4 Data Analysis: binomial tests

Data Analytics Module
Lecturer: Hans van der Zwan
Handout 04
Topic: binomial tests

Literature
Rumsey D. J. (2010). Statistical Essentials for Dummies. Hoboken: Wiley Publishing. Ismay C. & Kim A. Y. (2019). ModernDive. Statistical Inference for Data Science. https://moderndive.com.

Recommended literature

Preparation class
See module description
Ismay & Kim (2019), sections 10.1 to 10.4

4.1 Testing of hypotheses: the concepts

4.1.1 Introduction: example of formulating a research hypothesis

As part of a research on the economic effects of the outcome of the Brexit referendum in June 2016 the effect on house values in London is investigated. Because of the outcome of the Brexit referendum, the values of houses in London are expected to decrease. This is especially the case in the central districts of London, e.g. the City of Westminster district.
A sub research question is: what is the effect of the Brexit referendum on house values in the City of Westminster district in London.
To investigate the effect of the Brexit referendum on the values of houses in the City of the Westminster district, a comparison is made between the values in January 2016 and the values in January 2017, i.e. a half year before and a half year after the referendum.
As a measurement for the value of the houses, the selling prices of houses sold are used as an indicator. This means that the houses sold are considered to be a random sample of all houses in this district. Due to the outcome of the referendum, a decrease of the house values is expected.
There are many ways to operationalize this assumption. E.g. compare the average selling price in January 2016 with the average selling price in January 2017. Another possibility, and maybe a better choice, is to use the medians. It’s also possible to operationalize the assumption by comparing the proportion of houses sold for which the selling price is above 1 mln GBP. If house prices are decreasing it might be expected that the proportion of houses for which the selling price is more than 1 mln GPB, is decreasing as well. The price paid data for houses sold are collected from the open data at www.gov.uk. In the boxplots in Figure 1 the prices in January 2016 and January 2017 are compared. The median of the house prices in January 2017 is lower than in January 2016, the spread in the prices in January 2017 seems to be lower than in January 2016.
Table 3.1 gives an overview of some sample statistics. Indeed the average selling price in January 2017 is less than in January 2016.

Figure 1. Boxplots selling prices houses (GBP) in City of Westminster district in London. A comparison is made between the selling prices in January 2016 and in January 2017.

Table 2
Summary statistics prices houses sold in London, City of Westminster district in January 20016 and January 2017
year COUNT AVERAGE MEDIAN SD NUMBER_ABOVE_1MLN PERC_ABOVE_1MLN
2,016 333 1,490,100 999,000 2,042,500 164 49.2
2,017 236 1,478,800 1,087,500 1,394,700 123 52.1

If the proportion houses for which the selling price is above 1 mln GPB is used as metric for the house values, no support is found for a decrease in these values. After all the proportion with a selling price above 1 mln GPB increased.
The average selling price decreased. It is not unexpected that the two averages differ, because two more or less random groups are compared. Actually it would be very surprising if the two averages would be exactly the same. A lower mean selling price in January 2017, doesn’t necessarily mean that the average value of all houses in the district in January 2017 is lower than the average value in January 2016. The question is whether this difference between the two means is merely a matter of chance, or whether it is due to an underlying difference between the values in January 2017 and in January 2016. To find out if the difference is just a matter of chance, a so-called t-test can be used. The conclusion would be that the outcome of a t-test does not support the assumption that the mean value of houses in the city of Westminster district in January 2017 (M = 1,478,800; SD = 1,087,500) is less than in January 2016 (M = 1,490,100; SD = 999,000); t(566 ) = -.079; p = 0.469.

4.1.2 Hypotheses testing the concepts

4.1.2.1 Hypotheses testing in court

In a trial there is a statement (hypothesis) the prosecutor wants to prove: ‘the defendant is guilty’. As long as there is no evidence, the opposite is accepted: ‘the defendant is not guilty’. The defendant is found guilty only if the evidence that he is, is ‘beyond reasonable doubt’. After the trial there are four possibilities:
1. The defendant is not guilty and is found not guilty (right decision)
2. The defendant is not guilty, but is found guilty (wrong decision)
3. The defendant is guilty but is not found guilty (wrong decision)
(note: ‘not found guilty’ is not the same as found ‘not guilty’)
4. The defendant is guilty and is found guilty (right decision)

WHAT IS REALLY TRUTH
not guilty guilty
DECISION Acquitted right decision wrong decision,
second order error
Sentenced wrong decision,
first order error
right decision


4.1.2.2 Hypotheses testing in statistics

The principles for statistical testing of hypotheses are the same as these in court. The difference is that we are looking for statistical evidence instead of juridical. The researcher postulates a hypothesis he wants to prove, the so called HA (or H1) hypothesis. As long as there is no statistical evidence that this hypothesis is true, the opposite (the H0 hypothesis) is assumed to be true. After a testing procedure the researcher comes to the decision whether to reject or not reject the H0-hypothesis. In the same way as in court there are four possibilities, shown in the diagram below.

WHAT IS REALLY TRUTH
H0 is right H0 is not right
DECISION Do not reject H0 right decision wrong decision,
second order error
(\(\beta\)-risk)
Reject H0 wrong decision,
first order error (\(\alpha\)-risk)
right decision


In a statistical procedure, we examine whether there is statistical evidence that the data contradict the H0-hypothesis in favor of the HA. If that is the case the H0-hypothesis is rejected. Statistical evidence means that the \(\alpha\)-risk may not be greater than a pre-agreed value (usually 0.05). If the statistical evidence is not strong enough to reject the H0, we stay at the starting point of the procedure (H0 is true) and we do not reject H0. This does not mean that there is statistical evidence that supports H0, it does mean that there is no statistical reason to reject H0.
The way these statistical procedures are used in various statistical tests is discussed in what follows.

4.1.3 Proportion tests or binomial tests

As can be seen in the above example (house values in London) it is sometimes possible to operationalize a research question by using proportions and formulate a hypothesis about these proportions.

4.1.3.1 Proportion test against a standard

In a univariate analysis a test against a standard is in many cases a good technique to operationalize a research question. Some examples.

Example: predicting the outcome of flipping a coin (1)
Consider a person who claims to be clairvoyant (paranormal gifted) so that he can predict the outcome of flipping a coin.2

Experiment: ask the person to predict the outcome if the coin is flipped once.
Questions:
(i) If the predition is correct, would that be seen as support for the claim?
(ii) What if the experiment is flipping the coin twice and both predictions are correct?
(iii) What if the coin is flipped three times and all predictions are correct?
(iv) After how many correct predictions can the claim be honored?

Example: predicting the outcome of flipping a coin (2)
Consider a person who claims to be clairvoyant, and says he can better predict the outcome of flipping a coin than simply by guessing. In other words, the claim is that his predictions are correct in more than 50% of the cases.
Research experiment (data collection): the person is asked to make 100 predictions. Based on the number of correct predictions (k) it will be decided to honor his claim or not.
The hypothesis to test is HA: p > .5 where p is the probability of a correct prediction.
In the procedure we start with the assumption that the claim is not correct, in other words p = .5. To come to a decision the so called prob-value will be calculated, that is the probability to find a result as has been found, or even further away from what is expected, assuming p = .5 (i.e. assuming the claim is not correct). If this prob-value is low, than there are two options:
- the probability of a correct prediction is .5, the person was just lucky in guessing correct a great number of times;
- the probability of a correct prediction is more than .5, and that is why he predicted correct a great number of times.
If the prob-value is lower than a pre-agreed value, mostly .05 is used, the first option is rejected and it is said that the data support the HA-hypothesis.

For instance if he predicts correctly 60 out of 100 times the prob-value equals 0.028. The prob-value has been calculated using this webapplication.

Figure 2. Binomial distribution with n = 100, p = .50. The black area corresponds with the probability of 60 or more successes in 100 trials under the assumption that a probability of a success equals 0.50.

Example: filling packages of sugar
Packs of sugar are filled using a filling machine. Although on average the contents is 1000 gram there is always some variation in the contents of the packs of sugar. In the past 10% of the packages contain less than 995 grams. This was reason to buy a new packing machine. To test if this machine performs better than the old one, a random sample of 100 packs is drawn. From these packs, 6 contain less than 995 grams. Does this sample result give statistical evidence that the machine performs better? In other words: is this sample result enough evidence that in the population less than 10% of the packages contain less than 995 grams.

To answer this question we assume the null-hypothesis to be true and calculate the probability to find 6 or less packs with less than 995 grams under this assumption.

Figure 3 . Binomial distribution with n = 100, p = .10. The black area corresponds with the probability of 6 or less successes in 100 trials under the assumption that the probability of a success equals 0.10.

The prob-value in this case equals .117. In other words, at a .05 significance level, the data do not support the hypothesis that less than 10% op the packs contain less than 995 grams. So the data do not give support to the hypothesis that the new machine is better. This does not mean that the new machine is not better than the old one; it does say that the collected data do not give ‘statistical evidence’ that the new machine is better. If the data is expanded the conclusion can change. E.g. if n = 200 and k = 12 the result would be significant (prob-value = .032).

4.1.3.2 Writing it up

Harvard and APA style do not only apply to references, but also, among other things, to reporting results from hypotheses testing.
In the last example the result can be reported as follows.
Based on a one-sided binomial test the observed values (N = 100, K = 6) do not give significant support to the assumption that less than 10 percent of the packages filled by this machine, contain less than 1000 grams, p = .117.

4.2 Exercise

Excercise 1
Research sub question: did house prices in London increase in January 2019 relative to the prices of houses sold in 2018.
In 2018 the average selling Price in London was 740,824 GBP, the median was 515,000 GPB and 15.4% of the houses sold had a selling price above 1 mln GBP.
(i) This research question has been operationalized by stating the hypothesis that the proportions of houses with a selling price of more than 1 mln GBP in the City of London district in January 2018 was more than 16.3%.
Perform this test and report the conclusion as it would be done in a scientific report.
(ii) This research question can also be operationalized by stating the hypothesis that the proportions of houses with a selling price of more than 510,000 GBP (the median in 2018) in London in January 2018 was more than 50%.
Perform this test and report the conclusion as it would be done in a scientific report.

4.3 Proportion test: comparing two groups

In a statistical analysis it is quite common to compare different groups. For instance the differences between incomes in the profit and the not-for-profit sector, the air quality in different cities, and so on.
Comparing proportions in two different groups can sometimes be used to operationalize a research question about differences between two groups.

Example: influence website background on users
In an experiment data is collected about the influence of background colors on the attractiveness of a website. Attractiveness has been operationalized by the proportion of visitors that click on a button for more information.
The same information was presented on two different backgrounds (I and II). From former research it was expected that website I is more attractive than website II.
Data collected: of the 250 visitors to the website with background I, 40 have pressed a click button to get more information; of the 225 visitors to the website with background II, 25 pressed the button.
The question is whether this data support the hypothesis that background I is more attractive.
In a statistical sence, the hypothesis to be tested is:
HA: pI > pII; where p1 is the proportion visitors of website I that clicks on the button and pII the proportion of website II that clicks on this button.
This is an example of a two sample proportion test.
To find out if the data support the HA-hypothesis, the probability is calculated that assuming H0 is true, a difference between the two proportions will be found as has been found in the collected data.3 With the aid of this webapp this P-value can be found (.061). The conclusion would be: a one-sided independent proportion test did not result in a significant difference between the proportion of visitors of website I that clicked on the button (N = 250, K = 40, proportion = 0.16) with the proportion of visitors of website II who did so (N = 225, K = 25, proportion = 0.111), Z = 1.548, p = .061.

Exercise 2
Use the data in the file HP_LONDON_JAN19.xlsx.
(i) Test if there is a significant difference between the proportion flats of all houses sold in the BARNET district and in the GREENWICH district.
(ii) “The proportion houses with a selling price of more than 1 mln GBP for the category terraced houses is higher than for the category semi-detached houses.”
Test if the data support this statement.

Exercise 3
Operationalize the following research questions using a binomial test.
(i) Do female students at THUAS score better for statistics assignments as male students?
(ii) Did the quality of the service of an IT helpdesk in the company increased after the intervention?
(iii) Did the number of accidents in the Netherlands in which electric bicycles were involved increase in 2018 compared with 2017?

4.4 Homework assignment

Option 1
Researching air quality in the major Dutch cities.
Use a two-sample proporion test to test if the proportion days with an average PM10-value above 30 \(\mu g\) per m3 in Amsterdam is higher than in the city of your choice.
Use a proportion test to test if the air quality in Amsterdam in worse than in The Hague (Den Haag).

Option 2
Use your own data.
Think about a research question which can be operationalized with a proportion test that can be tested using this data set. Perform this test. Write up te result according Harvard/ APA style.


  1. The data set is not a random sample from all registered cars in the Netherlands; it is a random sample from registered cars from three brands, KIA, BMW and AUDI; because of didactic reasons, KIA PICANTO’s are excluded from the sample.

  2. P-values for common tests can be calculated with this web app.