Hypothesis Testing of proportion-based samples
Part 2 of our Introduction to Hypothesis Testing series.
by Daniel Bray, posted 20/03/2020
In part one of this series, I introduced the concept of hypothesis testing, and described the different elements that go into using the various tests. It ended with a cheat-sheet to help you choose which test to use based on the kind of data you’re testing.
In this second post I will go into more detail on proportion-based samples.
If any of the terms Null Hypothesis, Alternative Hypothesis, p-value are new to you, I’d suggest reviewing the first part of this series before moving on.
What is a proportion-based sample?
In these cases we’re interested in checking proportions. For example 17% of a sample matches some profile, and the rest does not. This could be a test comparing a single sample against some expected value, or comparing two different samples.
Note: These tests are only valid when there are only two possible options; and if the probability of one option is p, then the probability of the other must be (1 – p).
Requirements for the quality of the sample
For these tests the following sampling rules are required:
|Random||The sample must be a random sample from the entire population|
|Normal||The sample must reflect the distribution of the underlying population. For these tests a good rule of thumb is that:
For example: if a sample finds that 80% of issues were resolved in 5 days, and 20% were not, then that sample must have at least 10 issues resolved within 5 days, and at least 10 issues resolved in more than 5 days.
|Independent||The sample must be independent – for these tests, a good rule of thumb is that the sample size is less than 10% of the total population.|
Code Samples for Proportion-based Tests
Compare the proportion in a sample to an expected value
Here we have a sample and we want to see if some proportion of that sample is greater than/less than/different to some expected test value.
In this example:
- We expect more than 80% of the tests to pass, so our null hypothesis is: 80% of the tests pass
- Our alternative hypothesis is: more than 80% of the tests pass
- We sampled 500 tests, and found 410 passed
- We use a 1-sample z-test to check if the sample allows us to accept or reject the null hypothesis
To calculate the p-value in Python:
from statsmodels.stats.proportion import proportions_ztest # can we assume anything from our sample significance = 0.05 # our sample - 82% are good sample_success = 410 sample_size = 500 # our Ho is 80% null_hypothesis = 0.80 # check our sample against Ho for Ha > Ho # for Ha < Ho use alternative='smaller' # for Ha != Ho use alternative='two-sided' stat, p_value = proportions_ztest(count=sample_success, nobs=sample_size, value=null_hypothesis, alternative='larger') # report print('z_stat: %0.3f, p_value: %0.3f' % (stat, p_value)) if p_value > significance: print ("Fail to reject the null hypothesis - we have nothing else to say") else: print ("Reject the null hypothesis - suggest the alternative hypothesis is true")
Compare the proportions between 2 samples
Here we have two samples, defined by a proportion, and we want to see if we can make an assertion about whether the overall proportions of one of the underlying populations is greater than / less than / different to the other.
In this example, we want to compare two different populations to see how their tests relate to each other:
- We have two samples – A and B. Our null hypothesis is that the proportions from the two populations are the same
- Our alternative hypothesis is that the proportions from the two populations are different
- From one population we sampled 500 tests and found 410 passed
- From the other population, we sampled 400 tests and found 379 passed
- We use a 2-sample z-test to check if the sample allows us to accept or reject the null hypothesis
To calculate the p-value in Python:
from statsmodels.stats.proportion import proportions_ztest import numpy as np # can we assume anything from our sample significance = 0.025 # our samples - 82% are good in one, and ~79% are good in the other # note - the samples do not need to be the same size sample_success_a, sample_size_a = (410, 500) sample_success_b, sample_size_b = (379, 400) # check our sample against Ho for Ha != Ho successes = np.array([sample_success_a, sample_success_b]) samples = np.array([sample_size_, sample_size_b]) # note, no need for a Ho value here - it's derived from the other parameters stat, p_value = proportions_ztest(count=successes, nobs=samples, alternative='two-sided') # report print('z_stat: %0.3f, p_value: %0.3f' % (stat, p_value)) if p_value > significance: print ("Fail to reject the null hypothesis - we have nothing else to say") else: print ("Reject the null hypothesis - suggest the alternative hypothesis is true")
In the next post I will focus on hypothesis testing mean-based samples.