Dear community,
My question is about statistics.
I have a complex experimental dataset (hundreds of samples) in which I want to test several factors and crossed factors. I know that before running any statistical test, it is very common to reduce the library sizes (which are about 2500 transcripts per sample) by applying a filtering step on the fold-change. However, fold-change is tricky in my context since it should be done between two conditions of a factor, but I want to test for multiple factors and their interaction (eg: y ~ factorA + factorB + factorA:factorB + factorC).
So, to cope with this, I was considering to filter the library sizes by performing what I call a "statistical refinement", i.e., performing a first statistical test with the complete GLM model and from this I would subset all transcripts having p<0.05 for at least one of the factors (not padj).
It has the effect of reducing my library size to approximately 800 transcripts, which is, I think reasonable to perform my second statistical trial. From this second statistical trial, I will then consider deferentially expressed genes to be statistically significant when padj<0.05 for the corresponding factors.
I have not found studies having processed the data in a similar manner so I would like to know what is your opinion about such method. In my opinion, filtering steps are essential but (i) fold-changes filters are not always relevant (a small fold change of a transcript with very small variability across conditions could have dramatic biological effects and so should not be disregarded), (ii) fold change filters are appropriate only within a simple design (ie, 1 factor with 2 levels, eg: exposed vs non exposed), (iii) otherwise statistical refinement may be a good option to consider.
Many thanks in advance for your constructive arguments and if you have the references of studies showing similar methods (I have not found yet).
Kind regards
My question is about statistics.
I have a complex experimental dataset (hundreds of samples) in which I want to test several factors and crossed factors. I know that before running any statistical test, it is very common to reduce the library sizes (which are about 2500 transcripts per sample) by applying a filtering step on the fold-change. However, fold-change is tricky in my context since it should be done between two conditions of a factor, but I want to test for multiple factors and their interaction (eg: y ~ factorA + factorB + factorA:factorB + factorC).
So, to cope with this, I was considering to filter the library sizes by performing what I call a "statistical refinement", i.e., performing a first statistical test with the complete GLM model and from this I would subset all transcripts having p<0.05 for at least one of the factors (not padj).
It has the effect of reducing my library size to approximately 800 transcripts, which is, I think reasonable to perform my second statistical trial. From this second statistical trial, I will then consider deferentially expressed genes to be statistically significant when padj<0.05 for the corresponding factors.
I have not found studies having processed the data in a similar manner so I would like to know what is your opinion about such method. In my opinion, filtering steps are essential but (i) fold-changes filters are not always relevant (a small fold change of a transcript with very small variability across conditions could have dramatic biological effects and so should not be disregarded), (ii) fold change filters are appropriate only within a simple design (ie, 1 factor with 2 levels, eg: exposed vs non exposed), (iii) otherwise statistical refinement may be a good option to consider.
Many thanks in advance for your constructive arguments and if you have the references of studies showing similar methods (I have not found yet).
Kind regards