Downweighting outliers
Downweighting outliers reduces the chance that a single outlier for a feature is sufficient to reject the null hypothesis and to call a feature differentially expressed. In the presence of outliers, downweighting can increase both sensitivity and precision. When no outliers are present, downweighting typically leads to similar sensitivity and reduced precision. Tests on simulated data show that downweighting leads to a modest loss of control of the false discovery rate (FDR) regardless of whether outliers are present. For example, among features with FDR p-value below 0.05, 5% should be false positives, but when downweighting outliers a higher proportion will be false positives.
Because downweighting is only advantageous when outliers are actually present, we recommend using it only when a standard analysis is enriched for genes that are highly expressed in just one sample. This is often easiest to see in a heat map.
Note that downweighting outliers is not a way of handling low quality samples. If a single sample behaves very differently from others, consider removing it from the analysis.
The implementation of outlier downweighting is described in more detail in The GLM model.