. This isolates the reform’s effect, if any, on crime rates in New Mexico.
Table C.3: New Mexico, Colorado and Texas, Jan. 1, 2013, to June 30, 2017
Table C.4: New Mexico and Texas, Jan. 1, 2013, to Dec. 30, 2017
Table C.5: Sample limited to border counties in New Mexico, Colorado and Texas, Jan. 1, 2013, to June 30, 2017
Table C.6: Sample limited to border counties in New Mexico and Texas, Jan. 1, 2013, to Dec. 30, 2017