|Food and Nutrition Bulletin Volume 20, Number 1, 1999 (UNU, 1999, 181 pages)|
|Epidemiology of child developmental problems: The extent of the problems of poor development in children from deprived backgrounds|
Traditional regression models and mean difference models have historically been used to study the impact of poverty on child development. These methods may not be the most effective to analyze the problem. The effects of poverty on development can most effectively be understood in terms of an epidemiological multiple-risk-factor model. Such a model allows a differentiation between the impact of the risks associated with poverty on the individual and the consequences of poverty to the population. This distinction cannot be achieved using regression.
To distinguish between risk associated with the individual and risk to the population is important for making recommendations to clinicians and policy makers. The prevalence of a risk factor will govern its impact on the population. It is possible to have a rare but serious risk factor that is of great clinical importance when present in the individual but is of minor importance to the population. On the other hand, exposure of a large segment of a community to a risk factor can have a great impact on the occurrence of a disorder in the population. Such a relationship can be evident even when the association of risk with an individual appears to be relatively weak when examined in terms of regression coefficients or mean differences. Small mean differences or very modest correlations (even in the range generally considered negligible by researchers in child development) can have large effects on populations.
Much of the importance of population effects is associated with changes in the shape of the normal distribution. For instance, a large increase in the number of cases in the lower tail of the normal distribution will show only a small effect when expressed in terms of product-moment correlations or mean differences . A reader who finds this surprising might reflect on the observation  that the correlation between smoking and lung cancer results in a product-moment correlation of approximately r = 0.10. The difference between smokers and nonsmokers, expressed in terms of risk ratios from the same data, shows that smokers are approximately 11 times more likely to contract lung cancer than nonsmokers.