Figure A3.

Comparison of log fold change (logFC) estimates between crude unweighted models and models weighted for inverse probability of selection conditional on self-reported smoking status, logarithm cotinine, and their interaction. Differences in logFC estimates between unweighted and weighted models on the y-axis represent an approximation of the bias due to selection on smoking. Estimates are overall fairly concordant, with nearly all (99%) of OTU-variable pairs having an absolute difference in point estimate less than 0.35. Very few (n=10) hypotheses that were significant (FDR<0.01) in unweighted analysis were nonsignificant in weighted analysis. Of these, 9/10 had nearly identical point estimates but larger variance in the weighted models. In contrast, a large number of hypothesis tests that were nonsignificant in unweighted analysis were significant in weighted analysis. Specifically, weighting by selection for smoking identified 10 new significant OTUs for gender, 13 for age, 24 for education, 10 for income, 13 for marital status, 26 for race, and 8 for nativity. Where the two models disagreed on significance tests, the vast majority of disagreements were characterized by significance in the weighted model and nonsignificance in the unweighted model. Furthermore, the point estimates from the weighted models were more often further from the null than unweighted models.