Skip to contents

Combine evidence for an effect in every contrast

Source: R/find_consistent_hit.R
combine_consistent_hits.Rd

Tests whether each gene has an effect in a prespecified direction in all supplied contrasts, using an intersection-union test.

Usage

combine_consistent_hits(
  results,
  direction = c("Down", "Up"),
  fdr = 0.05,
  fdr_method = c("BH", "BY")
)

Arguments

results

A named list of at least two complete gene-level tables. Every table must contain the same unique, non-missing Gene identifiers and numeric PValue values between zero and one. PValue must be an unadjusted one-sided p-value for direction, including genes whose observed effect is opposite. Do not supply filtered hit lists or two-sided p-values. Optional Log2FC columns are retained.

direction

The common, prespecified direction, "Down" or "Up". This labels the hypothesis; it does not convert input p-values. The caller must ensure that all tables test this direction.

fdr

Maximum combined FDR to return (default 0.05); use 1 for all tested genes.

fdr_method

Multiple-testing correction across genes: "BH" (default, independence or suitable positive dependence) or "BY" (arbitrary dependence).

Value

A tibble ordered by combined PValue, with Gene, Direction (the tested direction), NContrasts, PValue, FDR, PValue_<contrast> and, when supplied, Log2FC_<contrast> columns. No pooled effect is computed.

Details

For each gene the null is that at least one contrast has no effect in the requested direction. The alternative is an effect in that direction in every contrast. The combined p-value is the maximum of the individual raw p-values. This combination allows arbitrary dependence between contrasts, provided the individual p-values are valid. It does not repair a misspecified model, an inappropriate reference distribution, or ignored repeated measurements.

FDR adjustment is applied to the combined p-values over all supplied genes, before filtering. Input FDR columns are ignored. Missing genes, missing p-values and duplicated identifiers are errors: a missing test cannot be silently omitted from an all-contrasts hypothesis. The function cannot detect tables filtered to the same subset of genes; supplying the complete testing universe is the caller's responsibility.

Consistency here means an effect in every specified contrast, not equal effect sizes, a time interaction, or independent replication.

References

Heller R, Golland Y, Malach R, Benjamini Y (2007). Conjunction group analysis: an alternative to mixed/random effect analysis. NeuroImage 37:1178-1185. https://pubmed.ncbi.nlm.nih.gov/17689266/

See also

find_consistent_hit, find_time_hit

Examples

t3 <- data.frame(Gene = c("A", "B"), PValue = c(0.001, 0.01))
t4 <- data.frame(Gene = c("B", "A"), PValue = c(0.8, 0.002))
combine_consistent_hits(list(T3 = t3, T4 = t4))
#> # A tibble: 1 × 7
#>   Gene  Direction NContrasts PValue   FDR PValue_T3 PValue_T4
#>   <chr> <chr>          <int>  <dbl> <dbl>     <dbl>     <dbl>
#> 1 A     Down               2  0.002 0.004     0.001     0.002