Letter: coffee consumption and gallstone disease – a cautionary note on the assignment of exposure values in dose–response meta‐analyses. Authors' reply

Abstract

We would like to thank Crippa and colleagues for their letter regarding our recent article.1, 2 They raised the point that different exposure assignment values might change the shape of the dose–response curve in meta-analyses and, thus, encouraged direct retrieval of the exposure distribution data from the original authors. We whole-heartedly agree with this point. However, discussion of this issue is warranted. In assigning a single value between boundaries in each category of exposure, the summary slope of individual dose–response studies might be sensitive to the method of assignment.3 According to Berlin et al.,4 assigning a mean value to each category may reduce the apparent slope of a dose–response model, whereas the use of a median method might reduce the sensitivity to a relatively small number of outlier individual exposure values. For our recent article,2 we did not obtain the original exposure distribution data (median and/or mean values) directly from the authors. Instead, we used the mid-point between the upper and lower boundaries, which is often the only method available for assigning exposure levels.4 In such cases, dose–response trends that are calculated using different methods of assigning values for open-ended categories will reflect the sensitivity of the conclusions.4 The nonlinear slope was changed in our analysis when using the other methods of exposure assignment for the open-ended categories as reported by Crippa et al.1 To test the sensitivity of the linear trends in our analysis, we re-analysed the data by setting the highest open-ended category to be the same amplitude as the preceding category. Statistical analyses were conducted using stata 12.0 (StataCorp, College Station, TX, USA). Compared with our method, the mid-point was set at 1.5-times the lower boundary for the highest open-ended category. However, the aforementioned method did not materially alter the inverse association [the previous relative risk (RR) was 0.95; 95% confidence interval (CI) 0.91–1.00; P = 0.049]. In fact, it strengthened the inverse association slightly, as the RR for an increment of one cup of coffee consumed per day was 0.94 (95% CI 0.91–0.97). Assigning values for the highest open-ended categories when original data cannot be retrieved from the authors remains controversial. Crippa et al. reported that setting the highest open-ended category to be the same amplitude as the preceding one might lead to more reasonable values for coffee consumption in our dose–response analyses (up to 6.5 cups/day). However, this assumption may have underestimated the values when coffee was consumed at a rate of at least 4 cups/day in the highest category (up to 4.5 cups/day). We believe both methods of assignment mentioned above are feasible. We agree with the recommendation that the exposure distribution data should be retrieved from the original studies. When original data cannot be retrieved, to make sound recommendations, further studies concerning the sensitivity of different exposure value assignments for the highest open-ended category in dose–response meta-analyses are needed. The author's declarations of personal and financial interests are unchanged from those in the original article.2