Lesson 99 – The Two-Sample Hypothesis Tests in R

Over the past seven lessons, we equipped ourselves with the necessary theory of the two-sample hypothesis tests.

Lessons 92 and 93 were about the hypothesis test on the difference in proportions. In Lesson 92, we learned Fisher’s Exact Test to verify the difference in proportions. In Lesson 93, we did the same using the normal distribution as a limiting distribution when the sample sizes are large.

Lessons 94, 95, and 96 were about the hypothesis test on the difference in means. In Lesson 94, we learned that under the proposition that the population variances of two random variables are equal, the test-statistic, which uses the pooled variance, follows a T-distribution with n_{1}+n_{2}-2 degrees of freedom. In Lesson 95, we learned that Welch’s t-Test could be used when we cannot make the equality of population proportions assumption. In Lesson 96, we learned the Wilcoxon’s Rank-sum Test that used a ranking approach to approximate the significance of the differences in means. This method is data-driven, needs no assumptions on the limiting distribution, and works well for small sample sizes.

Lesson 97 was about the hypothesis test on the equality of variances. Here we got a sneak-peak into a new distribution called F-distribution, which is the converging distribution of the ratio of two Chi-square distributions divided by their respective degrees of freedom. The test-statistic is the ratio of the sample variances, which we can verify against an F-distribution with n_{1}-1 numerator degrees of freedom and n_{2}-1 denominator degrees of freedom.

All these hypothesis tests can also be done using the bootstrap approach, which we learned in Lesson 98. It uses the data at hand to generate the null distribution of any desired statistic as long as it is computable from the data. It is very flexible. There is no need to make any assumptions on the data’s distributional nature or the limiting distribution for the test-statistic.

It is now time to put all of this learning to practice.


The City of New York Department of Sanitation (DSNY) archives data on the monthly tonnage of trash collected from NYC residences and institutions.

To help us with today’s workout, I dug up our gold mine, the “Open Data for All New Yorkers” page, and found an interesting dataset on the monthly tonnage of trash collected from NYC residences and institutions.

Here is a preview.

The data includes details on when (year and month) DSNY collected the trash, from where (borough and community district), and how much (tonnage of refuse, source-separated recyclable paper, and source-separated metal, glass, plastic, and beverage cartons). They also have other seasonal data such as the tonnage of leaves collected in November and December and tons of Christmas trees collected from curbside in January.

For our workout today, we can use this file named “DSNY_Monthly_Tonnage_Data.csv.” It is a slightly modified version of the file you will find from the open data page.

My interest is in Manhattan’s community district 9 — Morningside Heights, Manhattanville, and Hamilton Heights.

I want to compare the tonnage of trash collected in winter to that in summer. Let’s say we compare the data in Manhattan’s community district 9 for February and August to keep it simple.

Are you ready?

The Basic Steps

Step1: Get the data

You can get the data from here. The file is named DSNY_Monthly_Tonnage_Data.csv. It is a comma-separated values file which we can conveniently read into the R workspace. 

Step 2Create a new folder on your computer

And call this folder “lesson99.”
Make sure that the data file “DSNY_Monthly_Tonnage_Data.csv” is in this folder.

Step 3Create a new code in R

Create a new code for this lesson. “File >> New >> R script”. Save the code in the same folder “lesson99” using the “save” button or by using “Ctrl+S.” Use .R as the extension — “lesson99_code.R

Step 4Choose your working directory

“lesson99” is the folder where we stored the code and the input data file. Use setwd("path") to set the path to this folder. Execute the line by clicking the “Run” button on the top right. 

setwd("path to your folder")

Or, if you opened this “lesson99_code.R” file from within your “lesson99” folder, you can also use the following line to set your working directory.

setwd(getwd())

getwd() gets the current working directory. If you opened the file from within your director, then getwd() will resort to the working directory you want to set.

Step 5Read the data into R workspace

Execute the following command to read your input .csv file.

# Read the data file
nyc_trash_data = read.csv("DSNY_Monthly_Tonnage_Data.csv",header=T)

Since the input file has a header for each column, we should have header=T in the command when we read the file to ensure that all the rows are correctly read.

Your RStudio interface will look like this when you execute this line and open the file from the Global Environment.

Step 6Extracting a subset of the data

As I said before, for today’s workout, let’s use the data from Manhattan’s community district 9 for February and August. Execute the following lines to extract this subset of the data.

# Extract February refuse data for Manhattan's community district 9 
feb_index = which((nyc_trash_data$BOROUGH=="Manhattan") & (nyc_trash_data$COMMUNITYDISTRICT==9) & (nyc_trash_data$MONTH==2))

feb_refuse_data = nyc_trash_data$REFUSETONSCOLLECTED[feb_index]

# Extract August refuse data for Manhattan's community district 9
aug_index = which((nyc_trash_data$BOROUGH=="Manhattan") & (nyc_trash_data$COMMUNITYDISTRICT==9) & (nyc_trash_data$MONTH==8))

aug_refuse_data = nyc_trash_data$REFUSETONSCOLLECTED[aug_index]

In the first line, we look up for the row index of the data corresponding to Borough=Manhattan, Community District = 9, and Month = 2.

In the second line, we extract the data on refuse tons collected for these rows.

The next two lines repeat this process for August data.

Sample 1, February data for Manhattan’s community district 9 looks like this:
2953.1, 2065.8, 2668.2, 2955.4, 2799.4, 2346.7, 2359.6, 2189.4, 2766.1, 3050.7, 2175.1, 2104.0, 2853.4, 3293.2, 2579.1, 1979.6, 2749.0, 2871.9, 2612.5, 455.9, 1951.8, 2559.7, 2753.8, 2691.0, 2953.1, 3204.7, 2211.6, 2857.9, 2872.2, 1956.4, 1991.3

Sample 2, August data for Manhattan’s community district 9 looks like this:
2419.1, 2343.3, 3032.5, 3056.0, 2800.7, 2699.9, 3322.3, 3674.0, 3112.2, 3345.1, 3194.0, 2709.5, 3662.6, 3282.9, 2577.9, 3179.1, 2460.9, 2377.1, 3077.6, 3332.8, 2233.5, 2722.9, 3087.8, 3353.2, 2992.9, 3212.3, 2590.1, 2978.8, 2486.8, 2348.2

The values are in tons per month, and there are 31 values for sample 1 and 30 values for sample 2. Hence, n_{1}=31 and n_{2}=30.

The Questions for Hypotheses

Let’s ask the following questions.

  1. Is the mean of the February refuse collected from Manhattan’s community district 9 different from August?
  2. Is the variance of the February refuse collected from Manhattan’s community district 9 different from August?
  3. Suppose we set 2500 tons as a benchmark for low trash situations, is the proportion of February refuse lower than 2500 tons different from the proportion of August? In other words, is the low trash situation in February different than August?

We can visualize their distributions. Execute the following lines to create a neat graphic of the boxplots of sample 1 and sample 2.

# Visualizing the distributions of the two samples
boxplot(cbind(feb_refuse_data,aug_refuse_data),horizontal=T, main="Refuse from Manhattan's Community District 9")
text(1500,1,"February Tonnage",font=2)
text(1500,2,"August Tonnage",font=2)

p_threshold = 2500 # tons of refuse
abline(v=p_threshold,lty=2,col="maroon")

You should now see your plot space changing to this.

Next, you can execute the following lines to compute the preliminaries required for our tests.

# Preliminaries
  # sample sizes
  n1 = length(feb_refuse_data)
  n2 = length(aug_refuse_data)
  
  # sample means
  x1bar = mean(feb_refuse_data)
  x2bar = mean(aug_refuse_data)

  # sample variances
  x1var = var(feb_refuse_data)
  x2var = var(aug_refuse_data)

  # sample proportions
  p1 = length(which(feb_refuse_data < p_threshold))/n1
  p2 = length(which(aug_refuse_data < p_threshold))/n2

Hypothesis Test on the Difference in Means

We learned four methods to verify the hypothesis on the difference in means:

  • The two-sample t-Test
  • Welch’s two-sample t-Test
  • Wilcoxon’s Rank-sum Test
  • The Bootstrap Test 

Let’s start with the two-sample t-Tests. Assuming a two-sided alternative, the null and alternate hypothesis are:

H_{0}: \mu_{1} - \mu_{2}=0

H_{A}: \mu_{1} - \mu_{2} \neq 0

The alternate hypothesis indicates that substantial positive or negative differences will allow the rejection of the null hypothesis.

We know that for the two-sample t-Test, we assume that the population variances are equal, and the test-statistic is t_{0}=\frac{\bar{x_{1}}-\bar{x_{2}}}{\sqrt{s^{2}(\frac{1}{n_{1}}+\frac{1}{n_{2}})}}, where s^{2}=(\frac{n_{1}-1}{n_{1}+n_{2}-2})s_{1}^{2}+(\frac{n_{2}-1}{n_{1}+n_{2}-2})s_{2}^{2} is the pooled variance. t_{0} follows a T-distribution with n_{1}+n_{2}-2 degrees of freedom. We can compute the test-statistic and check how likely it is to see such a value in a T-distribution (null distribution) with so many degrees of freedom.

Execute the following lines in R.

pooled_var = ((n1-1)/(n1+n2-2))*x1var + ((n2-1)/(n1+n2-2))*x2var

t0 = (x1bar-x2bar)/sqrt(pooled_var*((1/n1)+(1/n2)))

df = n1+n2-2

pval = pt(t0,df=df)

print(pooled_var)
print(df)
print(t0)
print(pval)

We first computed the pooled variance s^{2} and, using it, the test-statistic t_{0}. Then, based on the degrees of freedom, we compute the p-value.

The p-value is 0.000736. For a two-sided test, we compare this with \frac{\alpha}{2}. If we opt for a rejection rate \alpha of 5%, then, since our p-value of 0.000736 is less than 0.025, we reject the null hypothesis that the means are equal.

All these steps can be implemented using a one-line command in R.

Try this:

t.test(feb_refuse_data,aug_refuse_data,alternative="two.sided",var.equal = TRUE)

You will get the following prompt in the console.

Two Sample t-test

data: feb_refuse_data and aug_refuse_data

t = -3.3366, df = 59, p-value = 0.001472

alternative hypothesis: true difference in means is not equal to 0

95 percent confidence interval:

 -658.2825 -164.7239

sample estimates:

mean of x mean of y 

 2510.697 2922.200 

As inputs of the function, we need to provide the data (sample 1 and sample 2). We should also indicate that we evaluate a two-sided alternate hypothesis and that the population variances are equal. This will prompt the function to execute the standard two-sample t-test.

While the function provides additional information, all we need for our test are pointers from the first line. t=-3.3366 and df=59 are the test-statistic and the degrees of freedom that we computed earlier. The p-value looks different than what we computed. It is two times the value we estimated. The function provides a value that is double that of the original p-value to allow us to compare it to \alpha instead of \frac{\alpha}{2}. Either way, we know that we can reject the null hypothesis.


Next, let’s implement Welch’s t-Test. When the population variances are not equal, i.e., when \sigma_{1}^{2} \neq \sigma_{2}^{2}, the test-statistic is t_{0}^{*}=\frac{\bar{x_{1}}-\bar{x_{2}}}{\sqrt{\frac{s_{1}^{2}}{n_{1}}+\frac{s_{2}^{2}}{n_{2}}}}, and it follows an approximate T-distribution with f degrees of freedom which can be estimated using the Satterthwaite’s equation: f = \frac{(\frac{s_{1}^{2}}{n_{1}}+\frac{s_{2}^{2}}{n_{2}})^{2}}{\frac{(s_{1}^{2}/n_{1})^{2}}{(n_{1} - 1)}+\frac{(s_{2}^{2}/n_{2})^{2}}{(n_{2}-1)}}

Execute the following lines.

f = (((x1var/n1)+(x2var/n2))^2)/(((x1var/n1)^2/(n1-1))+((x2var/n2)^2/(n2-1)))

t0 = (x1bar-x2bar)/sqrt((x1var/n1)+(x2var/n2))

pval = pt(t0,df=f)

print(f)
print(t0)
print(pval)

The pooled degrees of freedom is 55, the test-statistic is -3.3528, and the p-value is 0.000724. Comparing this with 0.025, we can reject the null hypothesis.

Of course, all this can be done using one line where you indicate that the variances are not equal:

t.test(feb_refuse_data,aug_refuse_data,alternative="two.sided",var.equal = FALSE)

When you execute the above line, you will see the following on your console.

Welch Two Sample t-test
data: feb_refuse_data and aug_refuse_data
t = -3.3528, df = 55.338, p-value = 0.001448
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
-657.4360 -165.5705
sample estimates:
mean of x mean of y
2510.697 2922.200

Did you notice that it runs Welch’s two-sample t-Test when we indicate that the variances are not equal? Like before, we can compare the p-value to 0.05 and reject the null hypothesis.

The mean of the February refuse collected from Manhattan’s community district 9 is different from August. Since the test-statistic is negative and the p-value is lower than the chosen level of rejection rate, we can conclude that the mean of February refuse tonnage is significantly lower than the mean of August refuse tonnage beyond a reasonable doubt.


How about Wilcoxon’s Rank-sum Test? The null hypothesis is based on the proposition that if x_{1} and x_{2} are samples from the same distribution, there will be an equal likelihood of one exceeding the other.

H_{0}: P(x_{1} > x_{2}) = 0.5

H_{0}: P(x_{1} > x_{2}) \neq 0.5

We know that this ranking and coming up with the rank-sum table is tedious for larger sample sizes. For larger sample sizes, the null distribution of the test-statistic W, which is the sum of the ranks associated with the variable of smaller sample size in the pooled and ordered data, tends to a normal distribution. This allows the calculation of the p-value by comparing W to a normal distribution. We can use the following command in R to implement this.

wilcox.test(feb_refuse_data,aug_refuse_data,alternative = "two.sided")

You will get the following message in the console when you execute the above line.

Wilcoxon rank sum test with continuity correction
data: feb_refuse_data and aug_refuse_data
W = 248, p-value = 0.001788
alternative hypothesis: true location shift is not equal to 0

Since the p-value of 0.001788 is lower than 0.05, the chosen level of rejection, we reject the null hypothesis that x_{1} and x_{2} are samples from the same distribution.


Finally, let’s take a look at the bootstrap method. The null hypothesis is that there is no difference between the means. 

H_{0}: P(\bar{x}_{Feb}>\bar{x}_{Aug}) = 0.5

H_{0}: P(\bar{x}_{Feb}>\bar{x}_{Aug}) \neq 0.5

We first create a bootstrap replicate of X and Y by randomly drawing with replacement n_{1} values from X and n_{2} values from Y.

For each bootstrap replicate from X and Y, we compute the statistics \bar{x}_{Feb} and \bar{x}_{Aug} and check whether \bar{x}_{Feb}>\bar{x}_{Aug}. If yes, we register S_{i}=1. If not, we register S_{i}=0.

We repeat this process of creating bootstrap replicates of X and Y, computing the statistics \bar{x}_{Feb} and \bar{x}_{Aug}, and verifying whether \bar{x}_{Feb}>\bar{x}_{Aug} and registering S_{i} \in (0,1) a large number of times, say N=10,000.

The proportion of times S_{i} = 1 in a set of N bootstrap-replicated statistics is the p-value.

Execute the following lines in R to implement these steps.

#4. Bootstrap
  N = 10000
  null_mean = matrix(0,nrow=N,ncol=1)
  null_mean_ratio = matrix(0,nrow=N,ncol=1)
 
 for(i in 1:N)
   {
     xboot = sample(feb_refuse_data,replace=T)
     yboot = sample(aug_refuse_data,replace=T)
     null_mean_ratio[i] = mean(xboot)/mean(yboot) 
     if(mean(xboot)>mean(yboot)){null_mean[i]=1} 
   }
 
 pvalue_mean = sum(null_mean)/N
 hist(null_mean_ratio,font=2,main="Null Distribution Assuming H0 is True",xlab="Xbar/Ybar",font.lab=2)
 abline(v=1,lwd=2,lty=2)
 text(0.95,1000,paste("p-value=",pvalue_mean),col="red")

Your RStudio interface will then look like this:

A vertical bar will show up at a ratio of 1 to indicate that the area beyond this value is the proportion of times S_{i} = 1 in 10,000 bootstrap-replicated means.

The p-value is close to 0. We reject the null hypothesis since the p-value is less than 0.025. Since more than 97.5% of the times, the mean of February tonnage is less than August tonnage; there is sufficient evidence that they are not equal. So we reject the null hypothesis.

In summary, while the standard t-Test, Welch's t-Test, and Wilcoxon's Rank-sum Test can be implemented in R using single-line commands,

t.test(x1,x2,alternative="two.sided",var.equal = TRUE)
t.test(x1,x2,alternative="two.sided",var.equal = FALSE)
wilcox.test(x1,x2,alternative = "two.sided")

the bootstrap test needs a few lines of code.

Hypothesis Test on the Equality of Variances

We learned two methods to verify the hypothesis on the equality of variances; using F-distribution and using the bootstrap method.

For the F-distribution method, the null and the alternate hypothesis are 

H_{0}:\sigma_{1}^{2}=\sigma_{2}^{2}

H_{A}:\sigma_{1}^{2} \neq \sigma_{2}^{2}

The test-statistic is the ratio of the sample variances, f_{0}=\frac{s_{1}^{2}}{s_{2}^{2}}

We evaluate the hypothesis based on where this test-statistic lies in the null distribution or how likely it is to find a value as large as this test-statistic f_{0} in the null distribution.

Execute the following lines.

# 1. F-Test
  f0 = x1var/x2var
  
  df_numerator = n1-1
  
  df_denominator = n2-1
  
  pval = 1-pf(f0,df1=df_numerator,df2=df_denominator)
  
  print(f0)
  print(df_numerator)
  print(df_denominator)
  print(pval)

The test-statistic, i.e., the ratio of the sample variances, is 1.814.

The p-value, which, in this case, is the probability of finding a value greater than the test-statistic (P(F > f_{0})), is 0.0562. When compared to a one-sided rejection rate (\frac{\alpha}{2}) of 0.025, we cannot reject the null hypothesis

We do have a one-liner command.

var.test(feb_refuse_data,aug_refuse_data,alternative = "two.sided")

You will find the following message when you execute the var.test() command.

F test to compare two variances
data: feb_refuse_data and aug_refuse_data
F = 1.814, num df = 30, denom df = 29, p-value = 0.1125
alternative hypothesis: true ratio of variances is not equal to 1
95 percent confidence interval:
0.8669621 3.7778616
sample estimates:
ratio of variances
1.813959

As usual, with the standard R function, we need to compare the p-value (which is two times what we computed based on the right-tail probability) with \alpha=0.05. Since 0.1125 > 0.05, we cannot reject the null hypothesis that the variances are equal.


Can we come to the same conclusion using the bootstrap test? Let’s try. Execute the following lines in R.

# 2. Bootstrap
 N = 10000
   null_var = matrix(0,nrow=N,ncol=1)
   null_var_ratio = matrix(0,nrow=N,ncol=1)
 
 for(i in 1:N)
   {
     xboot = sample(feb_refuse_data,replace=T)
     yboot = sample(aug_refuse_data,replace=T)
     null_var_ratio[i] = var(xboot)/var(yboot) 
     if(var(xboot)>var(yboot)){null_var[i]=1} 
   }
 
 pvalue_var = sum(null_var)/N
 hist(null_var_ratio,font=2,main="Null Distribution Assuming H0 is True",xlab="XVar/YVar",font.lab=2)
 abline(v=1,lwd=2,lty=2)
 text(2,500,paste("p-value=",pvalue_var),col="red")

Your RStudio interface should look like this if you correctly coded up and executed the lines. The code provides a way to visualize the null distribution.

The p-value is 0.7945. 7945 out of the 10,000 bootstrap replicates had s^{2}_{Feb}>s^{2}_{Aug}. For a 5% rate of error (\alpha=5\%), we cannot reject the null hypothesis since the p-value is greater than 0.025. The evidence (20.55% of the times) that the February tonnage variance is less than August tonnage is not sufficient to reject equality.

In summary, use var.test(x1,x2,alternative = "two.sided") if you want to verify based on F-distribution, or code up a few lines if you want to verify using the bootstrap method.

Hypothesis Test on the Difference in Proportions

For the hypothesis test on the difference in proportions, we can employ Fisher’s Exact Test, use the normal approximation under the large-sample assumption, or the bootstrap method.

For Fisher’s Exact Test, the test-statistic follows a hypergeometric distribution when H_{0} is true. We could assume that the number of successes is fixed at t=x_{1}+x_{2}, and, for a fixed value of t, we reject H_{0}:p_{1}=p_{2} for the alternate hypothesis H_{A}:p_{1}>p_{2} if there are more successes in random variable X_{1} compared to X_{2}.

The p-value can be derived under the assumption that the number of successes X=k in the first sample X_{1} has a hypergeometric distribution when H_{0} is true and conditional on a total number of t successes that can come from any of the two random variables X_{1} and X_{2}.

P(X=k) = \frac{\binom{t}{k}*\binom{n_{1}+n_{2}-t}{n_{1}-k}}{\binom{n_{1}+n_{2}}{n_{1}}}

Execute the following lines to implement this process, plot the null distribution and compute the p-value.

#Fisher's Exact Test
  x1 = length(which(feb_refuse_data < p_threshold))
  p1 = x1/n1
  
  x2 = length(which(aug_refuse_data < p_threshold))
  p2 = x2/n2
  
  N = n1+n2
  t = x1+x2
  
  k = seq(from=0,to=n1,by=1)
  p = k
  for(i in 1:length(k)) 
   {
    p[i] = (choose(t,k[i])*choose((N-t),(n1-k[i])))/choose(N,n1)
   }

  plot(k,p,type="h",xlab="Number of successes in X1",ylab="P(X=k)",font=2,font.lab=2)
  points(k,p,type="o",lty=2,col="grey50")
  points(k[13:length(k)],p[13:length(k)],type="o",col="red",lwd=2)
  points(k[13:length(k)],p[13:length(k)],type="h",col="red",lwd=2)
  pvalue = sum(p[13:length(k)])
  print(pvalue)

Your RStudio should look like this once you run these lines.

The p-value is 0.1539. Since it is greater than the rejection rate, we cannot reject the null hypothesis that the proportions are equal.

We can also leverage an in-built R function that does the same. Try this.

fisher_data = cbind(c(x1,x2),c((n1-x1),(n2-x2)))
fisher.test(fisher_data,alternative="greater")

In the first line, we create a 2×2 contingency table to indicate the number of successes and failures in each sample. Sample 1 has 12 successes, i.e., 12 times the tonnage in February was less than 2500 tons. Sample 2 has seven successes. So the contingency table looks like this:

    [,1]  [,2]
[1,] 12    19
[2,]  7    23

Then, fisher.test() funtion implements the hypergeometric distribution calculations.

You will find the following prompt in your console.

Fisher’s Exact Test for Count Data
data: fisher_data
p-value = 0.1539
alternative hypothesis: true odds ratio is greater than 1
95 percent confidence interval:
0.712854 Inf
sample estimates:
odds ratio
2.050285

We get the same p-value as before — we cannot reject the null hypothesis.


Under the normal approximation for large sample sizes, the test-statistic z = \frac{\hat{p_{1}}-\hat{p_{2}}}{\sqrt{p(1-p)*(\frac{1}{n_{1}}+\frac{1}{n_{2}})}} \sim N(0,1). p is the pooled proportion which can be compute as p = \frac{x_{1}+x_{2}}{n_{1}+n_{2}}.

We reject the null hypothesis when the p-value P(Z \ge z) is less than the rate of rejection \alpha.

Execute the following lines to set this up in R.

# Z-approximation
  p = (x1+x2)/(n1+n2)
  z = (p1-p2)/sqrt(p*(1-p)*((1/n1)+(1/n2)))
  pval = 1-pnorm(z)

The p-value is 0.0974. Since it is greater than 0.05, we cannot reject the null hypothesis.


Finally, we can use the bootstrap method as follows.

# Bootstrap
 N = 10000
 null_prop = matrix(0,nrow=N,ncol=1)
 null_prop_ratio = matrix(0,nrow=N,ncol=1)
 
 for(i in 1:N)
   {
     xboot = sample(feb_refuse_data,replace=T)
     yboot = sample(aug_refuse_data,replace=T)
 
     p1boot = length(which(xboot < p_threshold))/n1 
     p2boot = length(which(yboot < p_threshold))/n2 

     null_prop_ratio[i] = p1boot/p2boot 
    if(p1boot>p2boot){null_prop[i]=1} 
  }
 
 pvalue_prop = sum(null_prop)/N
 hist(null_prop_ratio,font=2,main="Null Distribution Assuming H0 is True",xlab="P1/P2",font.lab=2)
 abline(v=1,lwd=2,lty=2)
 text(2,250,paste("p-value=",pvalue_prop),col="red")

We evaluated whether the low trash situation in February is different than August.

H_{0}: P(p_{Feb}>p_{Aug}) = 0.5

H_{A}: P(p_{Feb}>p_{Aug}) > 0.5

The p-value is 0.8941. 8941 out of the 10,000 bootstrap replicates had p_{Feb}>p_{Aug}. For a 5% rate of error (\alpha=5\%), we cannot reject the null hypothesis since the p-value is not greater than 0.95. The evidence (89.41% of the times) that the low trash situation in February is greater than August is insufficient to reject equality.

Your Rstudio interface will look like this.

In summary, 
use fisher.test(contingency_table,alternative="greater") 
if you want to verify based on Fisher's Exact Test; 

use p = (x1+x2)/(n1+n2); 
    z = (p1-p2)/sqrt(p*(1-p)*((1/n1)+(1/n2))); 
    pval = 1-pnorm(z) 
if you want to verify based on the normal approximation; 

code up a few lines if you want to verify using the bootstrap method.

  1. Is the mean of the February refuse collected from Manhattan’s community district 9 different from August?
    • Reject the null hypothesis. They are different.
  2. Is the variance of the February refuse collected from Manhattan’s community district 9 different from August?
    • Cannot reject the null hypothesis. They are probably the same.
  3. Suppose we set 2500 tons as a benchmark for low trash situations, is the proportion of February refuse lower than 2500 tons different from the proportion of August? In other words, is the low trash situation in February different than August?
    • Cannot reject the null hypothesis. They are probably the same.

Please take the next two weeks to digest all the two-sample hypothesis tests, including how to execute them in R.

Here is the full code for today’s lesson.

In a little over four years, we reached 99 lessons 😎

I will have something different for the 100. Stay tuned …

If you find this useful, please like, share and subscribe.
You can also follow me on Twitter @realDevineni for updates on new lessons.

Lesson 98 – The Two-Sample Hypothesis Tests using the Bootstrap

Two-Sample Hypothesis Tests – Part VII

H_{0}: P(\theta_{x}>\theta_{y}) = 0.5

H_{A}: P(\theta_{x}>\theta_{y}) > 0.5

H_{A}: P(\theta_{x}>\theta_{y}) < 0.5

H_{A}: P(\theta_{x}>\theta_{y}) \neq 0.5

These days, a peek out of the window is greeted by chilling rain or warm snow. On days when it is not raining or snowing, there is biting cold. So we gaze at our bicycles, waiting for that pleasant month of April when we can joyfully bike — to work, or for pleasure.

Speaking of bikes, since I have nothing much to do today except watch the snow, I decided to explore some data from our favorite “Open Data for All New Yorkers” page.

Interestingly, I found data on the bicycle counts for East River Bridges. New York City DOT keeps track of the daily total of bike counts on the Brooklyn Bridge, Manhattan Bridge, Williamsburg Bridge, and Queensboro Bridge.

I could find the data for April to October during 2016 and 2017. Here is how the data for April 2017 looks.

They highlight all non-holiday weekdays with no precipitation in yellow.

Being a frequent biker on the Manhattan Bridge, my curiosity got kindled. I wanted to verify how different the total bike counts on the Manhattan Bridge are from the Williamsburg Bridge.

At the same time, I also wanted to share the benefits of the bootstrap method for two-sample hypothesis tests.

To keep it simple and easy for you to follow the bootstrap method’s logical development, I will test how different the total bike counts data on Manhattan Bridge are from that of the Williamsburg Bridge during all the non-holiday weekdays with no precipitation.

Here is the data of the total bike counts on Manhattan Bridge during all the non-holiday weekdays with no precipitation in April of 2017 — essentially, the data from the yellow-highlighted rows in the table for Manhattan Bridge.

5276, 6359, 7247, 6052, 5054, 6691, 5311, 6774

And the data of the total bike counts on Williamsburg Bridge during all the non-holiday weekdays with no precipitation in April of 2017.

5711, 6881, 8079, 6775, 5877, 7341, 6026, 7196

Their distributions look like this.

We are looking at the boxplots that present a nice visual of the data range and its percentiles. And we can compare one sample to another. Remember Lesson 14? There is a vertical line at 6352 bikes, the maximum number of bikes on Manhattan Bridge during weekends, holidays, or rainy days — i.e., the non-highlighted days.

I want answers to the following questions.

Is the mean of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge. \bar{x}_{M}=\bar{x}_{W}?
Is the median of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge. \tilde{x}_{M}=\tilde{x}_{W}?
Is the variance of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge. s^{2}_{M}=s^{2}_{W}?
Is the interquartile range of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge. IQR_{M}=IQR_{W}?
Is the proportion of the total bike counts on Manhattan Bridge less than 6352, different from that on Williamsburg Bridge. P(M<6352)=P(W<6352) or p_{M}=p_{W}?

What do we know so far? 

We know how to test the difference in means using the t-Test under the proposition that the population variances are equal (Lesson 94) or using Welch’s t-Test when we cannot assume equality of population variances (Lesson 95). We also know how to do this using Wilcoxon’s Rank-sum Test that uses the ranking method to approximate the significance of the differences in means (Lesson 96).

We know how to test the equality of variances using F-distribution (Lesson 97).

We know how to test the difference in proportions using either Fisher’s Exact test (Lesson 92) or using the normal distribution as the null distribution under the large-sample approximation (Lesson 93).

In all these tests, we made critical assumptions on the limiting distributions of the test-statistics.

  • What is the limiting distribution of the test-statistic that computes the difference in medians?
  • What is the limiting distribution of the test-statistic that compares interquartile ranges of two populations?
  • What if we do not want to make any assumptions on data distributions or the limiting forms of the test-statistics?

Enter the Bootstrap

I would urge you to go back to Lesson 79 to get a quick refresher on the bootstrap, and Lesson 90 to recollect how we used it for the one-sample hypothesis tests.

The idea of the bootstrap is that we can generate replicates of the original sample to approximate the probability distribution function of the population. Assuming that each data value in the sample is equally likely with a probability of 1/n, we can randomly draw n values with replacement. By putting a probability of 1/n on each data point, we use the discrete empirical distribution \hat{f} as an approximation of the population distribution f.

Take the data for Manhattan Bridge.
5276, 6359, 7247, 6052, 5054, 6691, 5311, 6774

Assuming that each data value is equally likely, i.e., the probability of occurrence of any of these eight data points is 1/8, we can randomly draw eight numbers from these eight values — with replacement.

It is like playing the game of Bingo where the chips are these eight numbers. Each time we get a number, we put it back and roll it again until we draw eight numbers.

Since each value is equally likely, the bootstrap sample will consist of numbers from the original data (5276, 6359, 7247, 6052, 5054, 6691, 5311, 6774), some may appear more than one time, and some may not show up at all in a random sample.

Here is one such bootstrap replicate.
6359, 6359, 6359, 6052, 6774, 6359, 5276, 6359

The value 6359 appeared five times. Some values like 7247, 5054, 6691, and 5311 did not appear at all in this replicate.

Here is another replicate.
6359, 5276, 5276, 5276, 7247, 5311, 6052, 5311

Such bootstrap replicates are representations of the empirical distribution \hat{f}, i.e., the proportion of times each value in the data sample occurs. We can generate all the information contained in the true distribution by creating \hat{f}, the empirical distribution.

Using the Bootstrap for Two-Sample Hypothesis Tests

Since each bootstrap replicate is a possible representation of the population, we can compute the relevant test-statistics from this bootstrap sample. By repeating this, we can have many simulated values of the test-statistics that form the null distribution to test the hypothesis. There is no need to make any assumptions on the distributional nature of the data or the limiting distribution for the test-statistic. As long as we can compute a test-statistic from the bootstrap sample, we can test the hypothesis on any statistic — mean, median, variance, interquartile range, proportion, etc.

Let’s now use the bootstrap method for two-sample hypothesis tests. Suppose there are two random variables, X and Y, and any statistic computed from them are \theta_{x} and \theta_{y}. We may have a sample of n_{1} values representing X and a sample of n_{2} values to represent Y.

\theta_{x},\theta_{y} can be mean, median, variance, proportion, etc. Any computable statistic from the original data is of the form \theta_{x},\theta_{y}.

The null hypothesis is that there is no difference between the statistic of X or Y.

H_{0}: P(\theta_{x}>\theta_{y}) = 0.5

The alternate hypothesis is

H_{A}: P(\theta_{x}>\theta_{y}) > 0.5

or

H_{A}: P(\theta_{x}>\theta_{y}) < 0.5

or

H_{A}: P(\theta_{x}>\theta_{y}) \neq 0.5

We first create a bootstrap replicate of X and Y by randomly drawing with replacement n_{1} values from X and n_{2} values from Y.

For each bootstrap replicate i from X and Y, we compute the statistics \theta_{x} and \theta_{y} and check whether \theta_{x}>\theta_{y}. If yes, we register S_{i}=1. If not, we register S_{i}=0.

For example, one bootstrap replicate for X (Manhattan Bridge) and Y (Williamsburg Bridge) may look like this:

xboot: 6691 5311 6774 5311 6359 5311 5311 6052
yboot: 6775 6881 7341 7196 6775 7341 6775 7196

Mean of this bootstrap replicate for X and Y are 5890 and 7035. Since \bar{x}^{X}_{boot}<\bar{x}^{Y}_{boot}, we register S_{i}=0

Another bootstrap replicate for X and Y may look like this:

xboot: 6774 6359 6359 6359 6052 5054 6052 6691
yboot: 6775 7196 7196 6026 6881 7341 6881 5711

Mean of this bootstrap replicate for X and Y are 6212.5 and 6750.875. Since \bar{x}^{X}_{boot}<\bar{x}^{Y}_{boot}, we register S_{i}=0

We repeat this process of creating bootstrap replicates of X and Y, computing the statistics \theta_{x} and \theta_{y}, and verifying whether \theta_{x}>\theta_{y} and registering S_{i} \in (0,1) a large number of times, say N=10,000.

The proportion of times S_{i} = 1 in a set of N bootstrap-replicated statistics is the p-value.

p-value=\frac{1}{N}\sum_{i=1}^{i=N}S_{i}

Based on the p-value, we can use the rule of rejection at the selected rate of rejection \alpha.

For a one-sided alternate hypothesis, we reject the null hypothesis if p-value < \alpha or p-value > 1-\alpha.

For a two-sided alternate hypothesis, we reject the null hypothesis if p-value < \frac{\alpha}{2} or p-value > 1-\frac{\alpha}{2}.

Manhattan Bridge vs. Williamsburg Bridge

Is the mean of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge?

H_{0}: P(\bar{x}_{M}>\bar{x}_{W}) = 0.5

H_{A}: P(\bar{x}_{M}> \bar{x}_{W}) \neq 0.5

Let’s take a two-sided alternate hypothesis.

Here is the null distribution of \frac{\bar{x}_{M}}{\bar{x}_{W}} for N = 10,000.

A vertical bar is shown at a ratio of 1 to indicate that the area beyond this value is the proportion of times S_{i} = 1 in a set of 10,000 bootstrap-replicated means.

The p-value is 0.0466. 466 out of the 10,000 bootstrap replicates had \bar{x}_{M}>\bar{x}_{W}. For a 10% rate of error (\alpha=10\%), we reject the null hypothesis since the p-value is less than 0.05. Since more than 95% of the times, the mean of the total bike counts on Manhattan Bridge is less than that of the Williamsburg Bridge; there is sufficient evidence that they are not equal. So we reject the null hypothesis.

Can we reject the null hypothesis if we select a 5% rate of error?



Is the median of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge?

H_{0}: P(\tilde{x}_{M}>\tilde{x}_{W}) = 0.5

H_{A}: P(\tilde{x}_{M}> \tilde{x}_{W}) \neq 0.5

The null distribution of \frac{\tilde{x}_{M}}{\tilde{x}_{W}} for N = 10,000.

The p-value is 0.1549. 1549 out of the 10,000 bootstrap replicates had \tilde{x}{M}>\tilde{x}{W}. For a 10% rate of error (\alpha=10\%), we cannot reject the null hypothesis since the p-value is greater than 0.05. The evidence (84.51% of the times) that the median of the total bike counts on Manhattan Bridge is less than that of the Williamsburg Bridge is not sufficient to reject equality.



Is the variance of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge?

H_{0}: P(s^{2}_{M}>s^{2}_{W}) = 0.5

H_{A}: P(s^{2}_{M}>s^{2}_{W}) \neq 0.5

The null distribution of \sqrt{\frac{s^{2}_{M}}{s^{2}_{W}}} for N = 10,000. We are looking at the null distribution of the ratio of the standard deviations.

The p-value is 0.4839. 4839 out of the 10,000 bootstrap replicates had s^{2}_{M}>s^{2}_{W}. For a 10% rate of error (\alpha=10\%), we cannot reject the null hypothesis since the p-value is greater than 0.05. The evidence (51.61% of the times) that the variance of the total bike counts on Manhattan Bridge is less than that of the Williamsburg Bridge is not sufficient to reject equality.



Is the interquartile range of the total bike counts on Manhattan Bridge different than that on Williamsburg Bridge?

H_{0}: P(IQR_{M}>IQR_{W}) = 0.5

H_{A}: P(IQR_{M}>IQR_{W}) \neq 0.5

The null distribution of \frac{IQR_{M}}{IQR_{W}} for N = 10,000. It does not resemble any known distribution, but that does not restrain us since the bootstrap-based hypothesis test is distribution-free.

The p-value is 0.5453. 5453 out of the 10,000 bootstrap replicates had IQR_{M}>IQR_{W}. For a 10% rate of error (\alpha=10\%), we cannot reject the null hypothesis since the p-value is greater than 0.05. The evidence (45.47% of the times) that the interquartile range of the total bike counts on Manhattan Bridge is less than that of the Williamsburg Bridge is not sufficient to reject equality.



Finally, is the proportion of the total bike counts on Manhattan Bridge less than 6352, different from that on Williamsburg Bridge?

H_{0}: P(p_{M}>p_{W}) = 0.5

H_{A}: P(p_{M}>p_{W}) \neq 0.5

The null distribution of \frac{p_{M}}{p_{W}} for N = 10,000.

The p-value is 0.5991. 5991 out of the 10,000 bootstrap replicates had p_{M}>p_{W}. For a 10% rate of error (\alpha=10\%), we cannot reject the null hypothesis since the p-value is greater than 0.05. The evidence (40.09% of the times) that the proportion of the total bike counts less than 6352 on Manhattan Bridge is less than that of the Williamsburg Bridge is not sufficient to reject equality.



Can you see the bootstrap concept’s flexibility and how widely we can apply it for hypothesis testing? Just remember that the underlying assumption is that the data are independent. 

To summarize,

Repeatedly sample with replacement from original samples of X and Y -- N times.
Each time draw a sample of size n_{1} from X and a sample of size n_{2} from Y.
Compute the desired statistic (mean, median, skew, etc.) from each
bootstrap sample.
The null hypothesis P(\theta_{x}>\theta_{y})=0.5 can now be tested as follows:
  • S_{i}=1 if \theta_{x}>\theta_{y}, else, S_{i}=0
  • p-value=\frac{1}{N}\sum_{i=1}^{N}S_{i} (average over all N bootstrap-replicated statistics)
  • If p-value < \frac{\alpha}{2} or p-value > 1-\frac{\alpha}{2}, reject the null hypothesis for a two-sided hypothesis test at a selected rejection rate of \alpha.
  • If p-value < \alpha, reject the null hypothesis for a left-sided hypothesis test at a selected rejection rate of \alpha.
  • If p-value > 1-\alpha, reject the null hypothesis for a right-sided hypothesis test at a selected rejection rate of \alpha.

After seven lessons, we are now equipped with all the theory of the two-sample hypothesis tests. It is time to put them to practice. Dust off your programming machines and get set.

If you find this useful, please like, share and subscribe.
You can also follow me on Twitter @realDevineni for updates on new lessons.

Lesson 97 – The Two-Sample Hypothesis Test – Part VI

On the Equality of Variances

using F-distribution

H_{0}:\sigma_{1}^{2}=\sigma_{2}^{2}

H_{A}:\sigma_{1}^{2}>\sigma_{2}^{2}

H_{A}:\sigma_{1}^{2}<\sigma_{2}^{2}

H_{A}:\sigma_{1}^{2} \neq \sigma_{2}^{2}

A good friend is a wall street day trader who has been trading BTC. Lately, he has observed higher volatility in the daily gains/losses, which he associates with celebrities and big banks’ newfound interest. Perhaps they are late to the party, or they quietly got in and are pumping it up.

We wanted to conduct a fun exercise to verify whether the recent period’s variability is more significant than a preceding period. We downloaded BTC price data for the past month and computed daily gains or losses as the difference between closing and opening prices. Here is how the gains/losses look like over the past month.

We highlight two samples, the gains/losses for the most recent week in red and the gains/losses for the week preceding it in blue.

The data from the first sample, i.e., BTC gains/losses for the most recent week, is:
$256, $7991, -$900, -$1357, $2496, $541, -$1094

The variability measured using standard-deviation of the seven values is $3299.

The data from the second sample, i.e., BTC gains/losses for the week preceding the current week, is:
-$429, $323, $2774, $1024, -$393, $986, -$943

Its variability is $1374.

At first glance, it seems evident that the variability in sample 1 is greater than sample2.

But, we know that glances and feelings are not enough. 

We need to establish evidence beyond a reasonable doubt. 

So my friend and I set up a hypothesis testing framework on the equality of variances.

The null hypothesis is that the population variances are equal.

H_{0}: \sigma_{1}^{2}=\sigma_{2}^{2}

To counter this, we write an alternate hypothesis that

H_{A}:\sigma_{1}^{2}>\sigma_{2}^{2}

Then, suppose we know the test-statistic and the null distribution, i.e., the limiting distribution that the test-statistic converges to. In that case, we can compute the p-value and reject H_{0} if it is less than our chosen rate of rejection, i.e., if p-value < \alpha. You and I know the whole process. 

Do you know what the test-statistic is?

.

.

.

Refer to Lesson 73. Can we say that \frac{(n-1)s^{2}}{\sigma^{2}} is a Chi-square distribution with (n-1) degrees of fredom?

So we can write,

\frac{(n-1)s^{2}}{\sigma^{2}}=\chi^{2}

For the two samples which have n_{1} and n_{2} values, we can write the following two equations.

\frac{(n_{1}-1)s_{1}^{2}}{\sigma_{1}^{2}}=\chi_{1}^{2}

\frac{(n_{2}-1)s_{2}^{2}}{\sigma_{2}^{2}}=\chi_{2}^{2}

If we take their ratio, we get

\frac{\frac{(n_{1}-1)s_{1}^{2}}{\sigma_{1}^{2}}}{\frac{(n_{2}-1)s_{2}^{2}}{\sigma_{2}^{2}}} = \frac{\chi_{1}^{2}}{\chi_{2}^{2}}

Let’s move the degrees of freedom terms to the right-hand side.

\frac{s_{1}^{2}/\sigma_{1}^{2}}{s_{2}^{2}/\sigma_{2}^{2}} = \frac{\chi_{1}^{2}/(n_{1}-1)}{\chi_{2}^{2}/(n_{2}-1)}

The ratio that we see on the right-hand side is the ratio of two Chi-square distributions divided by their respective degrees of freedom.

This ratio, \frac{\chi_{1}^{2}/(n_{1}-1)}{\chi_{2}^{2}/(n_{2}-1)} happens to be a special distribution called the F-distribution as presented by George W. Snedecor.

Snedecor named after Sir R.A. Fisher (F-distribution), who first introduced the ratio of sample variances for evaluating whether they are significantly different. He introduced it as the general z-distribution.

The probability distribution function of an F-distribution is:

f_{F}(f) = \frac{(\frac{n_{1}+n_{2}}{2}-1)!(\frac{n_{1}}{n_{2}})^{(\frac{n_{1}}{2})}f^{(\frac{n_{1}}{2}-1)}}{(\frac{n_{1}}{2}-1)!(\frac{n_{2}}{2}-1)!(1+\frac{n_{1}}{n_{2}}f)^{\frac{n_{1}+n_{2}}{2}}}; \hspace{5} 0 < f < \infty

In a later lesson, we will derive this from the first principles. We need to know some concepts on joint distributions to be able to do that. So acknowledge it for now, but look out for its derivation in one of the future lessons.

You can say that the F-distribution is a non-negative distribution since it is the ratio of the two non-negative Chi-square distributions. It is denoted as F_{df1,df2}, i.e., an F-distribution with df1 degrees of freedom in the numerator (corresponding to \chi_{1}^{2}) and df2 degrees of freedom in the denominator (corresponding to \chi_{2}^{2}).

Visually, it looks similar to the Chi-square distribution. See two examples here.

The distribution shown in black has six degrees of freedom, each in the numerator and the denominator. The distribution shown in red has six degrees of freedom in the numerator and 12 degrees of freedom in the denominator. The one with lower degrees of freedom has a fatter tail and more variance — to factor in more uncertainty due to smaller sample sizes.

At any rate, we can now see that

\frac{s_{1}^{2}/\sigma_{1}^{2}}{s_{2}^{2}/\sigma_{2}^{2}} = F_{n_{1}-1,n_{2}-1}

Now, if the null hypothesis is true, we have \sigma_{1}^{2}=\sigma_{2}^{2}; which means, \frac{s_{1}^{2}}{s_{2}^{2}} = F_{n_{1}-1,n_{2}-1}

So the test-statistic is the ratio of the sample variances.

f_{0}=\frac{s_{1}^{2}}{s_{2}^{2}}

We evaluate the hypothesis based on where this test-statistic lies in the null distribution or how likely it is to find a value as large as this test-statistic f_{0} in the null distribution.

For our BTC exercise, the alternate hypothesis H_{A} is \sigma_{1}^{2} > \sigma_{2}^{2}. So, we check how far away from unity is our test-statistic f_{0}, the ratio of the sample variances.

Suppose it lies in the critical rejection region based on our choice of \alpha (the rate of rejection); in that case, the p-value will be small, allowing us to reject the null hypothesis that the population variances are equal.

Let’s try it.

The sample variance, s_{1}^{2}, for the first sample, i..e, BTC gains/losses in the most recent week, is

s_{1}^{2} = \frac{1}{n_{1}-1}\sum_{i=1}^{i=n_{1}}(x_{i}-\bar{x})^{2}=10,884,592

The sample variance for the second sample (BTC gains/losses for the week preceding the current week) is

s_{2}^{2} = \frac{1}{n_{2}-1}\sum_{i=1}^{i=n_{2}}(x_{i}-\bar{x})^{2}=1,887,768

Their ratio, i.e., the test-statistic f_{0} is 5.77.

Since each sample has seven values, one value per day, the null distribution assuming that H_{0} is true is an F-distribution with six numerator and denominator degrees of freedom (F_{6,6}).

It looks like this.

As we can see, the test-statistic (f_{0}=5.77) is in the rejection region if we opt for a rejection rate of \alpha=5\%. The critical value at the 5% level is f_{critical}=4.28. Any value beyond this critical value indicates a deviation far enough from unity to reject the null hypothesis that the variances are equal.

Most statistics textbooks provide a tabulated version of f_{critical} for various \alpha values. Here is an example of how such a table looks for \alpha=5\%. It is a snapshot from Snedecor’s book entitled “Statistical Methods,” published in 1937-1938. Rows correspond to numerator degrees of freedom, and columns correspond to denominator degrees of freedom.

We can also compute the p-value, which, in this case, is the probability of finding a value greater than the test-statistic (P(F > f_{0})), from standard computer-based statistics programs. 

In our case, the p-value is 0.025, lower than the chosen rate of rejection of 5%. There is a less than 5 in 100 chance of drawing a sample having a larger F value from the null distribution at hand.

So we reject the null hypothesis that  \sigma_{1}^{2}=\sigma_{1}^{2}. The population variances are unequal. The variability of the recent period is greater than the variability of the preceding period. 

Perhaps the small, most recent sample is not enough to deduce the conclusion on the changes? Go ahead and test for yourself using more data. CoinDesk has data that goes back as far as October 2013, when the price of BTC was $123.

Will the music stop? Does the party come to an end? How much bitcoin do you have?

If you find this useful, please like, share and subscribe.
You can also follow me on Twitter @realDevineni for updates on new lessons.

Lesson 96 – Wilcoxon’s Rank-sum Test

Part V of the Two-Sample Hypothesis Tests

H_{0}: P(x_{1}>x_{2})=0.5

H_{A}: P(x_{1}>x_{2})>0.5

H_{A}: P(x_{1}>x_{2})<0.5

H_{A}: P(x_{1}>x_{2}) \neq 0.5

United Airlines is the best flight service out there. They always arrived on time or earlier when I traveled with them.

How can you say that? Did you compare it to any other airline? I can claim that Jet Blue is the best flight service. I traveled with them five times; the flight arrived twice ahead of schedule, and the most lengthy delay was only 15 minutes.

I seem to be a clear winner here. My delays are -4 and 0 minutes. 😎

My delays are -1, -3, -2, 15 and 8 minutes. Can you still claim you are the winner?

You seem to be bathing in hypothesis tests, and you know the idea of “beyond statistical doubt.”

Can you prove that United flights tend to arrive earlier than Jet Blue flights? Or rather, can you disprove that there is no significant difference between the delays of United flights and Jet Blue flights?

I can use the standard version of the two-sample t-Test or the version proposed by Welch to verify whether the difference in mean delays is significant or not.

Okay. But do you think it is reasonable to compute the sample statistics of mean and variance on such a small sample? At least I have a sample of five. You only have two.

😕 😕 😕

We could employ the rank-sum test developed by Frank Wilcoxon in 1945.

Go on…

Wilcoxon, in his paper titled “Individual Comparisons by Ranking Methods,” indicated the possibility of using ranking methods to approximate the significance of the differences in means. By ranking methods, he meant to replace the actual numerical data with their ranks.

How exactly is it done?

Let’s employ our small data samples and understand the logic and the procedure of the Wilcoxon’s Rank-sum Test.

But we need to start with defining the null and the alternate hypotheses. Take the delays of United flights to be x_{1} and Jet Blue’s delays to be x_{2}. As the null distribution, we assume that x_{1} and x_{2} are samples from the same distribution.

Then, P(x_{1}>x_{2})=0.5 since they will be intermingled. In other words, if x_{1} and x_{2} are samples from the same distribution, there will be an equal likelihood of one exceeding the other.

Let me reason out the alternate hypothesis. If the null hypothesis is not true, then we could have P(x_{1}>x_{2})>0.5 if the delays of United flights are generally greater than Jet Blue. More observations of United will be towards the right of the observations from Jet Blue if x_{1} is from a distribution that is generally greater than x_{2}.

We could have P(x_{1}>x_{2})<0.5 if the delays of United flights are generally smaller than Jet Blue. More observation of United will be towards the left of (or below) the observations of Jet Blue if x_{1} is from a distribution that is generally lower than x_{2}.

Or, we could say P(x_{1}>x_{2}) \neq 0.5 to establish the fact that United delays are different than Jet Blue delays.

Great. Let me summarize them.

H_{0}: P(x_{1}>x_{2})=0.5

H_{A}: P(x_{1}>x_{2})>0.5

H_{A}: P(x_{1}>x_{2})<0.5

H_{A}: P(x_{1}>x_{2}) \neq 0.5

Now, let’s pool all the data into a single set of seven values and order them from smallest to the largest. I will use red color to write the United’s delay times to know that this is from a different sample.

United
-4, 0
Jet Blue
-1, 3, -2, 15, 8


Pooled and ordered
-4, -2, -1, 0, 3, 8, 15

Can you look at the pooled and ordered data and tell me what the rank of -4 is?

That would be rank 1. The smallest value.

Right. And the rank of 15, the largest value in the pooled set, is 7. When two or more values are the same, then we assign the mean rank to all those values.

Would you agree that if either most of the smallest ranks or most of the largest ranks are coming from United flights, i.e., sample x_{1}, there would be more substantial evidence to reject the null hypothesis?

That is sensible.

So we can think of the sum of the ranks that correspond to values from x_{1} as the test-statistic. In our hypothesis test, the sum of the ranks of the values from United flights is 1 + 4 = 5. United flights have delays of -4 and 0 minutes. These values, in the pooled and ordered data, have ranks of 1 and 4. Their rank-sum is 5.

Test-statistic W = sum of the ranks associated with x_{1}, the variable with a smaller sample size, in the pooled and ordered data.

I get it. We should now check how likely it is to see a rank-sum value of 5 in the null distribution of the rank-sums under the assumption that H_{0} is true.

Yes, that is correct. Let’s outline all possible values of the rank-sums.

We have 2 values from x_{1} and 5 values from x_{2}. If the null hypothesis is true, i.e., if x_{1} and x_{2} are samples from the same distribution, the values will be intermingled and the values in x_{1} can take the following ranks.

Nice table. I can see that the smallest possible rank-sum is 3 when the two values in x_{1} take a rank of 1 and 2. If they are intermingled, it is also equally likely that they take on ranks of 1 and 7, in which case, the rank-sum will be 8.

Yup. Like this, I have written down all possible rank-combinations — from 1-2 that gives a rank-sum of 3 to 6-7 that gives a rank-sum of 13.

There are 21 such combinations.

We can also get the total combinations as \binom{7}{2}. Seven values are arranged two at a time.

Right. More generally, if there are n_{1} values in x_{1} and n_{2} values in x_{2}, the total rank-combinations are \binom{n_{1}+n_{2}}{n_{1}}.

From here, let me deduce the null distribution of the rank-sums. From the table, I can see that the possible values of rank-sums are 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, and 13.

The rank-sum values of 3 and 4 occur only one time.
The rank-sum value of 5 occurs two times.
The rank-sum value of 6 occurs two times.
The rank-sum value of 7 occurs three times.
The rank-sum value of 8 occurs three times.
The rank-sum value of 9 occurs three times.
The rank-sum value of 10 occurs two times.
The rank-sum value of 11 occurs two times.
The rank-sum values of 12 and 13 occur one time.

Here, I also prepared a summary table to shows the frequency, probability, and cumulative probability.

The null distribution looks like this.

Very nicely done! You are showing the possible values of the test-statistic (W) on the x-axis and their relative frequency or the probability P(W) on the y-axis. W ranges from 3 to 13.

More generally, the possible values of W range from \frac{n_{1}(n_{1}+1)}{2} (the smallest rank-sum) to \frac{n_{1}(n_{1}+2n_{2}+1)}{2} (the largest rank-sum).

How did you get that? 😕

The smallest possible rank-sum will occur when all n_{1} values of x_{1} are to the left of x_{2}. It is the sum of the first n_{1} ranks — \frac{n_{1}(n_{1}+1)}{2}.

In the same way, the largest possible rank-sum will occur when all n_{1} values of x_{1} are to the right of x_{2}. To get the rank-sum of the last n_{1} values, we should first compute the rank-sum of the whole set of n_{1}+n_{2} values and subtract the rank-sum of n_{2} values from it.

In our example, the largest rank-sum is 6+7=13, which is the sum of the last two ranks. We can get the sum of 13 by first getting the full rank-sum 1+2+3+4+5+6+7 = 28 and then subtracting the rank-sum of the first five values: 1+2+3+4+5 = 15.

The full rank-sum is the sum of n_{1}+n_{2} integers = \frac{(n_{1}+n_{2})(n_{1}+n_{2}+1)}{2}

The rank-sum of n_{2} values is \frac{n_{2}(n_{2}+1)}{2}. This means that n_{1} values are on the right side (in the pooled and ordered values), and we are only computing the rank-sum of the remaiming n_{2} values.

When we subtract this rank-sum from the full rank-sum we get \frac{n_{1}(n_{1}+2n_{2}+1)}{2}

Ah! That is clever.

Can you also see in your null distribution table, a symmetry around the value of \frac{3+13}{2}=8?

Yes.

Can you get a generalized version of this mid-point rank-sum?

It is the mid-point of the smallest rank-sum of \frac{n_{1}(n_{1}+1)}{2} and the largest rank-sum of \frac{n_{1}(n_{1}+2n_{2}+1)}{2}

0.5*(\frac{n_{1}(n_{1}+1)}{2}+\frac{n_{1}(n_{1}+2n_{2}+1)}{2})=\frac{n_{1}(n_{1}+n_{2}+1)}{2}

Nice! The generalization of the frequency for each rank-sum can also be derived like this. Wilcoxon showed it in his paper.

At any rate, the observed test-statistic for our experiment is 5. So we compute the probability of finding a value as large as 5 — P(W \le 5), in the null distribution. It turns out to be 0.1905.

If we opt for a 5% rate of rejection, we cannot reject the null hypothesis that x_{1} and x_{2} are samples from the same distribution. It is the case even for a 10% rate of rejection. The evidence against the null hypothesis is not convincing.

I will concede 🙁

But I like this method. It seems to have no assumptions on the limiting distributions, is a data-driven approach, and works very well for small sample sizes. The technique is a little tedious for larger sample sizes, isn’t it?

Yes. It is rather tedious to come up with the rank-sum table each time. But there are standardized lookup tables for it. There is also a large-sum approximation to a normal distribution. As you noticed, the null distribution is symmetric. It tends to a normal distribution for larger sample sizes, and we can derive the appropriate formula for the test-statistic. Most statistical programs use this approximation. We will look at how it is done in RStudio at a later time.

Did you notice that your rank-sum should have been 3 for you to reject the null hypothesis at the 5% rejection rate?

Ah. That is correct. It means that United Airlines’ delays should have always been lower than Jet Blue’s delays to disprove that there is no significant difference between the delays of United flights and Jet Blue flights.

When we are applying this test on small sample sizes, it seldom results in a highly compelling case for rejecting the null hypothesis H_{0}. If we add more data, the null distribution smoothens out, and it would be possible to attain lower p-values, which provides a stronger case against H_{0}.

By the way, it looks like we are competing for the last prize. Based on the Department of Transportation’s on-time ranking, United and Jet Blue are close to last. You have some brownie points in claiming that United is better than Jet blue.

If you find this useful, please like, share and subscribe.
You can also follow me on Twitter @realDevineni for updates on new lessons.

error

Enjoy this blog? Please spread the word :)