min_group_size parameter setting and zscore filtering

[lixingii]

Hi, Thank you for showing this method. When using the latest ctwas_multigroup function, could you please provide some suggestions regarding the setting of min_group_size in the ctwas_sumtats function? What is the minimum value that can be set for this parameter, and how significant is its impact on the results? And in the ctwas results, apart from filtering the PIP, how should we filter the zscore and p values of TWAS?

[kevinlkx]

Thanks for your questions. For the min_group_size, you can try ctwas:::get_group_size_from_region_data(region_data) or table(z_gene$group) and see the group size of your input data. The current default of min_group_size is 100, which is already pretty low. Because ctwas learns the parameters from the data, if there are too few variables in a group, the parameters would not be reliable.

Apart from filtering by PIP, one can also filter by credible set (CS), but typically no need to filter z-scores and p-values from TWAS.

[lixingii]

Thanks for your questions. For the min_group_size, you can try ctwas:::get_group_size_from_region_data(region_data) or table(z_gene$group) and see the group size of your input data. The current default of min_group_size is 100, which is already pretty low. Because ctwas learns the parameters from the data, if there are too few variables in a group, the parameters would not be reliable.

Apart from filtering by PIP, one can also filter by credible set (CS), but typically no need to filter z-scores and p-values from TWAS.

Thanks for your response. Since my input data is relatively sparse, with an average group_size of 40, do I necessarily need to use a parameter of 100? Can this parameter be appropriately lowered?

[kevinlkx]

Please first make sure you are using all the data genome-wide. We have seen users got small group size due to issues when preparing the input data.

If you set min_group_size to 100, the groups with group_size less than 100 will be dropped. Yes, you can set it to a lower number, e.g. 40. However, the parameters may not be very accurate.