Predicting crime trends is difficult and fraught with unfounded assumptions. There is some question as to whether this endeavour is of any value whatsoever.

One of the simplest predictive methods was presented in an article in the previous edition of the Index (Vol. 2, No 2). The projections made there were derived from examining established crime trends and extrapolating these to the future. The kind of analysis used was regression analysis, which is dependent on the assumption that past trends will continue apace — and by implication that changes to the underlying causes of crime will be similarly consistent. As such, and as that article made clear, the predictions were abstracted from the social context and dealt with only two determinants: a) the absolute size of the population and b) the rate of crime. Using only two factors weakens this method. Predictions for any region or offence are not linked to an underlying pathology. It is assumed that the profiles of victims and perpetrators as well as the social, moral and economic milieu within which they meet will continue to change consistently in rate and direction for the period of the projection.

The political, social and economic changes currently experienced by the country undermine any such premise. Political transformation has effected greater openness and with it, a fundamental redefinition of social roles. Heightened aspirations have been coupled with economic reforms that have not yet been associated with job creation and increasing incomes. What transpires is an increasing juxtaposition of affluence and poverty and a heightening of the sense many have of their own deprivation. Although 1992 is widely accepted as the juncture marking the onset of political and economic change, it is still far from clear how economic and political transformation will manifest themselves. It is, in part, this uncertainty coupled with the fluidity of social role definition that primarily informs crime trends.

To break from making predictions based on simple regression requires linking social attributes of perpetrator and victim to crime trends. This becomes a minefield in terms of both scientific constraints and political implications. The ascription of crime to particular sectors of the community is both methodologically suspect and politically questionable. Nevertheless such projections would ideally be based on indices reflecting perceptions, attitudes, need, etc. The near-impossibility of obtaining such an information base compels reliance on proxy data from census and other surveys. The latter include income profiles, unemployment rates, and various measures of need. If the relationship between these proxy measures and crime rates were described and quantified it would be possible to treat trends as a function of demographic, economic, social or political change.

Computing crime trends

One way of quantifying these relationships is via Artificial Neural Networks (ANN). ANNs operate by emulating human learning processes. These programmes "learn" from examples what the patterns between predictor and predicted variables are. In this instance it would learn the relationship between the proxy socio-economic baseline data and crime rates. Once the patterns have been learnt, predictions can be made for other datasets i.e. where the crime rate is not known. If the "learning" dataset contains a time series the ANN will be able to learn the pattern between input (predictor) variables and the output (predicted variable). When applied to a dataset, it predicts the next value in the time series. As with regression analysis, predictions are based on the assumption of a certain continuity in trends. However the continuity is now dependent on underlying socio-economic data rather than on the mere existence of a prior pattern.

The choice of time period is naturally important for the prediction of crime trends. For the purposes of this article, socio-economic profiles of police districts were generated from the 1991 census data. These profiles were then combined with the number of reported crimes (by type) for each year between 1994 and 1997. The ANN was then set to learn the pattern between the underlying socio-economic profile and the various crime rates. Predictions of crime rates for each of the police districts were then calculated for 1998 and the year 2000.

Two important observations can be made. First, the total number of incidences reported will increase by approximately 3 percent between 1997 (the date of the latest available information) and 1998. However this trend will thereafter be immediately reversed and by the year 2000 crime rates will return to slightly above current levels. This represents a real decrease in the crime rate because the population will expand by over five percent during the same period.



Figure 1 Crime and population growth rates

Source: Crime and population growth rates

Second, the predictions indicate an dependence on socio-economic criteria. Certain regions such as the Upper Karoo and Namaqualand currently have far lower crime rates than suggested by their socio-economic profiles. The ANN reflects these anomalies as an anticipated massive increase in crime rate — in both cases a doubling of crime rates is "anticipated". Anomalies such as this are a reflection of the baseline data’s failure to give a sense of social networks which militate against criminal activity. The other police regions are typified by marginal increases or decreases in crime rates with the overall effect indicated above.

No correlation was assumed between any one, or several, socio-economic factor/s and crime rates. The relationships were established solely by the ANN on examination of the established patterns. The ability of ANN to identify these patterns presents enormous opportunities. The 1996 census results, when available, can be used to better reflect recent social transition and will allow for a more nuanced exploration of the relationship between crime and socio-economic profiles. Furthermore if crime data was available at police station level it would then be possible, for example, to take into account the effect on crime rates of juxtaposing suburbs of poverty and of affluence. All in all the ANN method promises to greatly enhance the quality of crime projections. Such quality will be demonstrated by the year 2000 if victim and police reporting of crime does not deteriorate in the interim.

Michael O’Donovan,
Chief Researcher
Human Sciences Research Council