`credtools meta`¶

Run meta-analysis as a standalone step.

credtools meta INPUTS OUTDIR [OPTIONS]

Common Use¶

credtools meta work/chunks/loci_list.txt work/meta \
  --meta-method meta_all \
  --threads 4

Options¶

Option	Meaning	Default
`--threads`	worker count	`1`
`--meta-method`	`meta_all`, `meta_by_population`, or `no_meta`	`meta_all`
`--ld-weighting`	geometric LD weighting: `ess` or `se`	`ess`
`--calculate-lambda-s`	estimate lambda-s while loading loci	off
`--skip`	reuse completed loci from a previous run	off
`--log-file`	write logs to a file	none

Outputs¶

OUTDIR/
- meta_config.json
- loci_info.txt
- heterogeneity.txt.gz
- {locus_id}/

Use OUTDIR/loci_info.txt as the next input for qc or finemap.

Meta Methods¶

Method	Behavior
`meta_all`	combine all rows for each locus
`meta_by_population`	combine rows within each population
`no_meta`	keep rows separate

Matched LD Weighting¶

Both ess (default) and se first match GWAS and LD within each cohort, align allele directions, then use the union of those matches. A variant does not have to occur in every cohort. Only finite BETA and positive finite SE contribute. A cohort with no eligible matches contributes zero; an entirely empty output fails explicitly.

In both modes, meta BETA, SE and P are recomputed by inverse-variance weighting from exactly those matched contributions. The option changes LD weighting, not the summary-statistics estimator. Original discovery GWAS inputs are not overwritten; these are LD-supported fine-mapping inputs and can lose power.

For cohort k and SNP j, define w[k,j] as:

ess: the supplied cohort sample_size for an eligible contribution;
se: 1 / SE[k,j]^2 from the input summary statistics;
zero when that cohort/SNP contribution is unavailable.

With a[k,j] = sqrt(w[k,j] / sum_k w[k,j]), merge LD as R[i,j] = sum_k a[k,i] * a[k,j] * R_k[i,j]. There is no pairwise renormalization. Supply effective sample sizes in sample_size for ESS weighting; credtools does not infer case/control ESS from that field. Supply any desired LDSC-corrected SE upstream; the option does not apply a new correction. SE weights describe the covariance of IVW meta Z under independent cohorts and appropriate LD; ESS weights remain an approximation to that covariance. Neither option corrects sample overlap or poor reference LD.

# Default: matched geometric ESS LD
credtools meta work/chunks/loci_list.txt work/meta_ess --meta-method meta_all

# Matched SE LD, with the same matched IVW summary statistics
credtools meta work/chunks/loci_list.txt work/meta_se --ld-weighting se

# The same option works within each population
credtools meta work/chunks/loci_list.txt work/popmeta_se \
  --meta-method meta_by_population --ld-weighting se

no_meta retains its existing intersection behavior and ignores the weighting choice. A single-cohort population still uses the matched input contract, with its BETA/SE and LD unchanged apart from allele alignment/order and numeric rounding; P is computed from BETA/SE.

Migration and Audits¶

This changes the default ESS algorithm compared with releases through v0.9.4: matching occurs before aggregation, and geometric normalization replaces the old pairwise ESS average. Use a new output directory, even for default ESS. Keep the old release and old outputs for exact historical reproduction.

Output prefixes include .matched_ess or .matched_se. Each output has *.meta_audit.tsv.gz (per-SNP original valid information, matched information, fraction retained, and contributing-cohort counts) and *.meta_cohorts.tsv.gz (per-cohort valid/matched SNP counts). Information is 1/SE^2 in these audits even in ESS mode. The scalar output sample size is the sum of input cohort sample sizes, not a per-SNP observed sample size.

meta_config.json records the algorithm and weighting. A different mode or a legacy output without a manifest cannot be reused or overwritten in place; --skip only reuses compatible output. As before, use a new output directory when changing the source input files; the manifest is not an input-content hash.

Information and summary-statistics arithmetic use float64; LD accumulation uses float32 and .ld.npz storage remains float16. There is no automatic PSD repair. The existing file loader's NaN-to-zero LD policy is unchanged: matching an LD map entry does not establish that every pairwise LD cell was observed.

credtools meta¶