-- | -- Module : Statistics.Test.NonParametric -- Copyright : (c) 2010 Neil Brown -- License : BSD3 -- -- Maintainer : bos@serpentine.com -- Stability : experimental -- Portability : portable -- -- Functions for performing non-parametric tests (i.e. tests without an assumption -- of underlying distribution). module Statistics.Test.NonParametric (-- * Mann-Whitney U test (non-parametric equivalent to the independent t-test) mannWhitneyU, mannWhitneyUCriticalValue, mannWhitneyUSignificant, -- * Wilcoxon signed-rank matched-pair test (non-parametric equivalent to the paired t-test) wilcoxonMatchedPairSignedRank, wilcoxonMatchedPairSignificant, wilcoxonMatchedPairSignificance, wilcoxonMatchedPairCriticalValue, -- * Wilcoxon rank sum test wilcoxonRankSums) where import Control.Applicative ((<$>)) import Control.Arrow ((***)) import Data.Function (on) import Data.List (findIndex, groupBy, partition, sortBy) import Data.Ord (comparing) import qualified Data.Vector.Unboxed as U (length, toList, zipWith) import Statistics.Distribution (quantile) import Statistics.Distribution.Normal (standard) import Statistics.Math (choose) import Statistics.Types (Sample) -- | The Wilcoxon Rank Sums Test. -- -- This test calculates the sum of ranks for the given two samples. The samples -- are ordered, and assigned ranks (ties are given their average rank), then these -- ranks are summed for each sample. -- -- The return value is (W_1, W_2) where W_1 is the sum of ranks of the first sample -- and W_2 is the sum of ranks of the second sample. This test is trivially transformed -- into the Mann-Whitney U test. You will probably want to use 'mannWhitneyU' -- and the related functions for testing significance, but this function is exposed -- for completeness. wilcoxonRankSums :: Sample -> Sample -> (Double, Double) wilcoxonRankSums xs1 xs2 = ((sum . map fst) *** (sum . map fst)) . -- sum the ranks per group partition snd . -- split them back into left and right concatMap mergeRanks . -- merge the ranks of duplicates groupBy ((==) `on` (snd . snd)) . -- group duplicate values zip [1..] . -- give them ranks (duplicates receive different ranks here) sortBy (comparing snd) $ -- sort by their values zip (repeat True) (U.toList xs1) ++ zip (repeat False) (U.toList xs2) -- Tag each sample with an identifier before we merge them where mergeRanks :: [(AbsoluteRank, (Bool, Double))] -> [(AbsoluteRank, Bool)] mergeRanks xs = zip (repeat rank) (map (fst . snd) xs) where -- Ranks are merged by assigning them all the average of their ranks: rank = sum (map fst xs) / fromIntegral (length xs) -- | The Mann-Whitney U Test. -- -- This is sometimes known as the Mann-Whitney-Wilcoxon U test, and -- confusingly many sources state that the Mann-Whitney U test is the same as -- the Wilcoxon's rank sum test (which is provided as 'wilcoxonRankSums'). -- The Mann-Whitney U is a simple transform of Wilcoxon's rank sum test. -- -- Again confusingly, different sources state reversed definitions for U_1 and U_2, -- so it is worth being explicit about what this function returns. Given two samples, -- the first, xs_1, of size n_1 and the second, xs_2, of size n_2, this function -- returns (U_1, U_2) where U_1 = W_1 - (n_1*(n_1+1))\/2 and U_2 = W_2 - (n_2*(n_2+1))\/2, -- where (W_1, W_2) is the return value of @wilcoxonRankSums xs1 xs2@. -- -- Some sources instead state that U_1 and U_2 should be the other way round, often -- expressing this using U_1' = n_1*n_2 - U_1 (since U_1 + U_2 = n_1*n*2). -- -- All of which you probably don't care about if you just feed this into 'mannWhitneyUSignificant'. mannWhitneyU :: Sample -> Sample -> (Double, Double) mannWhitneyU xs1 xs2 = (fst summedRanks - (n1*(n1 + 1))/2 ,snd summedRanks - (n2*(n2 + 1))/2) where n1 = fromIntegral $ U.length xs1 n2 = fromIntegral $ U.length xs2 summedRanks = wilcoxonRankSums xs1 xs2 -- | Calculates the critical value of Mann-Whitney U for the given sample -- sizes and significance level. -- -- This function returns the exact calculated value of U for all sample sizes; -- it does not use the normal approximation at all. Above sample size 20 it is -- generally recommended to use the normal approximation instead, but this function -- will calculate the higher critical values if you need them. -- -- The algorithm to generate these values is a faster, memoised version of the -- simple unoptimised generating function given in section 2 of \"The Mann Whitney -- Wilcoxon Distribution Using Linked Lists\", Cheung and Klotz, Statistica Sinica -- 7 (1997), . mannWhitneyUCriticalValue :: (Int, Int) -- ^ The sample size -> Double -- ^ The p-value (e.g. 0.05) for which you want the critical value. -> Maybe Int -- ^ The critical value (of U). mannWhitneyUCriticalValue (m, n) p | p' <= 1 = Nothing | m < 1 || n < 1 = Nothing | otherwise = findIndex (>= p') $ let firstHalf = map fromIntegral $ take (((m*n)+1)`div`2) $ tail $ alookup !! (m+n-2) !! (min m n - 1) {- Original: [fromIntegral $ a k (m+n) (min m n) | k <- [1..m*n]] -} secondHalf | even (m*n) = reverse firstHalf | otherwise = tail $ reverse firstHalf in firstHalf ++ map (mnCn -) secondHalf where mnCn = (m+n) `choose` n p' = mnCn * p {- Original function, without memoisation, from Cheung and Klotz: a :: Int -> Int -> Int -> Int a u bigN m | u < 0 = 0 | u >= (m * smalln) = floor $ fromIntegral bigN `choose` fromIntegral m | m == 1 || smalln == 1 = u + 1 | otherwise = a u (bigN - 1) m + a (u - smalln) (bigN - 1) (m-1) where smalln = bigN - m -} -- Memoised version of the original a function, above. -- -- outer list is indexed by big N - 2 -- inner list by m (we know m < bigN) -- innermost list by u -- -- So: (alookup ! (bigN - 2) ! m ! u) == a u bigN m alookup :: [[[Int]]] alookup = gen 2 [1 : repeat 2] where gen bigN predBigNList = let bigNlist = [ let limit = round $ fromIntegral bigN `choose` fromIntegral m in [amemoed u m | u <- [0..m*(bigN-m)]] ++ repeat limit | m <- [1..(bigN-1)]] -- has bigN-1 elements in bigNlist : gen (bigN+1) bigNlist where amemoed :: Int -> Int -> Int amemoed u m | m == 1 || smalln == 1 = u + 1 | otherwise = let (predmList : mList : _) = drop (m-2) predBigNList -- m-2 because starts at 1 -- We know that predBigNList has bigN - 2 elements -- (and we know that smalln > 1 therefore bigN > m + 1) -- So bigN - 2 >= m, i.e. predBigNList must have at least m elements -- elements, so dropping (m-2) must leave at least 2 in (mList !! u) + (if u < smalln then 0 else predmList !! (u - smalln)) where smalln = bigN - m -- | Calculates whether the Mann Whitney U test is significant. -- -- If both sample sizes are less than or equal to 20, the exact U critical value -- (as calculated by 'mannWhitneyUCriticalValue') is used. If either sample is -- larger than 20, the normal approximation is used instead. -- -- If you use a one-tailed test, the test indicates whether the first sample is -- significantly larger than the second. If you want the opposite, simply reverse -- the order in both the sample size and the (U_1, U_2) pairs. mannWhitneyUSignificant :: Bool -- ^ Perform one-tailed test (see description above). -> (Int, Int) -- ^ The sample size from which the (U_1,U_2) values were derived. -> Double -- ^ The p-value at which to test (e.g. 0.05) -> (Double, Double) -- ^ The (U_1, U_2) values from 'mannWhitneyU'. -> Maybe Bool -- ^ Just True if the test is significant, Just -- False if it is not, and Nothing if the sample -- was too small to make a decision. mannWhitneyUSignificant oneTail (in1, in2) p (u1, u2) | in1 > 20 || in2 > 20 --Use normal approximation -- = (n1*(n1+1))/2 - u1 - (n1*(n1+n2))/2 -- = (n1*(n1+1))/2 - (-2*u1 + n1*(n1+n2))/2 -- = (n1*(n1+1) - 2*u1 + n1*(n1+n2))/2 -- = (n1*(2*n1 + n2 + 1) - 2*u1)/2 = let num = (n1*(2*n1 + n2 + 1)) / 2 - u1 denom = sqrt $ n1*n2*(n1 + n2 + 1) / 12 z = num / denom zcrit = quantile standard (1 - if oneTail then p else p/2) in Just $ (if oneTail then z else abs z) > zcrit | otherwise = do crit <- fromIntegral <$> mannWhitneyUCriticalValue (in1, in2) p return $ if oneTail then u2 <= crit else min u1 u2 <= crit where n1 = fromIntegral in1 n2 = fromIntegral in2 -- | The Wilcoxon matched-pairs signed-rank test. -- -- The value returned is the pair (T+, T-). T+ is the sum of positive ranks (the -- ranks of the differences where the first parameter is higher) whereas T- is -- the sum of negative ranks (the ranks of the differences where the second parameter is higher). -- These values mean little by themselves, and should be combined with the 'wilcoxonSignificant' -- function in this module to get a meaningful result. -- -- The samples are zipped together: if one is longer than the other, both are truncated -- to the the length of the shorter sample. -- -- Note that: wilcoxonMatchedPairSignedRank == (\(x, y) -> (y, x)) . flip wilcoxonMatchedPairSignedRank wilcoxonMatchedPairSignedRank :: Sample -> Sample -> (Double, Double) wilcoxonMatchedPairSignedRank a b -- Best to read this function bottom to top: = (sum *** sum) . -- Sum the positive and negative ranks separately. partition (> 0) . -- Split the ranks into positive and negative. None of the -- ranks can be zero. concatMap mergeRanks . -- Then merge the ranks for any duplicates by taking -- the average of the ranks, and also make the rank -- into a signed rank groupBy ((==) `on` abs . snd) . -- Now group any duplicates together -- Note: duplicate means same absolute difference zip [1..] . -- Add a rank (note: at this stage, duplicates will get different ranks) dropWhile (== 0) . -- Remove any differences that are zero (i.e. ties in the -- original data). We know they must be at the head of -- the list because we just sorted it, so dropWhile not filter sortBy (comparing abs) . -- Sort the differences by absolute difference U.toList $ -- Convert to a list (could be done later in the pipeline?) U.zipWith (-) a b -- Work out differences where mergeRanks :: [(AbsoluteRank, Double)] -> [SignedRank] mergeRanks xs = map ((* rank) . signum . snd) xs -- Note that signum above will always be 1 or -1; any zero differences will -- have been removed before this function is called. where -- Ranks are merged by assigning them all the average of their ranks: rank = sum (map fst xs) / fromIntegral (length xs) type AbsoluteRank = Double type SignedRank = Double -- | The coefficients for x^0, x^1, x^2, etc, in the expression -- \prod_{r=1}^s (1 + x^r). See the Mitic paper for details. -- -- We can define: -- f(1) = 1 + x -- f(r) = (1 + x^r)*f(r-1) -- = f(r-1) + x^r * f(r-1) -- The effect of multiplying the equation by x^r is to shift -- all the coefficients by r down the list. -- -- This list will be processed lazily from the head. coefficients :: Int -> [Int] coefficients 1 = [1, 1] -- 1 + x coefficients r = let coeffs = coefficients (r-1) (firstR, rest) = splitAt r coeffs in firstR ++ add rest coeffs where add (x:xs) (y:ys) = x + y : add xs ys add xs [] = xs add [] ys = ys -- This list will be processed lazily from the head. summedCoefficients :: Int -> [Double] summedCoefficients = map fromIntegral . scanl1 (+) . coefficients -- | Tests whether a given result from a Wilcoxon signed-rank matched-pairs test -- is significant at the given level. -- -- This function can perform a one-tailed or two-tailed test. If the first -- parameter to this function is False, the test is performed two-tailed to -- check if the two samples differ significantly. If the first parameter is -- True, the check is performed one-tailed to decide whether the first sample -- (i.e. the first sample you passed to 'wilcoxonMatchedPairSignedRank') is -- greater than the second sample (i.e. the second sample you passed to -- 'wilcoxonMatchedPairSignedRank'). If you wish to perform a one-tailed test -- in the opposite direction, you can either pass the parameters in a different -- order to 'wilcoxonMatchedPairSignedRank', or simply swap the values in the resulting -- pair before passing them to this function. wilcoxonMatchedPairSignificant :: Bool -- ^ Perform one-tailed test (see description above). -> Int -- ^ The sample size from which the (T+,T-) values were derived. -> Double -- ^ The p-value at which to test (e.g. 0.05) -> (Double, Double) -- ^ The (T+, T-) values from 'wilcoxonMatchedPairSignedRank'. -> Maybe Bool -- ^ Just True if the test is significant, Just -- False if it is not, and Nothing if the sample -- was too small to make a decision. wilcoxonMatchedPairSignificant oneTail sampleSize p (tPlus, tMinus) -- According to my nearest book (Understanding Research Methods and Statistics -- by Gary W. Heiman, p590), to check that the first sample is bigger you must -- use the absolute value of T- for a one-tailed check: | oneTail = ((abs tMinus <=) . fromIntegral) <$> wilcoxonMatchedPairCriticalValue sampleSize p -- Otherwise you must use the value of T+ and T- with the smallest absolute value: | otherwise = ((t <=) . fromIntegral) <$> wilcoxonMatchedPairCriticalValue sampleSize (p/2) where t = min (abs tPlus) (abs tMinus) -- | Obtains the critical value of T to compare against, given a sample size -- and a p-value (significance level). Your T value must be less than or -- equal to the return of this function in order for the test to work out -- significant. If there is a Nothing return, the sample size is too small to -- make a decision. -- -- 'wilcoxonSignificant' tests the return value of 'wilcoxonMatchedPairSignedRank' -- for you, so you should use 'wilcoxonSignificant' for determining test results. -- However, this function is useful, for example, for generating lookup tables -- for Wilcoxon signed rank critical values. -- -- The return values of this function are generated using the method detailed in -- the paper \"Critical Values for the Wilcoxon Signed Rank Statistic\", Peter -- Mitic, The Mathematica Journal, volume 6, issue 3, 1996, which can be found -- here: . -- According to that paper, the results may differ from other published lookup tables, but -- (Mitic claims) the values obtained by this function will be the correct ones. wilcoxonMatchedPairCriticalValue :: Int -- ^ The sample size -> Double -- ^ The p-value (e.g. 0.05) for which you want the critical value. -> Maybe Int -- ^ The critical value (of T), or Nothing if -- the sample is too small to make a decision. wilcoxonMatchedPairCriticalValue sampleSize p = case critical of Just n | n < 0 -> Nothing | otherwise -> Just n Nothing -> Just maxBound -- shouldn't happen: beyond end of list where m = (2 ** fromIntegral sampleSize) * p critical = subtract 1 <$> findIndex (> m) (summedCoefficients sampleSize) -- | Works out the significance level (p-value) of a T value, given a sample -- size and a T value from the Wilcoxon signed-rank matched-pairs test. -- -- See the notes on 'wilcoxonCriticalValue' for how this is calculated. wilcoxonMatchedPairSignificance :: Int -- ^ The sample size -> Double -- ^ The value of T for which you want the significance. -> Double -- ^^ The significance (p-value). wilcoxonMatchedPairSignificance sampleSize rank = (summedCoefficients sampleSize !! floor rank) / 2 ** fromIntegral sampleSize