p-values > 1

[bschilder]

After munging a bunch of summary stats from OpenGWAS, and trying to convert them to MAGMA files, I noticed MAGMA was throwing this error for a number of datasets, indicating that some values in the "P" col are >1 even after munging. Do we have any checks in place for this situation?

Sources

Possible reasons why this might occur in these cases:

• "P" column means something other than p-value.
• A bug in their p-value calculation that doesn't prevent the upper limit from being >1.
• Some transformation we did in the course of munging (eg logging/anti-logging when we shouldn't have)?

Here's one of the problematic files: https://gwas.mrcieu.ac.uk/datasets/ubm-a-103/

I read the munged file in and got some more info on this P column. Based on this distribution, it definitely does seem like something is awry.

Solutions

• Check that all p-values are between 0-1, and it they aren't provide a warning to the user with some info on this (e.g. min/max values in this col, the number of rows with abnormal p-values) so that they can be aware if there's any larger issues with how their sumstats files were created to begin with.
• If they meet some criterion (e.g. no negative values) we can automatically cut off the "p-values" at 1.

[Al-Murphy]

No this isn't something I thought to add but definitely sounds reasonable. This check should come after the small p-value check though: https://github.com/neurogenomics/MungeSumstats/blob/master/R/check_small_p_val.R as in certain instances the p column could be read in as a character field up to this point.

Do you want to add this to the branch you were working on? You should be able to copy the template from check_small_p_val() (make sure to include the imputation indicator - line 49-51) and then add the warnings and a parameter to remove p>1 and less than 0 with a default of TRUE?

[bschilder]

Sounds good, I was actually thinking the same thing. I'll work on that today.

[bschilder]

Added the changes to the NEWS, but here's the part relevant to this Issue:

• Added checks for p-values >1 or <0 via args convert_large_p and convert_neg_p, respectively. These are both handled by the new internal function check_range_p_val, which also reports the number of SNPs found meeting these criteria to the console/logs.
• check_small_p_val records which SNPs were imputed in a more robust way, by recording which SNPs met the criteria before making the changes (as opposed to inferred this info from which columns are 0 after making the changes). This function now only handles non-negative p-values, so that rows with negative p-values can be recorded/reported separately in the check_range_p_val step.
• check_small_p_val now reports the number of SNPs <= 5e-324 to console/logs.
• Unit tests have been added for both check_range_p_val and check_small_p_val.
• parse_logs can now extract information reported by check_range_p_val and check_small_p_val.
• New internal function logs_example provides easy access to log file stored in inst/extdata, and includes documentation on how it was created.
• Both check_range_p_val and check_small_p_val now use #' @inheritParams format_sumstats to improve consistency of documentation.