I have created a decision tree which can be used to select a proper statistical test for a statistical hypothesis. If we have a one sample of data and we want to check a hypothesis about a expected value, we can use z-test. This test assumes a variance of the test data is known. If the variance is not known one sample t-test needs to be used. The other assumption is that data has a normal distribution. z-test is one of the simplest statistical tests so we will use it to explain main principles of the hypothesis testing.
Let’s assume we are asked to solve this example: We have a machine which produces components of a specific length. The desired length is 190 mm. The inaccuracy of the machine is known and it is constant and characterized by standard deviation mm. The machine was set up by an employee and we want to check whether it was set correctly. To check this, we measured length of 20 sample components.
The measured values are in a following table.
Note: I tried to write this article as simple as possible. It contains links to more detail information. It uses no special software. A short tutorial for Excel and Python will be written.
At first, we need to formulate hypotheses. Two hypotheses are usually formulated: a null hypothesis and an alternative hypothesis (or ). The null hypothesis usually has an equal sign and the alternative hypothesis always contradicts the null hypothesis. The hypotheses for our examples are:
- The null hypothesis : The expected value of components is 190 mm. ()
- The alternative hypothesis : The expected value of components is not 190 mm. ()
As we know, the length of components will not be exactly 190 mm because it is affected by the inaccuracy of the machine. Even the average length will not be exactly 190 mm. But the key point of the testing is to say whether the to difference could by explained by the inaccuracy or it must have been caused by an error of the machine setting.
For example, if the average length of the component was 150 mm, it would be obviously caused by error. On the other hand, the average length 190.01 would suggest right setting. But what about 189.6 or 190.9? In these cases, it is impossible to decide out of one’s head and the hypothesis testing comes handy.
Before starting the actual calculation, we need to realize one more think. The outcome of our calculation does not be necessarily right. The reason is we base our decision only on a small sample (20 components), not all of them. For example, we might select a lot of shorter components and then the average length would be significantly lower than 190 mm and we would keen to reject the null hypothesis . This is called Type I error.
On the other hand, another error may happen. If the configuration of the machine was only slightly different (for example 189.99), we may not detect such a small difference. This situation is called Type II error. You can see all possible situation of the table below.
|is true||is false|
|is not rejected||Correct decision||Type II error|
|is rejected||Type I error||Correct decision|
The good news is that one can set the probability of Type I error. The probability of this error is called a level of significance and it is denoted by . On the contrary, the probability of Type II error is unknown.
Computation of the Test
No we will go through the computation itself. We will start with a classical method which consists of following steps:
- Definition of hypotheses
- Selection of a test statistics
- Computation of a critical region
- Computation of a value of the test statistics
- Interpretation of the result
We have already defined the hypotheses so we will go to the second point.
Selection of a Test Statistics
The test statistics is basically a formula. Each statistical test has its own test statistics so we basically select the formula by selecting the test. The test statistics of the z-test is
where denotes an average of the sample data, the hypothetical expected value (from ), denotes the standard deviation and a number of observations in the sample. We will use this formula in the 4th step.
Please note the bigger is the difference between the hypothetical expected value and the average value the further is the value of the statistics from 0.
Computation of a Critical Region
The critical region is used for the decision whether is rejected. There is a simple rule: If the value of the statistics is on element of the critical region, is rejected. Otherwise it is not rejected.
There is a simple logic behind calculation of the critical region: Assuming the is true, the critical region contains the least probable values of the test statistics. To be more specific, if is true then the measured average value will be probably close to the theoretical expected value. On the other hand, the big difference will be improbable. As we know, a big difference between these two values causes the value of the statistics to be significantly different from 0. So the extremely low and extremely high values are improbable and thus they should be in the critical region.
So we will simply cut the least probable value of the statistics. To do this, we need to know a statistical distribution of the statistics. The statistics of the z-test has normal (Gauss) distribution.
Let’s set the level of significance to . Both high and low values of the test statistics are suspicious so we will split the critical region into two parts. The first part will contain the least probable low values and the second part will contain the lest probable high probables.
We will split the the level of significance equally into two regions. So we need to identify the lowest values with total probability 2.5 % and the highest values with total probability 2.5 %. To do this, we will use the quantile function. We will denote the quantile function by . So we can write down a formula for the critical region as:
There are many ways how to get values of the quantile function. We can use tabled values which are part of each textbook, software like Microsoft Excel or programming language like Python or R. Let’s start with the most old-fashion way – the statistical table. We can use the table here. The desired quantile is . Because the normal distribution is symmetric the statistical tables contains values for quantile 0.5 and higher. So we need to get a value for quantile . Now we can find the desired value: 1.96.
The lower border value of the critical region is – 1.96. Now we can write down the critical region:
Computation of a Value of the Test Statistics
The computation is quite easy. We can substitute , and . The arithmetical mean of values is (you can check it in Excel or with a calculator). So the value of the statistics is
Interpretation of the Result
The interpretation is simple. The value of the statistics is not element of the critical area so we do not reject (at ).
It is never said that . As we said earlier, our result may be wrong because of possibility of Type II error. So we do not know a probability that our outcome is true.
Conclusion and Other Resources
This example was quite simple and many more things may be shown: one-tailed tests, calculation of p-value and tests for many other hypothesis.