Hi ~
Thank you for providing such a useful tool for cell-type proportion estimation and downstream analysis. I have been using scDC(scDC_noclustering) to analyze my single-cell dataset and encountered a question when interpreting the results of the GLM modeling step (fitGLM function).
In your code, it appears that both fixed-effect and random-effect models are fitted when subject_effect = TRUE. However, in my analysis, I consistently observe that the fixed-effect model produces more significant results and is better able to detect changes in major cell types, whereas the random-effect model often yields non-significant p-values, even for cell types with large observed differences.
Given this discrepancy, I would like to ask:
-
How do you recommend interpreting the results when the fixed and random effect models disagree?
-
Under what conditions would you suggest using the random-effects results instead of the fixed-effects ones?
-
Is there a standard way or recommended metric (e.g., subject-level variance or heterogeneity) to decide whether subject_effect should be considered?
I’m currently exploring whether meta-analysis-style heterogeneity metrics (such as I² or variance estimates from random effects) might help guide the model choice, but I would appreciate your guidance on the intended or best-practice usage within the scDC framework.
Thank you for your time and the great work on this package!
Best regards,
Yingxue Xiao
Hi ~
Thank you for providing such a useful tool for cell-type proportion estimation and downstream analysis. I have been using scDC(scDC_noclustering) to analyze my single-cell dataset and encountered a question when interpreting the results of the GLM modeling step (fitGLM function).
In your code, it appears that both fixed-effect and random-effect models are fitted when subject_effect = TRUE. However, in my analysis, I consistently observe that the fixed-effect model produces more significant results and is better able to detect changes in major cell types, whereas the random-effect model often yields non-significant p-values, even for cell types with large observed differences.
Given this discrepancy, I would like to ask:
How do you recommend interpreting the results when the fixed and random effect models disagree?
Under what conditions would you suggest using the random-effects results instead of the fixed-effects ones?
Is there a standard way or recommended metric (e.g., subject-level variance or heterogeneity) to decide whether subject_effect should be considered?
I’m currently exploring whether meta-analysis-style heterogeneity metrics (such as I² or variance estimates from random effects) might help guide the model choice, but I would appreciate your guidance on the intended or best-practice usage within the scDC framework.
Thank you for your time and the great work on this package!
Best regards,
Yingxue Xiao