Suppose there's a treatment applied to a cell culture to see if there's an effect on gene expression. The measurements are performed at two time points: 10 and 20 days after treatment. For the sake of simplicity, suppose that I work with the following model:

$$ log y_{s, g} = \beta_{0} + \beta_{t}x_{t} + \beta_{20d}x_{20d} + \beta_{int}x_{t}x_{20d} $$

where $x_{t}$ is 0 for control and 1 for treatment and $x_{20d}$ is 0 for 10 days after treatment and 1 for 20 days. $s$ and $g$ are sample and gene index, respectively.

I want to select all genes where effect of treatment is the same between two time points. I can do it like so:

  1. Use the full model and select genes where $\beta_{int}$ is not significant
  2. For those genes I use the simpler model $log y_{s, g} = \beta_{0} + \beta_{t}x_{t}$ to select genes for which treatment $\beta_{t}$.

However, is there a way to write a single contrast that I can use to select genes with stable treatment effect over time and avoid doing multiple tests for a gene?



1 Answer 1


There are some misconceptions here. If you want to select genes where the effect is the same, then your step

  1. Use the full model and select genes where $\beta_{int}$ is not significant

is not that. A non-significant test does not mean that the effect is the same. For example, consider a case where there is a difference but you have very little data and thus not enough power to detect it: the test will be not significant but there still is a difference.

Perhaps a better question to ask is whether the difference in gene expression after the 10th and the 20th day is significantly affected by treatment? To this end you could compute a new variable, which for each gene is the difference between the effect on the 10th and 20th day, i.e.

$$ y_{new} = y_{10d} - y_{20d}$$,

and then check whether there is a significant effect of treatment:

$$ y_{new} = \beta_0 + \beta_t x_t + \epsilon$$,

  • $\begingroup$ "...the test will be not significant but there still is a difference." - well, that's conceptually true. But I'm ok with having some false negatives. What you're proposing seems to be a different hypothesis. I want to specifically select genes where treatment effect is the same for time points. Why isn't the interaction term a good choice? $\endgroup$ Commented Dec 24, 2020 at 19:58
  • $\begingroup$ in that case why don't you just test whether $\beta_{int}$ is not significant and $\beta_t$ is significant in your full model? $\endgroup$
    – Jan
    Commented Dec 24, 2020 at 22:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.