Trend tests in epidemiology: P-values or confidence intervals?

Frequently, p-values are used in reporting epidemiological trend data. Because a p-value is a confounded mixture of effect size and sample size in dichotomous data, LANG, ROTHMAN, and CANN (1998) recommended the slope of a trend line with its standard error and a graphical presentation containing the rate ratios as a function of mid-exposure levels. However, the slope contains the assumption of a dose-response function. This article discusses a proposal based on odds ratios and the corresponding one-sided lower confidence intervals for pair-wise comparisons ('exposure levels with zero exposure') as well as comparisons between incremental exposure levels. The proposed method allows both decisions on the global trend, the lowest-observed-adverse-effect-level (LOAEL) and the non-observed-adverse-effect-level (NOAEL), and a simple exploratory analysis.