Search
Question: Use interactions to get the difference in the effect of two treatments
1
5 months ago by
rrcutler50
rrcutler50 wrote:

Hello all,

I have two experimental conditions and one control, I want to compare the difference of these effects in reference to the control. A simple way to do this would to get the gene sets that result from comparing each experimental level to the control separately, and then take the difference of these sets. However, this might miss where the difference in effects is direction or magnitude.

However, I am wondering if this can also be legitimately done using an interaction term? One would duplicate the control condition so that the setup would look like so (ignoring replicates here):

design(dds) <- ~Condition * Experimental_Condition
Sample  Condition Experimental_Condition
cond1 Experimental cond1
cntr Control cond1
cond2 Experimental cond2
cntr(duplicated) Control cond2

Note that the second control is just a duplicate of the first.

The resulting genes from the interaction term, conditionExperimental.Experimental_Conditioncond2, would be where the effect was significantly different between the condition effects in reference to the control. I realize this is similar to the LRT test, where we would be looking to see if there are any differences at all between all the levels. But, we will still not get only the genes where the effect is different. Please correct me if I'm wrong.

I also realize that this might mess with the dispersion estimates, as there is now a pseudo replicate.

Thanks,

-R

modified 5 months ago by Michael Love18k • written 5 months ago by rrcutler50

One control and two treatments, I think that's the same situation as in this post: Should I contrast 2 treatment groups using the control group as reference or directly against each other?

1
5 months ago by
Michael Love18k
United States
Michael Love18k wrote:

"Note that the second control is just a duplicate of the first. "

You mean, the identical samples? So the columns of counts are identical?

Note the following re-arrangement of terms:

(X-Z) - (Y-Z) = X-Y

If you want to compare condition 2 to condition 1, and they share the identical control, then it drops out of the numerator and denominator of the fold change. Am I reading your question correctly?

Hi Michael,

Yes, the columns of counts are indeed identical. Your equation makes this situation clear. So in order to find the difference in effect between the conditions, the comparison of two conditions to the control is actually the same as the direct comparison of the conditions to each other? This seems also to be what you are saying in the link in your comment, If I want the difference between the effects in reference to the control, it would be best to compare the treatments directly. This would be more robust than comparing the gene list resulting from the effects of the conditions on the control?

-R

1

Yes. Just compare the treatments directly. Adding extra columns invalidates the statistical methods because the additional, duplicated samples are not independent, they are perfectly correlated with the originals.