NCJ Number
221979
Journal
Journal of Quantitative Criminology Volume: 24 Issue: 1 Dated: March 2008 Pages: 51-72
Date Published
March 2008
Length
22 pages
Annotation
Using a dataset of homicide rates and a vector of explanatory variables for 400 U.S. counties over a 15-year period, this study compared the estimates obtained for fixed-effects and random-effects models for pooled cross-sectional and time series data, as well as latent growth curve models for panel data in special cases of a more general model.
Abstract
The comparison of the two modeling procedures suggests that each has some advantages for the researcher. Pooling methods offer greater flexibility in the handling of trends due to unmeasured variables and more easily allow unmeasured stable characteristics of cases to be controlled. Although these stable characteristics can be introduced into a growth trajectory model, the procedure for doing this can be cumbersome in some software packages. On the other hand, multilevel modeling more easily handles random variation in regression coefficients. Growth curve models indicate whether there is significant variability across cases in the values of the intercepts and slopes; and if so, how much. These models also show whether certain variables explain this variability and by how much. The importance of these differences rests on the types of questions a researcher asks. Most criminological theories focus on the impact of one or more independent variables on a dependent variable, such as a crime rate. Consequently, researchers attempting to test these theories empirically are primarily interested in the magnitude, sign, and statistical significance of the coefficients. For this purpose, a pooled model is the most direct way of reaching the research goal. Growth curve models focus on whether there is significant variability across cases in the values of the intercepts and slopes; and if so, how much. 5 tables and 56 references