Kickstarting R - Multiple comparisons

Significance testing is one of the more lively areas of statistics. In general, the idea is not to make too many mistakes in our conclusions. If this only applied to Type I errors, we could all relax, apply the most conservative tests of significance possible and restrict ourselves to the study of the glaringly obvious. R attends to the problem of significance testing in some ways, but sensibly avoids prescribing methods which may not be appropriate for particular analyses.

Adjustments for multiple comparisons

The basic method of adjusting for multiple comparisons is to define the group of comparisons that are to be tested and select an appropriate method of adjustment and the overall probability of a Type I error (perhaps considering the implications of Type II errors). Then, either define a critical probability which any test in that group must exceed or adjust the probability of each test individually and compare that to the selected overall probability of Type I errors. The latter method has been established in R in the function p.adjust(), but it's a bit awkward to integrate with functions like anova() that may produce a table with a number of probabilities.

Using the infert data set, we'll apply the Bonferroni correction to multiple tests of the prevalence of induced labor within groups defined by educational attainment in the infert data set. First let's go through the function group.prop.test() that I found useful for repetitive testing of groups of Bernoulli trial (success/failure) data where the outcome of interest was which groups differed from the overall proportion, that is, which groups were better or worse than the average level of success by a fairly conservative test.

The usual checks of the input data are performed, then the overall proportion is calculated and the result list is set up, filled with blanks and zeros. For each group defined by the grouping vector by, a test of proportions is conducted, and the adjusted probability stored in the appropriate element of gptest. Notice that the formatting of the group names was performed after the calculation. Otherwise the comparison performed by subset would have failed. After the calculation, the results are printed out and the list of results is returned invisibly. By playing around with this data, you may discover that a simple test of the contingency table indicates that the groups do not come from the same population, but in fact none differ from the average prevalence of induced labor, at least by this test.

When I originally wrote the function, it simply printed out the critical (corrected) p-value at the top of the table, and all of the observed values were compared with that.

Back to Table of Contents