Permutation Hypothesis Test in R Programming
In simple words, the permutation hypothesis test in R is a way of comparing a numerical value of 2 groups. The permutation Hypothesis test is an alternative to:
- Independent two-sample t-test
- Mann-Whitney U aka Wilcoxon Rank-Sum Test
Let’s implement this test in R programming.
Why use the Permutation Hypothesis Test?
- Small Sample Size.
- Assumptions(for parametric approach) not met.
- Test something other than classic approaches comparing Means and Medians.
- Difficult to estimate the SE for test-statistic.
Permutation Hypothesis Test Steps
- Specify a hypothesis
- Choose test-stat(Eg: Mean, Median, etc. )
- Determine Distribution of test-stat
- Convert test-stat to P-value
Note:
P-value = No. of permutations having a test-stat value greater than observed test-stat value/ No. of permutations.
Implementation in R
- Dataset: Chicken Diet Data. This dataset is a subset of the “chickwts” data in the “R dataset package”. Download the data set here.
- Hypothesis: The weight of the chicken is independent of the type of diet.
Test-Statistics
- Test-Statistics #1: The absolute value of the difference in mean weights for the two diets | Y1 – Y2 |. This is the same test statistics as the independent two-sided two-sample t-test.
- Test-Statistics #2: The absolute value of the difference in median weights for the two diets | Median1 – Median2 |
R
# R program to illustrate # Permutation Hypothesis Test # load the data set d <- read.table (file = "ChickData.csv" , header = T, sep = "," ) # print the dataset print (d) # check the names names (d) levels (d$feed) # how many observations in each diet? table (d$feed) # let's look at a boxplot of weight gain by those 2 diets boxplot (d$weight~d$feed, las = 1, ylab = "weight (g)" , xlab = "feed" , main = "Weight by Feed" ) # calculate the difference in sample MEANS mean (d$weight[d$feed == "casein" ]) # mean for casein mean (d$weight[d$feed == "meatmeal" ]) # mean for meatmeal # lets calculate the absolute diff in means test.stat1 <- abs ( mean (d$weight[d$feed == "casein" ]) - mean (d$weight[d$feed == "meatmeal" ])) test.stat1 # calculate the difference in sample MEDIANS median (d$weight[d$feed == "casein" ]) # median for casein median (d$weight[d$feed == "meatmeal" ]) # median for meatmeal # lets calculate the absolute diff in medians test.stat2 <- abs ( median (d$weight[d$feed == "casein" ]) - median (d$weight[d$feed == "meatmeal" ])) test.stat2 # Permutation Test # for reproducability of results set.seed (1979) # the number of observations to sample n <- length (d$feed) # the number of permutation samples to take P <- 100000 # the variable we will resample from variable <- d$weight # initialize a matrix to store the permutation data PermSamples <- matrix (0, nrow = n, ncol = P) # each column is a permutation sample of data # now, get those permutation samples, using a loop # let's take a moment to discuss what that code is doing for (i in 1:P) { PermSamples[, i] <- sample (variable, size = n, replace = FALSE ) } # we can take a quick look at the first 5 columns of PermSamples PermSamples[, 1:5] # initialize vectors to store all of the Test-stats Perm.test.stat1 <- Perm.test.stat2 <- rep (0, P) # loop thru, and calculate the test-stats for (i in 1:P) { # calculate the perm-test-stat1 and save it Perm.test.stat1[i] <- abs ( mean (PermSamples[d$feed == "casein" ,i]) - mean (PermSamples[d$feed == "meatmeal" ,i])) # calculate the perm-test-stat2 and save it Perm.test.stat2[i] <- abs ( median (PermSamples[d$feed == "casein" ,i]) - median (PermSamples[d$feed == "meatmeal" ,i])) } # before going too far with this, # let's remind ourselves of # the TEST STATS test.stat1; test.stat2 # and, take a look at the first 15 # permutation-TEST STATS for 1 and 2 round (Perm.test.stat1[1:15], 1) round (Perm.test.stat2[1:15], 1) # and, let's calculate the permutation p-value # notice how we can ask R a true/false question (Perm.test.stat1 >= test.stat1)[1:15] # and if we ask for the mean of all of those, # it treats 0 = FALSE, 1 = TRUE mean ((Perm.test.stat1 >= test.stat1)[1:15]) # Calculate the p-value, for all P = 100,000 mean (Perm.test.stat1 >= test.stat1) # and, let's calculate the p-value for # option 2 of the test statistic (abs diff in medians) mean (Perm.test.stat2 >= test.stat2) |
Output:
> print(d) weight feed 1 325 meatmeal 2 257 meatmeal 3 303 meatmeal 4 315 meatmeal 5 380 meatmeal 6 153 meatmeal 7 263 meatmeal 8 242 meatmeal 9 206 meatmeal 10 344 meatmeal 11 258 meatmeal 12 368 casein 13 390 casein 14 379 casein 15 260 casein 16 404 casein 17 318 casein 18 352 casein 19 359 casein 20 216 casein 21 222 casein 22 283 casein 23 332 casein > names(d) [1] "weight" "feed" > levels(d$feed) [1] "casein" "meatmeal" > table(d$feed) casein meatmeal 12 11
> mean(d$weight[d$feed == "casein"]) # mean for casein [1] 323.5833 > mean(d$weight[d$feed == "meatmeal"]) # mean for meatmeal [1] 276.9091 > test.stat1 [1] 46.67424 > median(d$weight[d$feed == "casein"]) # median for casein [1] 342 > median(d$weight[d$feed == "meatmeal"]) # median for meatmeal [1] 263 > test.stat2 [1] 79 > PermSamples[, 1:5] [,1] [,2] [,3] [,4] [,5] [1,] 379 283 380 352 206 [2,] 380 303 258 260 380 [3,] 257 206 379 380 153 [4,] 283 242 222 404 359 [5,] 222 260 325 258 258 [6,] 315 352 153 379 263 [7,] 352 263 263 325 325 [8,] 153 325 315 359 216 [9,] 368 379 344 242 260 [10,] 344 258 368 368 257 [11,] 359 257 206 257 315 [12,] 206 153 404 222 303 [13,] 404 344 303 390 390 [14,] 325 318 318 303 352 [15,] 242 404 332 263 404 [16,] 390 380 257 206 379 [17,] 260 332 216 315 318 [18,] 303 359 352 344 368 [19,] 263 222 242 283 222 [20,] 332 368 260 332 344 [21,] 318 315 283 318 283 [22,] 216 390 390 153 332 [23,] 258 216 359 216 242 > test.stat1; test.stat2 [1] 46.67424 [1] 79 > round(Perm.test.stat1[1:15], 1) [1] 17.1 32.4 17.6 47.1 56.1 28.9 31.0 40.8 6.8 13.8 9.1 46.5 28.9 50.9 32.7 > round(Perm.test.stat2[1:15], 1) [1] 61.0 75.0 4.5 59.0 78.0 17.0 62.0 38.5 4.5 16.0 23.0 60.5 63.5 75.0 37.0 > (Perm.test.stat1 >= test.stat1)[1:15] [1] FALSE FALSE FALSE TRUE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE FALSE > mean((Perm.test.stat1 >= test.stat1)[1:15]) [1] 0.2 > mean(Perm.test.stat1 >= test.stat1) [1] 0.09959 > mean(Perm.test.stat2 >= test.stat2) [1] 0.05407
Please Login to comment...