Hello Lydia,

I'm trying to run a jackknife sensitivity analysis by implementing multiple linear models, as per your suggestion above. I've set up dummy variables for each article, where all articles are coded as 1 except for the excluded study, which is coded as 0.

However, I have some questions regarding how to properly run and interpret the analysis:
1. Should I define a model and hypothesis, or use these variables as filters in the linear model?

• I tried running the analysis by selecting the variable (e.g. study_1_excluded) as a model predictor and setting the hypothesis to 1 (as per the GUI baseline recommendation). However, this resulted in an error: "ERROR: matrix is singular Default GSL error handler invoked"


• If this error is due to the lack of variance in the dummy variable (since only one study differs) how should I run the Leave-one-out analysis?


• Should I instead use the filter variable option in the GUI instead?

2. Could you please elaborate on how the linear model is calculated in this specific case?

• Since only one study has a different value, does the model still perform the regression at each voxel?


• How does SDM calculate the voxel-wise slope (β1eta_1β1​) when one group consists of just one study?

3. Interpreting the Results:

• What does a significant β1eta_1β1​ coefficient indicate in this context?


• If a voxel has a large β1eta_1β1​, should I conclude that the omitted study strongly influenced that region?


• Would it be better to compare the intercept maps (across different LOO runs) instead of focusing on the slopes?

I appreciate any guidance you can provide on best practices for running LOO sensitivity analysis using meta-regression in SDM.

I used the SDM Linux version: v6.21
Thank you so much!