IV. PROGRAM EVALUATION

C. Alcohol-Related Fatalities

Alcohol-related fatality data was taken from NHTSA’s Fatality Analysis Reporting System (FARS) for 2000 through the preliminary 2004 data.  Crashes involving pedalcyclists and pedestrians were included because they are considered active road users and could either be responsible for a fatal crash in which alcohol was determined to be involved or could be struck and fatally injured by a drinking driver.  The alcohol-related fatality trend was analyzed using an interrupted time series design.  A separate interrupted time series analysis was used to analyze the alcohol-related fatality trend for men 21 to 34 years old because they were the focus of the media efforts to increase awareness of the enforcement campaign.

Using this design, the Autoregressive Integrated Moving Average (ARIMA) method was able to determine if there was a change in the number of alcohol-related fatalities starting at a point in time coincident with the beginning of the first campaign crackdown in July of 2003 and sustained through December 31st of 2004.  ARIMA modeling required the selection of a model that controlled for periodic fluctuations in the data series.   That is, combinations of parameters were entered into the analysis such that systematic fluctuations in the data (i.e., monthly “lags”) were reduced to nonsignificance.  Lags were judged to be nonsignificant based on exploration of autocorrelations (AC) and partial autocorrelations (PAC) where the monthly lags were deemed to be random with 95 percent confidence.  The parameters used to control the lags, as required, significantly affected the series in order to be considered valid for inclusion in the model.  Analyses were conducted using the “Trends” module of the software package SPSS 11.5.  

The ARIMA modeling process applies parameters to account for periodic fluctuations in monthly alcohol-related fatalities.  For instance, alcohol-related fatalities tend to increase sharply over the summer months. There is also the possibility of nonperiodic fluctuations that might occur due to random noise or simply different numbers of weekend days (when drinking and driving are more prevalent) in a given month.  The modeling process accounts for these periodic variations in the series by including the appropriate parameter. Additionally, multivariate ARIMA models, like the one used in this study for analyzing all alcohol-related fatalities, allow for the addition of a “covariate” which examines change in a series in the context of changes in a similar comparison series.  For instance, drinking and driving fatalities can be affected by the weather, economic conditions, regionwide trends in drinking and driving, and regionwide efforts to combat drinking and driving.  Thus, using the alcohol-related fatality totals for all contiguous counties from neighboring States may help to remove the impacts of these three sources of variation on the number of alcohol-related fatalities each month.  The covariate used here for analyzing all alcohol-related fatalities in Connecticut was the combined total alcohol-related fatalities each month for all contiguous counties from the three surrounding States: New York, Massachusetts, and Rhode Island.  The five New York counties were Suffolk, Nassau, Westchester, Putnam, and Duchess.  The three Massachusetts counties were Berkshire, Hampden, and Worcester.  The three Rhode Island counties were Providence, Kent, and Washington.