Quality Control

QC asks a practical question: do the summary statistics and LD look consistent enough to trust the fine-mapping result?

Run QC before interpreting a locus, especially for multi-ancestry analysis.

Run QC

credtools qc loci_list.txt qc_results --threads 4

The main summary is:

qc_results/qc.txt.gz

For each locus, CREDTOOLS also writes detailed tables under:

qc_results/{locus_id}/

Run QC With Outlier Removal

credtools qc loci_list.txt qc_cleaned \
  --threads 4 \
  --remove-outlier

If outliers are removed, CREDTOOLS writes cleaned files and a new input list:

qc_cleaned/cleaned/cleaned_loci_info.txt.gz

You can pass that file to finemap:

credtools finemap qc_cleaned/cleaned/cleaned_loci_info.txt.gz finemap_cleaned \
  --tool susie

What the Checks Mean

Check	What it can reveal
Kriging RSS / lambda-s	summary statistics and LD disagree
Dentist-S	variants with suspicious association patterns
MAF comparison	allele frequency mismatch
SNP missingness	variants present in one study but missing in another
LD decay and LD moments	unusual LD structure
Cochran's Q	heterogeneity across studies

Kriging RSS and Dentist-S

Kriging RSS compares observed z-scores with LD-implied expected z-scores. The main detailed file is:

qc_results/{locus_id}/expected_z.txt.gz

Important columns:

Column	Meaning
z	observed z-score after CREDTOOLS standardization
condmean	LD-implied expected z-score
condvar	conditional variance
z_std_diff	standardized residual
logLR	evidence for allele-switch-like LD mismatch
lambda_s	locus-level RSS regularization estimate

Dentist-S checks whether each variant is consistent with the lead variant and its LD to that lead variant. Its detailed file is:

qc_results/{locus_id}/dentist_s.txt.gz

The default Dentist-S rule counts variants with -log10p_dentist_s >= 4 and r2 >= 0.6.

Use both checks together. A high lambda_s with many kriging RSS outliers often points to broad LD/reference mismatch. A small number of Dentist-S outliers can be local allele, build, or imputation problems around a strong lead signal.

Useful Flags

Flag	When to touch it
--logLR-threshold	LD mismatch calls are too strict or too loose
--z-threshold	z-score outlier calls need adjustment
--dentist-pvalue-threshold	Dentist-S calls need adjustment
--enable-c1b	you want the extra high-z residual rule
--adaptive-qc	you want a two-stage cleanup for elevated lambda-s

Most users should start with defaults.

How to Use QC in Practice

1. Run the pipeline or standalone QC.
2. Check qc_run_summary.log for failures.
3. Inspect qc.txt.gz.
4. Plot the locus if a metric looks odd.
5. Use outlier removal only when you understand what was removed.

Do not blindly remove outliers

Outlier removal can help with clear mismatches, but it can also remove real signal if thresholds are too aggressive. Keep the original run and compare.