Comparing prevalence / incidence rates

Some cross sectional studies may collect information on presumptive risk factors as well as health outcomes.  For instance, winter loss surveys may collect information on management practices utilized in addition to health outcome (mortality). When the study permits the population to be divided based on different “exposures”, the measures of disease outcomes (prevalence or incidence rates) can be   compared. When prevalence is the measure of comparison,  differences in exposure between two groups separated by risk factor exposure can be compared using a Chi-Square test, or in cases where fewer than 5 cases were expected in a given cell, the Fisher’s exact test. Resulting from this approach is a p value, which simply provides a goal post by which we can assert that the populations differ significantly (typically when a p ≤ 0.05 is calculated, the prevalence rates in two populations are considered to be significantly different). However, this approach does not give any indication as to the size of the effect of exposure to the risk factor. The magnitude of this effect can be gleaned by comparing the 95% CI of the point prevalence estimates. Generally speaking, populations that have point estimates with overlapping 95% CI are not significantly different, while those who do not have overlapping populations are.  More importantly, the 95% CI aid in the interpretation of any exposure effect in that it puts the upper and lower bounds on possible magnitude of any effect (Gardner and Altman, 1986).

When cross sectional studies result in incidence rates (e.g. from winter loss surveys), rates between groups separated by exposure can be compared using ANOVA and other basic parametric tests. As is the case for the non-parametric tests mentioned in the above paragraph, these will result in a P value which indicates if the incidence rates in the populations differ. This result is of limited value because not only is it of interest that the populations are different; the magnitude of the difference is of note. Calculating and comparing 95% CI for the point estimate of Incidence rates has more meaning than stating that the two groups within a population are different or not based on a statistical test (Box 6).

Box 6.

A winter loss survey was conducted to determine the winter mortality (Oct 1 – April 1) of US beekeepers over the winter of 2010-2011 (vanEngelsdorp et al., 2012). A subset of these respondents also answered various questions regarding their management practices. In all 1,074 beekeepers indicated they had used a known varroa mite control product in a majority of their hives over the previous year, while 1,675 responding beekeepers reported not using any known varroa mite control product in any of their hives. Beekeepers who used a known varroa mite control product suffered an average loss of 29.5% (95% CI 27.5 - 31.4%) of their colonies, while those who did not indicate they used a known varroa mite control product suffered an average loss of 36.7% (95% CI 34.9 - 38.55) (BeeInformed.org Report 30).

Box 6.1

As the two confidence intervals do not overlap we know the two populations are different, we can say that beekeepers who treated with a known varroa control product lost 7 fewer overwintering colonies per 100 than those who did not; in other words beekeepers who treated with a known varroa control product lost 20% (difference in average loss rate / average loss rate in those not treating = 7/37*100 %) fewer colonies than those who did not.