Merge pull request #374 from fslaborg/repo-assist/improve-dunn-sidak-20260412-ff9a8aac-3956323be622a182

[Repo Assist] feat: add dunnSidakFWER and holmSidakFWER corrections to MultipleTesting

Author: annacwe

src/FSharp.Stats/Testing/MultipleTesting.fs

@@ -78,6 +78,52 @@ module MultipleTesting =
             result
 
 
+    /// Šidák single-step FWER adjustment (NaN-safe).
+    /// Each p-value is adjusted independently as p_adj = 1 - (1 - p)^m, where m is the number of
+    /// non-NaN tests. This is slightly less conservative than Bonferroni.
+    /// NaN p-values are preserved at their original indices.
+    let inline dunnSidakFWER (p : float[]) : float[] =
+        let total  = p.Length
+        let indexed = p |> Array.mapi (fun i v -> i, v)
+        let valid   = indexed |> Array.filter (fun (_, v) -> not (Double.IsNaN v))
+        let m       = valid.Length
+        let result  = Array.create total Double.NaN
+        if m = 0 then result else
+            valid |> Array.iter (fun (i, pv) ->
+                result.[i] <- min 1.0 (1.0 - (1.0 - pv) ** float m))
+            result
+
+    /// Holm–Šidák (step-down) FWER adjustment (NaN-safe).
+    /// Uses the Šidák formula at each step of the Holm procedure, giving a slightly more
+    /// powerful test than Holm–Bonferroni while still controlling the FWER.
+    /// NaN p-values are ignored in the computation and preserved at their original indices.
+    let inline holmSidakFWER (p : float[]) : float[] =
+        let total  = p.Length
+        let indexed = p |> Array.mapi (fun i v -> i, v)
+        let valid   = indexed |> Array.filter (fun (_, v) -> not (Double.IsNaN v))
+        let m       = valid.Length
+        let result  = Array.create total Double.NaN
+        if m = 0 then result else
+            let sorted = valid |> Array.sortBy snd
+
+            // raw[i] = 1 - (1 - p_(i+1))^(m-i) for i = 0..m-1  (rank in 1-indexed sorted order = i+1)
+            let raw =
+                sorted
+                |> Array.mapi (fun i (_, pv) -> 1.0 - (1.0 - pv) ** float (m - i))
+
+            // running max from the left, capped at 1.0
+            let adjAsc =
+                raw
+                |> Array.scan (fun runningMax r -> max runningMax r) 0.0
+                |> Array.tail
+                |> Array.map (min 1.0)
+
+            Array.zip (sorted |> Array.map fst) adjAsc
+            |> Array.iter (fun (i, adj) -> result.[i] <- adj)
+
+            result
+
+
     /// Benjamini-Hochberg Correction (BH)
     /// 'projection' should return a tuple of any identifier and the pValues as float, when applied to 'rawP'
     /// This function applies the Benjamini-Hochberg multiple testing correcture and returns all False Discovery Rates to which the given p-values are still 

tests/FSharp.Stats.Tests/Testing.fs

@@ -497,6 +497,50 @@ let hochbergTests =
 
     ]
 
+[<Tests>]
+let dunnSidakTests =
+    // p = [0.01; 0.04; 0.03; 0.1; 0.5], m=5
+    // expected values verified with Python: 1 - (1-p)^m
+    let pValues = [| 0.01; 0.04; 0.03; 0.1; 0.5 |]
+    let pNaN    = [| 0.01; nan;  0.03; 0.1; 0.5 |]
+
+    testList "Testing.MultipleTesting.DunnSidak" [
+        testCase "singleStepBasic" <| fun () ->
+            let result = MultipleTesting.dunnSidakFWER pValues
+            // 1 - (1-p)^5 for each p
+            let expected = [| 0.049010; 0.184627; 0.141266; 0.40951; 0.96875 |]
+            Array.iter2 (fun r e ->
+                Expect.floatClose Accuracy.low r e "Single-step Šidák adjusted p-values should match."
+            ) result expected
+        testCase "singleStepNaN" <| fun () ->
+            let result = MultipleTesting.dunnSidakFWER pNaN
+            // m=4 valid values; 1 - (1-p)^4
+            Expect.floatClose Accuracy.low result.[0] 0.039404 "p[0] with NaN should match."
+            Expect.isTrue (Double.IsNaN result.[1]) "NaN position should remain NaN."
+            Expect.floatClose Accuracy.low result.[2] 0.114707 "p[2] with NaN should match."
+            Expect.floatClose Accuracy.low result.[3] 0.34390  "p[3] with NaN should match."
+            Expect.floatClose Accuracy.low result.[4] 0.93750  "p[4] with NaN should match."
+        testCase "holmSidakBasic" <| fun () ->
+            let result = MultipleTesting.holmSidakFWER pValues
+            // sorted p: [0.01, 0.03, 0.04, 0.1, 0.5]
+            // raw: [1-0.99^5, 1-0.97^4, 1-0.96^3, 1-0.9^2, 1-0.5^1]
+            // running max (already monotone): [0.04901, 0.11471, 0.11526, 0.19, 0.5]
+            // back to original order: [0.04901, 0.11526, 0.11471, 0.19, 0.5]
+            let expected = [| 0.049010; 0.115264; 0.114707; 0.19; 0.5 |]
+            Array.iter2 (fun r e ->
+                Expect.floatClose Accuracy.low r e "Holm–Šidák adjusted p-values should match."
+            ) result expected
+        testCase "holmSidakNaN" <| fun () ->
+            let result = MultipleTesting.holmSidakFWER pNaN
+            // m=4 valid values; sorted: [0.01, 0.03, 0.1, 0.5]
+            // raw: [1-0.99^4, 1-0.97^3, 1-0.9^2, 1-0.5^1] = [0.039404, 0.087327, 0.19, 0.5]
+            Expect.floatClose Accuracy.low result.[0] 0.039404 "p[0] Holm-Šidák NaN"
+            Expect.isTrue (Double.IsNaN result.[1]) "NaN position should remain NaN."
+            Expect.floatClose Accuracy.low result.[2] 0.087327 "p[2] Holm-Šidák NaN"
+            Expect.floatClose Accuracy.low result.[3] 0.19000  "p[3] Holm-Šidák NaN"
+            Expect.floatClose Accuracy.low result.[4] 0.50000  "p[4] Holm-Šidák NaN"
+    ]
+
 [<Tests>]
 let benjaminiHochbergTests =