Boxplots of the log-odds scores (A) and log-ratios (B) obtained after fitting a linear model to all genes across all arrays in the spike experiment and making contrasts between 3 pM and 1 pM. A separate box is shown for each background correction method with and without bead variances as inverse weights in the linear model. Two separate boxplots are shown for each method and weighting scheme to indicate the log-odds scores for the spikes (bold colours) and non-spikes (transparent colours). The use of weights improves the log-odds scores for the spikes without increasing the log-odds for the non-spikes, which represents an increase in power to detect true DE. In panel B, we show that the log-ratios for the spikes are under-estimated when the data is not background adjusted, whereas the background subtracted and normexp processed data recover values much closer to the true log fold-change (dashed line, M = 1.73).