Hi,

I have a experiment where ERCC mix1 and mix2 were applied cross samples. I was trying to remove unwanted factors using ERCCs in class B which are supposed to be no fold change between mix1 and mix2. After that, I wanna integrate the RUVg-adjusted factors (W_1 as in the manual) into a GLMM formula by lme4 package.  I have the following questions:

During modelling GLMM by lme4, should I use W_1 as a block in the fixed effect part or as an additional offset (current offset is library size) ? As the values in W_1 seems already been log-transformed, I guess it's better to be used as offset?  

Thanks a lot.

We don't have any direct experience using mixed models and RUV, so you should be careful and do extensive EDA and goodness of fit analyses (residuals etc) to make sure that the fitted model is reasonable, as our results with the fixed effects GLM may not translate directly.

That said, I have two comments:

1. You are estimating W_1 with very few negative controls, since if I recall correctly there are only 20 or so spike ins in class B. So be extra careful in making sure that the negative controls are really capturing unwanted variation and they are not capturing the biology of interest. In our experience, with fewer than 100 negative controls, the estimates of W can be really noisy.

2. I would use W_1 as a fixed effect in the model, as this is an (estimated) covariate that captures the unwanted effects present in the data. Using it as an offset is not appropriate because the estimated W alpha at the first step of RUV are unidentifiable, hence, you need to re-estimate alpha in the full model once you fixed a particular W. This will also help if W and your effects of interest are correlated.