logfoldchange always -0.430725 for non significant classes

[theislab]

logfoldchange always -0.430725 for non significant classes

[johannesostner]

Yes, this works as intended. The log-fold change column is simply the log-fold change between the "expected_sample" column of the effect results and the "expected_sample"-column of the intercept results.

The expected samples in the intercept results is imply the mean of a Dirichlet-multinomial distribution over the MCMC-sampled parameters for the intercepts:

alpha_par = alphas_df["final_parameter"]
alphas_exp = np.exp(alpha_par)
alpha_sample = (alphas_exp / np.sum(alphas_exp) * y_bar).values

For the effects, we have the same formula but we add the siginificant effects: alpha_par + betas_df["final_parameter"]. This results in the log-fold change being the same for all non-significant cell types.

[le-ander]

Hm, I see.
I guess this is very intuitive when you fully grasp what the underlying model is doing (which I don't and I guess not all future users of the tool will be).

I guess I would have just expected a simple mean log-fold change between conditions over all samples here. Might be worth a discussion what makes most sense with the users in mind here. I wonder what Maren's @mbuttner take is on this.