- 🔥 IHW (Independent Hypothesis Weighting) 🔥
- 💡Idea: Method uses a data-driven approach to calculate weights using independent covariates and then applies a group-weighted BH method
- Features Size investing
-
Assumptions:
- covariate must be independent of p-value under null
- tests are independent under null
- 👍Good: Improves power and controls FDR at alpha level
- 👎Bad: does not maintain FDR when (1) insufficient power to detect false hypotheses (2) all hypotheses are true
- 🔥 ASH (Adaptive SHrinkage) 🔥
- 💡Idea:
- 👍Good:
- 👎Bad:
- 🔥 Boca-Leek 🔥
- 💡Idea:
- Estimates FDR conditional on covariates in a multiple testing framework by (1) estimating the null proportion of hypotheses with logistic regression and (2) multiplying the BH-adjusted p-values by the estimated null proportion.
- Assumes that the p-values are independent of the covariates conditional on the null or alternative. In other words, the probability that a test statistic or p-value is drawn from the null or non-null distribution depends on the covariates (but not the realized value of the test statistic or p-value itself).
- 👍Good:
- FDR control can be maintained even when tests are moderately correlated
- If hypotheses are independent, can use bootstrapping to obtain CIs for the null proportion of hypotheses
- Increasing the number of tests can lead to improvement in FDR control
- 👎Bad:
- Improvement over Storey (2002) is minor
- Scott (2015) is superior when test statistics are normally distributed
- Does not control FDR when hypotheses are highly correlated
- Requires tuning of
$\lambda$ parameter and careful specification of covariates (smoothing splines? categorization?)
- 💡Idea:
-
Benjamini and Hochberg (BH)
- 💡Idea: Uses p-values. Assumes independence of tests (or assumes all hypotheses are exchangeable), but has been shown to be robust even with correlated tests.
- 👍Good: Easy to use. More powerful than FWER-based methods
- 👎Bad: Sub-optimal power (and can't prioritize tests) when the individual tests differ in statistical properties such as sample size, true effect size, signal-to-noise ratio or prior probability of being false (see Ignatiadis et al. 2016)
- To increase power while controlling FDR, can use independent covariates to prioritize tests
-
Weighted BH (basis of IHW)
- 💡Idea: Associate each test with a non-negative weight such that weights averahe to 1. Hypotheses with higher weights are prioritized.
- 👍Good: Controls FDR
- 👎Bad: Weights must be prespecified, are independent of data
-
Greedy Independent Filtering (compared in IHW paper)
- 💡Idea: Filter all p-values using an independent covariate (
$X$ ) such that$X$ is less than some threshold$x$ . Greedy if researcher tests all possible thresholds and picks one that maximizes number of discoveries. However, greedy approaches do not control for FDR at alpha level. - 👍Good: Controls FDR at alpha level IF researcher ONLY uses a pre-specified threshold
- 👎Bad: Not automated, very subjective, can lead to p-hacking
- 💡Idea: Filter all p-values using an independent covariate (
-
LSL-GBH, TST-GBH (compared in IHW paper)
- 💡Idea:
- 👍Good:
- 👎Bad:
-
SBH (compared in IHW paper)
- 💡Idea: Use a categorial (or binned continuous) covariate to stratify tests, apply BH within each strata, combine significant tests
- 👍Good:
- 👎Bad: Loss of FDR control for the null tests because different strata are treated equally
- To increase power, different strata should be prioritized differently
-
Clfdr (Cai's local FDR) (compared in IHW paper)
- 💡Idea:
- 👍Good:
- 👎Bad:
-
FDRreg (compared in IHW paper, swfdr paper)
- 💡Idea:
- 👍Good:
- 👎Bad:
-
local false discovery rate (compared in ASH paper)
- 💡Idea: local FDR is estimated separately within each group and then estimates are grouped together
- 👍Good:
- 👎Bad: FDR isn't controlled at alpha level if alternative distributions across groups are different
-
mixture false discovery rate (compared in ASH paper)
- 💡Idea:
- 👍Good:
- 👎Bad: