Statistical testing

Summary: this tutorial explains how to perform various statistical tests with FSharp.Stats.

Table of contents

• Test Statistics
  • T-Test
  • Anova
  • F-Test
  • H-Test
  • Friedman-Test
  • Wilcoxon signed-rank Test
  • Chi-Squared Test
  • Bartlett
• PostHoc
  • Fisher's LSD
  • Hays
  • Tukey HSD
  • Dunnetts test
  • Fisher Hotelling
• Multiple testing
  • Benjamini-Hochberg
  • Q Value
  • SAM

FSharp.Stats provides hypothesis tests for different applications. A hypothesis test is a statistical test that is used to determine whether there is enough evidence in a sample of data to infer that a certain condition is true for the entire population. A hypothesis test examines two opposing hypotheses about a population: the null hypothesis and the alternative hypothesis.

Test Statistics

T-Test

By using a t test a difference of means can be evaluated. There are different kinds of t test designs implemented in FSharp.Stats.

1. One sample t test:

   • Is the population mean equal to the value of H0?
   • e.g. “Is my grade different from the distribution mean?”

2. Two sample t test with equal variances:

   • Prerequisite: equal variances
   • Is the mean of Population A equal to the mean of Population B?

   • 2.1 unpaired t test:

     • e.g.: Does the cell size differ between wildtype and mutant?

   • 2.2 paired t test:

     • e.g.: Does the medication influence the blood pressure? Measurement of the same persons before and after medication.

3. Two sample t test with unequal variances

   • Welch test (unpaired)

Case 1: One sample t test

open FSharp.Stats
open FSharp.Stats.Testing

let sampleA = vector [|4.5; 5.1; 4.8; 4.4; 5.0|]

// calculates a one sample t test with a given sample and the fixed value the sample should be compared with
let oneSampleTTest = TTest.oneSample sampleA 5.

(*
    The test returns no significant p value:
    oneSampleTTest.PValue = 0.1533
*)

Case 2: Two sample t test with equal variances (unpaired)

A standard two sample t test expects the samples to be taken from populations with equal standard deviations. Violations of this requirement result in an inflated false positive rate.

let sample1 = vector [|4.9;5.0;6.7;4.8;5.2|]
let sample2 = vector [|3.9;4.9;3.8;4.5;4.5|]

let twoSampleTTest = TTest.twoSample true sample1 sample2

(*
    The test returns a significant p value:
    twoSampleTTest.PValue = 0.0396
*)

Case 3: Two sample t test with equal variances (paired)

A paired t-test is used to compare two population means where you have two samples in which observations in one sample can be paired with observations in the other sample. Examples of where this might occur are:

• Before-and-after observations on the same subjects (e.g. students’ diagnostic test results before and after a particular module or course).
• A comparison of two different methods of measurement or two different treatments where the measurements/treatments are applied to the same subjects (e.g. blood pressure measurements using a stethoscope and a dynamap).

let sampleP1 = vector [18.;21.;16.;22.;19.;24.;17.;21.;23.;18.;14.;16.;16.;19.;18.;20.;12.;22.;15.;17.;]
let sampleP2 = vector [22.;25.;17.;24.;16.;29.;20.;23.;19.;20.;15.;15.;18.;26.;18.;24.;18.;25.;19.;16.;]

let paired = TTest.twoSamplePaired sampleP1 sampleP2

(*
    The test returns a significant p value:
    paired.PValue = 0.00439
*)

Case 4: Two sample t test with unequal variances (Welch test)

If you are unsure about the nature of the underlying population, you may ask if the theoretical population distributions you want to compare do have the same standard deviations.

If not the welch test can serve as an appropriate hypothesis test for mean differences.

let sampleW1 = vector [0.8;0.9;1.0;1.1;1.2]
let sampleW2 = vector [0.8;1.1;1.3;1.5;2.0]

let welch = TTest.twoSample false sampleW1 sampleW2

(*
    The test returns a not significant p value:
    welch.PValue = 0.1725626595
*)

Anova

let dataOneWay =
    [|
    [|0.28551035; 0.338524035; 0.088313218; 0.205930807; 0.363240102;|];
    [|0.52173913; 0.763358779; 0.32546786; 0.425305688; 0.378071834; |];
    [|0.989119683; 1.192718142; 0.788288288; 0.549176236; 0.544588155;|];
    [|1.26705653; 1.625320787; 1.266108976; 1.154187629; 1.268498943; 1.069518717;|];
    |]

let contrastMatrix = 
    [| 
    [|1.0;-1.0;0.0;0.0;|]
    [|1.0;0.0;-1.0;0.0;|]
    [|1.0;0.0;0.0;-1.0;|]
    [|0.0;1.0;-1.0;0.0;|]
    [|0.0;1.0;0.0;-1.0;|]
    [|0.0;0.0;1.0;-1.0;|]
    |]

let oneWayResult = Anova.oneWayAnova dataOneWay

{ Factor = { DegreesOfFreedom = 3.0
             MeanSquares = 1.081456009
             Significance = 9.257958872e-07
             Source = BetweenGroups
             Statistic = 27.59425833
             SumOfSquares = 3.244368027 }
  Error = { DegreesOfFreedom = 17.0
            MeanSquares = 0.03919134177
            Significance = nan
            Source = WithinGroups
            Statistic = nan
            SumOfSquares = 0.6662528101 }
  Total = { DegreesOfFreedom = 20.0
            MeanSquares = 0.1955310419
            Significance = nan
            Source = Total
            Statistic = nan
            SumOfSquares = 3.910620837 } }

(*
anovaResult.Factor.Statistic = 27.594
The factor statistic indicates how much more variability there is between the the samples 
than within the samples.
anovaResult.Factor.PValue = 9.257958872e-07
A strong significant p value in the factor field indicates that one or more means differ from each other
*)
// http://astatsa.com/OneWay_Anova_with_TukeyHSD/
// https://www.wessa.net/rwasp_Two%20Factor%20ANOVA.wasp

let data =
    [
        (0.28, 'A', 'M');
        (0.95, 'A', 'M');
        (0.96, 'A', 'M');
        (0.97, 'A', 'M');
        (0.40, 'A', 'M');
        (0.18, 'A', 'M');
        (0.12, 'A', 'M');
        (0.62, 'A', 'M');
        (1.81, 'A', 'F');
        (1.51, 'A', 'F');
        (1.41, 'A', 'F');
        (1.39, 'A', 'F');
        (1.20, 'A', 'F');
        (1.55, 'A', 'F');
        (1.48, 'A', 'F');
        (1.25, 'A', 'F');
        (0.95, 'B', 'M');
        (1.33, 'B', 'M');
        (0.92, 'B', 'M');
        (0.85, 'B', 'M');
        (1.06, 'B', 'M');
        (0.69, 'B', 'M');
        (0.70, 'B', 'M');
        (0.79, 'B', 'M');
        (2.93, 'B', 'F');
        (3.24, 'B', 'F');
        (3.42, 'B', 'F');
        (2.79, 'B', 'F');
        (2.54, 'B', 'F');
        (3.28, 'B', 'F');
        (2.80, 'B', 'F');
        (3.40, 'B', 'F');
    ]
    //f1
    |> Seq.groupBy (fun (v,f1,f2) -> f1)
    //f2
    |> Seq.map (fun (k,vls) -> 
        vls 
        |> Seq.groupBy (fun (v,f1,f2) -> f2)
        |> Seq.map (fun (k,vls') -> vls' |> Seq.map (fun (v,f1,f2) -> v) |> Seq.toArray)
        |> Seq.toArray
        ) 
    |> Seq.toArray
    

Anova.twoWayANOVA Anova.TwoWayAnovaModel.Mixed data

// http://statweb.stanford.edu/~susan/courses/s141/exanova.pdf
// http://scistatcalc.blogspot.de/2013/11/two-factor-anova-test-calculator.html#

let data' =
    [|
      // super
        // cold super
      [|[|4.;5.;6.;5.;|];
        // warm super
        [|7.;9.;8.;12.;|];
        // hot super
        [|10.;12.;11.;9.; |]|];
      // best 
        // cold best
      [|[|6.;6.;4.;4.;|];
        // warm best
        [|13.;15.;12.;12.;|];
        // hot best
        [|12.;13.;10.;13.;|]|]
    |]

Anova.twoWayANOVA Anova.TwoWayAnovaModel.Mixed data'

F-Test

The F-test is a method to determine if the variances of two samples are homogeneous. Also, ANOVA (analysis of variance) is based on the F-test and is used for the comparison of more than two samples. Knowing if your variances are equal (H₀ is true) helps to decide which test should be performed next. To determine if your variances are in fact equal you can perform a F-test.

Prerequisites:

• population distributions are normally distributed
• samples are independent

Possible further tests:

• two sample t-test with equal variances
• two sample t-test with unequal variances (Welch-test)
• ...

Note that there is no information about the direction of variance difference (e.g. Zimmermann 2004). In this implemented test the larger variance is always the numerator, therefore the comparison to F_{df1,df2,1-(alpha/2)} is used for a two sided test.

Important note: The effectiveness of a preliminary test of variances is discussed. The F-test is extremely sensitive to normality-violations, and even if the samples follow a normal distribution, it often does not detect situations where a t-test should be avoided.

References:

• Jürgen Bortz & Christof Schuster, Statistik für Human- und Sozialwissenschaftler (2010) Chapter 8.6
• Markowski, Conditions for the Effectiveness of a Preliminary Test of Variance (1990)
• Shlomo S. Sawilowsky, The Probable Difference Between Two Means When σ₁²≠σ₂² (2002)
• Ronald Ley, F curves have two tails but the F test is a one-tailed two-tailed test (1979) + Reviewer comments
• Donald W. Zimmermann, A note on preliminary tests of equality of variances (2004)

F-Test from data:

let sampleFA = vector [|5.0; 6.0; 5.8; 5.7|] 
let sampleFB = vector [|3.5; 3.7; 4.0; 3.3; 3.6|]

// comparison of sample variances 
let fTestFromData = FTest.testVariances sampleFA sampleFB
(* 
    { Statistic = 2.823383085
    DegreesOfFreedom1 = 3.0
    DegreesOfFreedom2 = 4.0
    PValue = 0.1708599931 }
    Using a significance level of 0.05 the sample variances do differ significantly.

*F-Test from given parameters:*
*)

// sample properties are given as (variance,degree of freedom) 
let sampleF1 = (0.1, 15.)
let sampleF2 = (0.05, 7.)

// comparison of sample variances 
let fTestFromParameters = FTest.testVariancesFromVarAndDof sampleF1 sampleF2

{ Statistic = 2.0
  DegreesOfFreedom1 = 15.0
  DegreesOfFreedom2 = 7.0
  PValue = 0.17963663
  PValueTwoTailed = 0.35927326 }

Using a significance level of 0.05 the sample variances do differ significantly.

H-Test

The H-test is also known as Kruskal-Wallis one-way analysis-of-variance-by-ranks and is the non-parametric equivalent of one-way ANOVA. It is a non-parametric test for comparing the means of more than two independent samples (equal or different sample size), and therefore is an extension of Wilcoxon-Mann-Whitney two sample test. Testing with H test gives information whether the samples are from the same distribution.

A benefit of the H-test is that it does not require normal distribution of the samples. The downside is that there is no information which samples are different from each other, or how many differences occur. For further investigation a post hoc test is recommended.

The distribution of the H-test statistic is approximated by chi-square distribution with degrees of freedom = sample count - 1.

The implemented H-test is testing for duplicate values in the data. Duplicates lead to ties in the ranking, and are corrected by using a correction factor.

Prerequisites:

• random and independent samples
• observations are from populations with same shape of distribution
• nominal scale, ordinal scale, ratio scale or interval scale data

References:

• E. Ostertagová, Methodology and Application of the Kruskal-Wallis Test (2014)
• Y. Chan, RP Walmsley, Learning and understanding the Kruskal-Wallis one-way analysis-of-variance-by-ranks test for differences among three or more independent groups (1997)

let groupA = [44.;44.;54.;32.;21.;28.]
let groupB = [70.;77.;48.;64.;71.;75.]
let groupC = [80.;76.;34.;80.;73.;80.] 
let samples = [groupA;groupB;groupC]

// calculation of the H-test 
let hResult = HTest.createHTest samples 

{ Statistic = 9.781466113
  DegreesOfFreedom = 2.0
  PValueLeft = 0.9924840891
  PValueRight = 0.007515910866
  PValue = 0.01503182173 }

PValueRight is significant at a alpha level of 0.05

A suitable post hoc test for H-tests is Dunn's test.

Friedman-Test

The Friedman-Test is a nonparametric method to detect differences in more than two related samples. It's used for dependent samples, e.g. with measurements of different treatments or at different timepoints. The test can be performed even on small sample sizes. One benefit of the implemented version in FSharp.Stats is that a correction factor is applied if there are any ties in the measurements.

Requirements:

• sample sizes must be identical
• repeated measurements
• no need for normal distribution of the samples
• samples must be at least ordinal scaled

The idea of the test is to rank each value in every ID (as seen below). The test statistic approximately follows a chi-squared distribution (dof = k-1). The recommended Post-Hoc-Test is Wilcoxon-signed-rank-test or a Bonferroni-correction. Example taken from Universitaet Zuerich - Methodenberatung Friedman-Test ( https://www.methodenberatung.uzh.ch/de/datenanalyse_spss/unterschiede/zentral/friedman.html ) with a modification for 3 ties in one ID.

References:

• Viv Bewick, Liz Cheek & Jonathan Ball, Statistics review 10: Further nonparametric methods (2004)
• Salvador García, Alberto Fernández, Julián Luengo, Francisco Herrera, Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power (2010)

ID   |  pre   | month 1| month 2| month 3| month 4
1       275     273      288      273      273

2       292     283      284      285      329

3       281     274      298      270      252

4       284     275      271      272      258

5       285     294      307      278      275
6       283     279      301      276      279
7       290     265      298      291      295
8       294     277      295      290      271
9       300     304      293      279      271
10      284     297      284      292      284

Ranking the results - average if values appear multiple times in one ID

ID   |  pre   | month 1| month 2| month 3| month 4
1       4       2        5        2        2      
2       4       1        2        3        5
3       4       3        5        2        1          
4       5       4        2        3        1          
5       3       4        5        2        1
6       4       2.5      5        1        2.5
7       2       1        5        3        4
8       4       2        5        3        1
9       4       5        3        2        1
10      2       5        2        4        2     
rank-sums
        36      29.5     39       25       20.5

// The data have to be entered in this format: 

let A = [|275.;273.;288.;273.;273.|]
let B = [|292.;283.;284.;285.;329.|]
let C = [|281.;274.;298.;270.;252.|]
let D = [|284.;275.;271.;272.;258.|]
let E = [|285.;294.;307.;278.;275.|]
let F = [|283.;279.;301.;276.;279.|]
let G = [|290.;265.;298.;291.;295.|]
let H = [|294.;277.;295.;290.;271.|]
let I = [|300.;304.;293.;279.;271.|]
let J = [|284.;297.;284.;292.;284.|]

// add everything in one sequence
let samplesMany =  [|A;B;C;D;E;F;G;H;I;J|]

// create the test 
let friedmanResult = FriedmanTest.createFriedmanTest samplesMany 

// results 

{ Statistic = 9.738219895
  DegreesOfFreedom = 4.0
  PValueLeft = 0.9549240734
  PValueRight = 0.04507592663
  PValue = 0.09015185326 }

Wilcoxon-Test

The Wilcoxon signed-rank test is a non-parametric alternative to the paired t-test. The test is used to determine if two samples stem from a similar distributed population. It calculates the differences between sets of pairs and ranks them to analyze if they are statistically significantly different. If at least one of the assumptions for the t-test is not given the Wilcoxon signed-rank test can be used. This implemented test uses a normal distribution approximation (validated against SciPy with mode="approx"). It also includes automatic tie correction as well as optional continuity correction.

Cases where Wilcoxon signed-rank test can be used:

• Compare 2 dependent samples
• Repeated measurements on single samples
• Matched samples

Assumptions for the test:

• Data are paired and from the same population
• Each pair is chosen randomly and independently
• At least interval scaled data

let before = [78.;24.;64.;45.;64.;52.;30.;50.;64.;50.;78.;22.;84.;40.;90.;72.]
let after  = [78.;24.;62.;48.;68.;56.;25.;44.;56.;40.;68.;36.;68.;20.;58.;32.]

// create the test
// Results with continuity correction: 
let wilcoxonResult = WilcoxonTest.createWilcoxonTest before after true 
// results 

{ Statistic = 2.072645167
  PValueLeft = 0.019102592
  PValueRight = 0.980897408
  PValueTwoTailed = 0.038205184 }

// Results without continuity correction: 
let wilcoxonResult2 = WilcoxonTest.createWilcoxonTest before after false 
// results 

{ Statistic = 2.104048882
  PValueLeft = 0.01768702961
  PValueRight = 0.9823129704
  PValueTwoTailed = 0.03537405922 }

//alternatively you can use just the differences 
let differences = [0.;0.;2.;-3.;-4.;-4.;5.;6.;8.;10.;10.;-14.;16.;20.;32.;40.]
// results from differences
let wilcoxonResult3 = WilcoxonTest.createWilcoxonTestFromDifferences differences true 
// results 

{ Statistic = 2.072645167
  PValueLeft = 0.019102592
  PValueRight = 0.980897408
  PValueTwoTailed = 0.038205184 }

References:

• Concepts & Applications of Inferential Statistics - VassarStats
• Wilcoxon Test, Adam Hayes, 13 April 2020, accessed 20.09.2021

Chi-Squared Test

WIP

Bartlett

WIP

PostHoc

ANOVA provides the user with a global statement if samples differ from each other. It does not provide detailed information regarding differences of the single samples.

If the H₀ hypothesis is neglected (so significant differences are assumed), a post hoc test (multiple comparison test) allows the pairwise comparison of the individual groups.

Reference: What is the proper way to apply the multiple comparison test?, Sangseok Lee and Dong Kyu Lee, 2018

Fisher's LSD

The simplest method is Fisher's least significant difference (Fisher's LSD). It calculates Student's t tests for all pairwise comparisons. But instead of estimating the variance for each sample separately it takes all groups into account. Violations of the homogeneity of variances reduce the test power. Since no correction for multiple comparisons is performed, the resulting p values must be corrected (for example with Benjamini-Hochberg method).

Important: Fishers LSD is dependent to a significant ANOVA (ANOVA-protected post-hoc test).

open PostHoc

let lsd = Testing.PostHoc.fishersLSD contrastMatrix dataOneWay

// For multi comparison correction, the p values are adjusted by the Benjamini-Hochberg approach
let (index,pValue,pValueAdj) = 
    lsd
    |> Testing.MultipleTesting.benjaminiHochbergFDRBy (fun x -> x,x.Significance)
    |> List.sortBy (fun (x,_) -> x.Index)
    |> List.map (fun (x,pValAdj) -> x.Index, x.Significance, pValAdj)
    |> List.unzip3

let lsdCorrected =
    let header = ["<b>Contrast index</b>";"<b>p Value</b>";"<b>p Value adj</b>"]
    let rows = 
        [
            [sprintf "%i" index.[0];sprintf "%.6f" pValue.[0];sprintf "%.6f" pValueAdj.[0];]    
            [sprintf "%i" index.[1];sprintf "%.6f" pValue.[1];sprintf "%.6f" pValueAdj.[1];]    
            [sprintf "%i" index.[2];sprintf "%.6f" pValue.[2];sprintf "%.6f" pValueAdj.[2];]       
            [sprintf "%i" index.[3];sprintf "%.6f" pValue.[3];sprintf "%.6f" pValueAdj.[3];]       
            [sprintf "%i" index.[4];sprintf "%.6f" pValue.[4];sprintf "%.6f" pValueAdj.[4];]    
            [sprintf "%i" index.[5];sprintf "%.6f" pValue.[5];sprintf "%.6f" pValueAdj.[5];]
        ]
    
    Chart.Table(
        header, 
        rows,
        HeaderFillColor = Color.fromHex "#45546a",
        CellsFillColor = Color.fromColors [Color.fromHex "#deebf7";Color.fromString "lightgrey"]
        )

Hays

Testing.PostHoc.hays contrastMatrix dataOneWay 


//let m1 = Seq.mean dataOneWay.[0]
//let m2 = Seq.mean dataOneWay.[1]

//let s1 = Seq.sum dataOneWay.[0]
//let s2 = Seq.sum dataOneWay.[1]
//let s3 = Seq.sum dataOneWay.[2]


//let d = (m1-m2)**2.0
//d / (3.926003843 * (1./5. + 1./5.))

// http://www.statisticslectures.com/topics/posthoconewayanova/
let dmg = 
    [|
     [|9.;8.;7.;8.;8.;9.;8.;|];
     [|7.;6.;6.;7.;8.;7.;6.;|];
     [|4.;3.;2.;3.;4.;3.;2.;|] ; 
    |]

let contrastMatrixDmg = 
    [| 
     [|1.0;-1.0;0.0;|] ; 
     [|1.0;0.0;-1.0;|] ; 
     [|0.0;1.0;-1.0;|]   
    |]


Testing.PostHoc.hays contrastMatrixDmg dmg 

Anova.oneWayAnova dmg

Tukey HSD

Tukeys honestly significant difference (HSD) can be used to inspect a significant ANOVA result for underlying causes.

Important note: Various discussions question the application of Tukeys HSD only to significant ANOVA results (Anova-protected post-hoc test), since Tukeys HSD already controls for multiple testing. Inform yourself prior to using an post hoc test appropriate to your experimental design. Fishers LSD however is dependent to a significant ANOVA.

Using this post hoc test you can determine which of the means differ significantly from each other. In the classis Tukeys HSD approach, the population variances are pooled for a more robust estimation (especially with small sample sizes). If the population variances differ, Tukeys HSD is not appropriate.

The implemented Tukey-Kramer-Method can be applied on unequal sample sizes and estimates the variance based on the means to compare. The Tukey-Kramer method can be used as stand-alone method for comparing multiple means.

A comparison of ANOVA and Tukey-Kramer-HSD with simulations for the robustness of normality-violations of the data can be found in:

Robustness of the ANOVA and Tukey-Kramer Statistical Tests, Wade C. Driscoll, Computers ind. Engng Vol 31, No. 1/2, pp. 265 - 268, 1996

Multiple-to-one comparisons can alternatively performed with Dunnet's test, which was designed for performing k-1 tests, while Tukeys HSD performs k((k-1)/2) tests.

Task: It should be tested if one or more means of samples taken from different mutants differ from each other.

// Example values from: https://brownmath.com/stat/anova1.htm
let hsdExample = 
    [|
        [|64.; 72.; 68.; 77.; 56.; 95.;|] // sample of mutant/drug/factor 1 
        [|78.; 91.; 97.; 82.; 85.; 77.;|] // sample of mutant/drug/factor 2 
        [|75.; 93.; 78.; 71.; 63.; 76.;|] // sample of mutant/drug/factor 3 
        [|55.; 66.; 49.; 64.; 70.; 68.;|] // sample of mutant/drug/factor 4    
    |]

let anovaResult = Anova.oneWayAnova hsdExample

(*
    anovaResult.Factor.Statistic = 5.41
    The factor statistic indicates how much more variability there is between the the samples 
    than within the samples.
    anovaResult.Factor.PValue = 0.0069
    A significant p value in the factor field indicates that one or more means differ from each other
*)

For tukey HSD calculations you need a contrast matrix, that defines the groups you want to compare for more detailed information.

Every contrast has as many entries as there are groups (samples). The groups, that should be compared are labelled with -1 or 1 respectively.

// In this contrast matrix every possible scenario is covered.
let contrastMatrixHSD = 
    [| 
    [|1.;-1.; 0.; 0.;|] // sample 1 is compared to sample 2
    [|1.; 0.;-1.; 0.;|] // sample 1 is compared to sample 3
    [|1.; 0.; 0.;-1.;|] 
    [|0.; 1.;-1.; 0.;|]
    [|0.; 1.; 0.;-1.;|]
    [|0.; 0.; 1.;-1.;|] // sample 3 is compared to sample 4
    |]

let hsdResult = tukeyHSD contrastMatrixHSD hsdExample 

"index	meanDiff	qStatistic	pValue
    0	13.000000	3.170103	0.146193
    1	4.000000	0.975416	0.899806
    2	10.000000	2.438541	0.337815
    3	9.000000	2.194687	0.427072
    4	23.000000	5.608644	0.003906
    5	14.000000	3.413957	0.106557"

(*
    For every generated contrast an output p value is calculated.
    e.g.
    hsdResult.[0].Significance = 0.0364
    hsdResult.[1].Significance = 0.4983 
    hsdResult.[2].Significance = 0.1001
    hsdResult.[3].Significance = 0.1364
    hsdResult.[4].Significance = 0.0008
    hsdResult.[5].Significance = 0.0255
*)

//TTest.twoSample true (vector hsdExample.[0]) (vector hsdExample.[1])
//TTest.twoSample true (vector hsdExample.[0]) (vector hsdExample.[2])
//TTest.twoSample true (vector hsdExample.[0]) (vector hsdExample.[3])
//TTest.twoSample true (vector hsdExample.[1]) (vector hsdExample.[2])
//TTest.twoSample true (vector hsdExample.[1]) (vector hsdExample.[3])
//TTest.twoSample true (vector hsdExample.[2]) (vector hsdExample.[3])

Dunnett's (t) test

When there is one control group which should be compared with all treatment-groups, Tukeys HSD would lead to an explosion of comparisons if the number of conditions increases. If just the comparison of each treatment to an control is required you can use Dunnett's test. It is a multiple-to-one post hoc test for homoscedastic samples with equal variance that has a higher power than Tukey's HSD since fewer comparisons have to be performed, and therefore the Confidence limits are wider than necessary. "ANOVA is not a necessary part of the multiple comparisons procedure" (Dunnett, 1964).

Note: Dunnett's test is designed for equal group sizes and will only provide approximate values when group sizes differ (Dunnett 1955).

Reference:

• A Multiple Comparison Procedure for Comparing Several Treatments with a Control; CW Dunnett; Journal of the American Statistical Association; Dec. 1955
• New Tables for Multiple Comparisons with a Control; CW Dunnett; Biometrics; Sep. 1964

let dunnetExample = 
    [|
        [|1.84;2.49;1.50;2.42;|]
        [|2.43;1.85;2.42;2.73;|]
        [|3.95;3.67;3.23;2.31;|]
        [|3.21;3.20;2.32;3.30;|]
        //only gives aproximate results when group sizes are unequal
        [|3.21;3.13;2.32;3.30;3.20;2.42;|]
    |]

//first sample serves as control
let dunnetContrastMatrix = 
    [|                
        [|-1.;1.;0.;0.;0.|]
        [|-1.;0.;1.;0.;0.|]
        [|-1.;0.;0.;1.;0.|]
        [|-1.;0.;0.;0.;1.|]
    |]

let dunnettResult = 
    PostHoc.dunnetts dunnetContrastMatrix dunnetExample Tables.dunnettsTwoSided095

[|{ Index = 0
    L = 0.295
    DegreesOfFreedom = 4.0
    MeanSquares = 0.2483897059
    Significance = false
    Statistic = 0.8370862672
    SumOfSquares = 4.222625
    CriticalValue = 2.691 }; { Index = 1
                               L = 1.2275
                               DegreesOfFreedom = 4.0
                               MeanSquares = 0.2483897059
                               Significance = true
                               Statistic = 3.483130146
                               SumOfSquares = 4.222625
                               CriticalValue = 2.691 };
  { Index = 2
    L = 0.945
    DegreesOfFreedom = 4.0
    MeanSquares = 0.2483897059
    Significance = false
    Statistic = 2.681513636
    SumOfSquares = 4.222625
    CriticalValue = 2.691 }; { Index = 3
                               L = 0.8675
                               DegreesOfFreedom = 4.0
                               MeanSquares = 0.2483897059
                               Significance = true
                               Statistic = 2.696548946
                               SumOfSquares = 4.222625
                               CriticalValue = 2.691 }|]

Fisher Hotelling

let d1 = [159.;179.;100.;45.;384.;230.;100.;320.;80.;220.;320.;210.;]
let d2 = [14.4;15.2;11.3;2.5;22.7;14.9;1.41;15.81;4.19;15.39;17.25;9.52; ]
    
Testing.FisherHotelling.test d1 d2

Multiple testing

When conducting multiple hypothesis test the α-error accumulates. This is because the p value just describes the probability for a false positive for one single test. If you perform 10 t-test at an α level of 0.05, the probability of getting a significant result by chance is 40.1% [ (1-(1-α)^k ].

let aErrorAcc = 
    [1. .. 100.]
    |> List.map (fun x -> x,(1. - 0.95**x))
    |> Chart.Line
    |> Chart.withXAxisStyle ( "number of tests (k)" )
    |> Chart.withYAxisStyle ( "probability of at least one false positive test")

To compensate this inflation, several multiple testing corrections were published. The most conservative method is the Bonferroni correction, where the used α level is divided by the number of performed tests.

A modern correction approach is the Benjamini-Hochberg method also known as FDR (false discovery rate).

Benjamini-Hochberg

let pValues =
    [|
        0.000002;0.000048;0.000096;0.000096;0.000351;0.000368;0.000368;0.000371;0.000383;0.000383;0.000705;0.000705;0.000705;0.000705;0.000739;0.00101;0.001234;0.001509;0.001509;0.001509;0.001509;0.001686;0.001686;0.001964;0.001964;0.001964;0.001964;0.001964;0.001964;0.001964;0.002057;0.002295;0.002662;0.002662;
        0.002662;0.002662;0.002662;0.002662;0.002662;0.002672;0.002714;0.002922;0.00348;0.004066;0.004176;0.004176;0.004562;0.004562;0.005848;0.005848;0.006277;0.007024;0.007614;0.007614;0.007614;0.007614;0.007614;0.00979;0.01049;0.01049;0.012498;0.012498;0.012498;0.017908;0.018822;0.019003;0.019003;0.019003;
        0.020234;0.02038;0.021317;0.023282;0.026069;0.026773;0.027255;0.027255;0.027255;0.027255;0.0274;0.030338;0.03128;0.034516;0.034516;0.037267;0.037267;0.040359;0.042706;0.043506;0.04513;0.04513;0.047135;0.049261;0.049261;0.049261;0.049261;0.049333;0.050457;0.052112;0.052476;0.060504;0.063031;0.063031;
        0.063031;0.063031;0.065316;0.065316;0.066751;0.067688;0.069676;0.073043;0.078139;0.078594;0.078594;0.095867;0.098913;0.102606;0.102606;0.102606;0.107444;0.116213;0.126098;0.135099;0.135099;0.159786;0.179654;0.199372;0.203542;0.203542;0.211249;0.211968;0.226611;0.228287;0.238719;0.247204;0.263942;
        0.263942;0.289175;0.306064;0.330191;0.330191;0.340904;0.343869;0.350009;0.355614;0.355614;0.359354;0.386018;0.386018;0.434486;0.438791;0.464694;0.471015;0.4715;0.479307;0.490157;0.505652;0.539465;0.539465;0.558338;0.558338;0.601991;0.61052;0.634365;0.637835;0.677506;0.678222;0.727881;0.748533;
        0.753718;0.758701;0.810979;0.838771;0.854833;0.872159;0.878727;0.890621;0.916361;0.954779;0.98181;0.985365;0.986261;0.98958;0.99861;0.99861;0.999602;0.999895
    |] |> Array.sort

let pValsAdj =
    MultipleTesting.benjaminiHochbergFDR pValues

let bhValues =
    [
        Chart.Line(pValues,pValues,Name="diagonal")
        Chart.Line(pValues,pValsAdj,Name="adj")
    ]
    |> Chart.combine
    |> Chart.withXAxisStyle "pValue" 
    |> Chart.withYAxisStyle "BH corrected pValue"

Q Value

A q value defines the proportion of false positives there are within all discoveries that were called significant up to the current item. It serves as multiple testing correction procedure if a great proportion of tests are reported to be significant, and therefore Benjamini-Hochberg-correction would underestimate the true positive test results. Qvalues with pi0 = 1 is equivalent to Benjamini-Hochber-Method. For an extensive introduction and application example, see q value blog post at fslabs.

\(m = \#tests\)

qValues

\(qvalue_p = \frac{\#(false positives)}{\#positives}\)

qValues robust

Corrects for small p values especially if the number of tests is low or the population distributions do not follow a normal distribution. The correction takes into account the probability of a false positive being reported in the first place.

\(qvalueRobust_p = \frac{\#(false positives)}{\#positives * (1 - (1 - p)**m)}\)

let pi0 = 
    pValues
    |> MultipleTesting.Qvalues.pi0Bootstrap 

pi0

let qValues = 
    pValues
    |> MultipleTesting.Qvalues.ofPValues pi0

let qValuesRob =
    pValues
    |> MultipleTesting.Qvalues.ofPValuesRobust pi0 

let qChart =    
    [
        Chart.Line(pValues,qValues,Name="qValue")
        Chart.Line([(0.,pi0);(1.,pi0)],Name="pi<sub>0</sub>",LineDash=StyleParam.DrawingStyle.Dash)
        Chart.Line(pValues,qValuesRob,Name="qValueRobust")
    ]
    |> Chart.combine
    |> Chart.withTemplate ChartTemplates.lightMirrored
    |> Chart.withXAxisStyle ("p value",MinMax=(0.,1.))
    |> Chart.withYAxisStyle ("q Values",MinMax=(0.,1.))


let qHisto =
    [
        Chart.Histogram(pValues,XBins=TraceObjects.Bins.init(0.,1.,0.05),Name="pValues",HistNorm=HistNorm.ProbabilityDensity)
        Chart.Line([(0.,pi0);(1.,pi0)],Name="pi<sub>0</sub>",LineDash=StyleParam.DrawingStyle.Dash)
    ]
    |> Chart.combine
    |> Chart.withTemplate ChartTemplates.lightMirrored
    |> Chart.withXAxisStyle "p value"
    |> Chart.withYAxisStyle "density"

SAM

SAM (Significance Analysis of Microarrays) is a method developed to overcome multiple testing problems. It was proposed, but not restricted for, microarray analysis. It serves as a blue print for a variety of permutation test. Therefore, high throughput experiments can be analysed using a combined permutation-bootstrap-method. For more in depth information see the SAM BlogPost.

Data:

To use SAM, expression or intensity data need to be in the format ('id * float[])[], where 'id is a generic type implementing System.IComparable.

In this example, 'id is the name of the gene, and float array being the replicates.

One way of achieving this is the following data preparation:

Columns are samples, here 1 and 2, representing control and treatment groups. Rows are transcript counts (here indicated with gene identifier). The next step is to read in the data, e.g. via deedle, and to create a dataframe. The rows are indexed by the sample name and the rowkeys are extracted. Here is the location where the dataset can be accessed: tests\FSharp.Stats.Tests\data\TestDataSAM.txt

let df:Frame<string,string> = 
    Frame.ReadCsv(@"data/TestDataSAM.txt",hasHeaders=true,separators = "\t")
    // here, the name of A1 is needed. 
    |> Frame.indexRows "gene"

// get Rowkeys as Array
let rowheader :string[] = df.RowKeys |> Seq.toArray

// to separate control and factor sets (sample1 and sample2, respectively), 
// the datasets are chunked by the number of replicates (here: triplicates -> chunkBySize 3).
let (preData1,preData2) :float[][] * float [][]=  
    df
    |> Frame.getRows
    |> Series.values
    |> Seq.map (Series.values >> Seq.toArray >> Array.chunkBySize 3 >> fun x -> x.[0],x.[1])
    |> Array.ofSeq
    |> Array.unzip

// After chunking, the datasets are separated by factor and the rowkey is added to the data for later identification.
let data1 = Array.zip rowheader preData1 
let data2 = Array.zip  rowheader preData2 

Optional: Data can be normalised by median centering using the following function. Thereby the median of each column is determined and subtracted it from each value in the respective column. Since samples do not influence each other it makes no difference if you center each column individually or center the whole data matrix prior to segragation.

With data1 and data2, or corrected1 and corrected2, SAM can be performed.

open SAM
let corrected1 = 
    let medCorrect = 
        medianCentering data1 
    Array.zip rowheader medCorrect

let corrected2 = 
    let medCorrect = 
        medianCentering data2
    Array.zip rowheader medCorrect
 
let res = FSharp.Stats.Testing.SAM.twoClassUnpaired 100 0.05 data1 data2 (System.Random(1234))

Besides the data itself SAM requires the definition of some parameters:

• number of permutations: It seems to be sufficient to use ~ 100 permutations (also default in samR).
• desired FDR: While FDRs are given for each BioItem it is recommended to set a global FDR for the generation of the SAM plot.
• random seed: The seed is used for randomization of the permutations (System.Random()), or can be fixed to achieve the same results multiple times (e.g. System.Random(1234)).

The SAMResult<_> type summarizes all information required to construct the typical SAM plot:

• s0: fudge factor as described by Tusher et al. (2001)
• pi0: {pi0 ∈ R ∣ 0 < x < 1} estimated for the determination of qvalues
• delta: distance between the bisector and the parallel margins that separate nonsignificant from significant results
• upper/lower cut: y-axis intercepts that define significance-thresholds for the observed statistics
• positive/negative/non-significant bioitems
• False Discovery Rate (FDR)
• median number of false positives

The following snippet helps to generate the typical SAM plot.

let SAMChart = 

    let observed = [| res.NegSigBioitem; res.NonSigBioitem; res.PosSigBioitem|] |> Array.concat 
    let obs = observed |> Array.map (fun x -> x.Statistics) 
    let expected = res.AveragePermutations |> Array.map (fun x -> x.Statistics)
    let minDi = Seq.min obs
    let maxDi = Seq.max obs


    // positive significant changes 
    let posExpected = expected.[res.NegSigBioitem.Length + res.NonSigBioitem.Length .. res.NegSigBioitem.Length + res.NonSigBioitem.Length + res.PosSigBioitem.Length-1]
    let posChart = 
        Chart.Point(posExpected,res.PosSigBioitem |> Array.map (fun x -> x.Statistics))
        |> Chart.withLineStyle(Color=Color.fromKeyword Green)
        |> Chart.withTraceInfo("positive change",Visible = Visible.True )


    // no significant changes
    let nonex = expected.[res.NegSigBioitem.Length .. res.NegSigBioitem.Length + res.NonSigBioitem.Length-1]
    let nonchart = 
        Chart.Point(nonex,res.NonSigBioitem |> Array.map (fun x -> x.Statistics))
        |> Chart.withLineStyle(Color=Color.fromKeyword Gray)
        |> Chart.withTraceInfo("no change",Visible = Visible.True)

    // negative significant changes 
    let negex = expected.[0 .. res.NegSigBioitem.Length-1]
    let negchart = 
        Chart.Point(negex,res.NegSigBioitem |> Array.map (fun x -> x.Statistics))
        |> Chart.withLineStyle(Color=Color.fromKeyword Red)
        |> Chart.withTraceInfo("negative change",Visible = Visible.True)

    let samValues = 
        [
            negchart
            nonchart
            posChart
        ]
        |> Chart.combine
        
    let chartConfig =
        let svdConfig =
            ConfigObjects.ToImageButtonOptions.init(
                Format = StyleParam.ImageFormat.SVG)
        Config.init (
            ToImageButtonOptions = svdConfig,
            ModeBarButtonsToAdd=[ModeBarButton.HoverCompareCartesian]
            
        )

    let cutLineUp = [(minDi + res.Delta) ; (maxDi + res.Delta)]
    let cutsUp = 
        Chart.Line(cutLineUp,[minDi;maxDi])
        |> Chart.withLineStyle(Color=Color.fromKeyword Purple,Dash = StyleParam.DrawingStyle.Dash, Width = 0.5)
        |> Chart.withTraceInfo("delta",Visible = Visible.True)
    let cutLineLow = [(minDi - res.Delta) ; (maxDi - res.Delta)]
    let cutsLow = 
        Chart.Line(cutLineLow,[minDi;maxDi])
        |> Chart.withLineStyle(Color=Color.fromKeyword Purple,Dash = StyleParam.DrawingStyle.Dash, Width = 0.5)
        |> Chart.withTraceInfo("delta",Visible = Visible.True)
    let linechart = 
        Chart.Line([minDi;maxDi], [minDi;maxDi])
        |> Chart.withTraceInfo("bisecting angle",Visible = Visible.True)
        |> Chart.withLineStyle(Color=Color.fromKeyword Black, Width = 1)

    let uppercut = 
        let xAnchorUppercut = [minDi .. 5. .. maxDi]
        Chart.Line (xAnchorUppercut, List.init xAnchorUppercut.Length (fun x -> res.UpperCut))

        |> Chart.withTraceInfo("upper cut",Visible = Visible.True)
        |> Chart.withLineStyle(Color=Color.fromKeyword Black,Dash = StyleParam.DrawingStyle.Dash, Width = 0.3)

    let lowercut = 
        Chart.Line([minDi;maxDi],[res.LowerCut;res.LowerCut])
        |> Chart.withTraceInfo("lower cut",Visible = Visible.True)
        |> Chart.withLineStyle(Color=Color.fromKeyword Black,Dash = StyleParam.DrawingStyle.Dash,Width = 0.3)
        //|> Chart.withXAxisStyle(MinMax = (-15,20))
        
        |> Chart.withTraceInfo("lower cut",Visible = Visible.True)
    let plot = 
        [linechart;
        samValues;
        cutsUp;
        cutsLow;
        uppercut;
        lowercut]
        |> Chart.combine
        |> Chart.withTitle(title = "SAM results")
        |> Chart.withXAxisStyle("expected Score")
        |> Chart.withYAxisStyle ("observed Score")
        |> Chart.withConfig(chartConfig)
        |> Chart.withTemplate(ChartTemplates.lightMirrored)
    plot