I worked a bit to develop a regression model for annual violent crime data in DC. I sought to create a core model that estimated and accounted for the basic variations in annual violent crime. Once this was done, I added ME variables for the intervention year, 1993, plus lagged ME variables to test if there is a continuing effect of the ME in subsequent years.
For the core model I tested if the non-violent aka personal crimes (PC) were significant in explaining variations in violent crimes. Both in the current year, and also lagged variables, to see if personal crimes affected violent crimes in subsequent years. I hypothesized that these PCs could be correlated to strong control variables such as weather, police on the street, LE funding -- but which I have not yet aquired. This correlation hypothesis makes sense in that PC variables should also go up in hot weather and down with increased police and LE $. (And in regression, a variable that is correlated to a "good" control (independent) variable can be an effective replacement for the actual control variable.) And I pulled down 30 or so national economic variables -- not ideal, but the best I have at the moment -- no DC specific control data yet. This national economic variables should correlate, to a degree, with specific DC economic data, so it is a reasonable first step. I found a five variable model with a fairly good fit (1964-2003) with an adjusted R^2 of .89. That means the model explains 89% of the variation in violent crime. And each variable was statistically significant -- the t-values for each var was >2, meaning the is less than a 5% chance that the variable's contribution is just a fluke, a random chance effect. Several other diagnostics were run: the durbin-watson stat was good, showing that the model did not have undue levels of autocorrelation (variables were not correlated to their previous values -- t-1, t-2, etc.) And there was low correlation between independent variables, called non-collinearity -- an important characteristic for models to have. Another diagnostic, hetroscadasticity was mild. So generally, for a quick model, it was fairly strong. In the below table, for years 1990-2003 you can see the actual annual changes in violent crime in DC in the first column, and the estimated or predicted series from the model in the second column. As you can see, it generally rises and falls in synch with the actual crime data. Having a good core model as a baseline, I then added some ME variables. First 3 variables -- one for the intervention year, and two for subsequent years, to see if there were continuing effects. And then another model with 5 ME variables -- for the test year and 4 subsequent years. The ME effects were interesting. Their effect in the model was to show about a 4% increase in violent crime from the ME in the test year, and then a continuing decreasing crime impact of about 5% in the next 2-4 years. The hypothesis could be that in the intervention ME year, things are stirred up in the "collective consciousness" -- social unstressing so to speak, and then good effects emerge in the subsequent years. However, the significance of the ME variables was weak. There is a 20-50% chance that they are having no more impact than random chance. It could jsut be the abortion effect we have discussed (per Levitt) or other untested factors. Better DC specific data, more economic and demographic and weather data, and the acquistion of monthly data should shed light on this and determine better if there is a significant ME effect that can be demonstrated by this type of analysis. For now, its an interesting and thought-provoking result. Haiglin and all may have been looking at the wrong thing -- current ME effects, instead of where the action may really be --- future effects. That is, ME may have its effect via long-term structural changes in collective consciousness, not immediate ones -- which actually may be negative (washing machine effect, perhaps). Actual ----- Predicted ---- No ME ME 3 ME 5 1990 14.8% 12.3% 12.0% 7.8% 1991 -0.2% 1.6% 1.4% 4.4% 1992 15.5% 4.5% 4.5% 9.6% 1993 3.1% -2.0% 3.1% 4.9% 1994 -8.9% -6.2% -8.9% -10.6% 1995 0.0% 6.3% 0.0% 0.0% 1996 -7.2% -2.6% -2.6% -0.7% 1997 -18.0% -15.7% -16.0% -11.1% 1998 -15.1% -12.6% -12.6% -7.4% 1999 -5.3% -8.2% -8.1% -6.6% 2000 -7.4% -5.6% -5.6% -6.1% 2001 15.2% 8.2% 8.1% 8.0% 2002 -5.7% -0.5% -0.6% 8.2% 2003 -1.8% 0.0% -0.1% -7.5% ------ Estimated Independent Variables I Vars Beta Std. Error t-value Sig LAR 0.146493072 0.086854908 1.68664126 0.101712298 ROB 0.540438633 0.054472023 9.92139825 3.86814E-11 MT-2 0.118127037 0.05135057 2.300403597 0.028316925 UNEI 0.029748611 0.020410475 1.457516846 0.155030356 ME 0.041842476 0.049411284 0.846820246 0.40358555 ME-1 -0.035634855 0.050277557 -0.708762667 0.483768091 ME-2 -0.067917481 0.050595442 -1.342363634 0.189226664 ME3 -0.0532541 0.049249815 -1.081305573 0.287901536 ME4 -0.038371741 0.053936408 -0.711425589 0.48213999 ------------------------ Yahoo! Groups Sponsor --------------------~--> Get fast access to your favorite Yahoo! Groups. Make Yahoo! your home page http://us.click.yahoo.com/dpRU5A/wUILAA/yQLSAA/JjtolB/TM --------------------------------------------------------------------~-> To subscribe, send a message to: [EMAIL PROTECTED] Or go to: http://groups.yahoo.com/group/FairfieldLife/ and click 'Join This Group!' Yahoo! Groups Links <*> To visit your group on the web, go to: http://groups.yahoo.com/group/FairfieldLife/ <*> To unsubscribe from this group, send an email to: [EMAIL PROTECTED] <*> Your use of Yahoo! Groups is subject to: http://docs.yahoo.com/info/terms/