Hypothesis testing concepts
Studying this course? Keep your progress in one place.
Key facts
- A hypothesis test assumes and rejects it only if the result is very unlikely.
- Hypotheses are about a parameter ( or ), and gives a single value.
- The critical region is where is rejected; its probability is at most the significance level.
- Two-tailed tests split the significance level between the tails.
- Conclude in context: "sufficient" or "insufficient" evidence, never "proved".
What a test does
It decides whether sample data make a claim about a parameter unlikely enough to reject.
A hypothesis test uses sample data to decide whether a claim about a population parameter is believable. Assume the claim is true, work out how likely the observed result would be, and reject the claim only if the result would be very unlikely.
A test can show evidence for or against a claim, but it never proves it. The conclusion is worded as sufficient or insufficient evidence.
- 1
State hypotheses
and
- 2
Assume H₀ is true
Find the distribution
- 3
Find the critical region
At the significance level
- 4
Decide
Is the result inside?
- 5
Conclude
In context
A test result is not in the critical region. Which conclusion is correct?
Null and alternative hypotheses
Both are statements about a parameter; gives a single value.
The null hypothesis, , is the claim assumed to be true, and gives a single value of the parameter, for example . The alternative hypothesis, , says how the parameter differs: , or .
Define the parameter in context, such as "let be the probability that the coin lands heads".
Null hypothesis
- The claim assumed true
- A single value, e.g.
Alternative
- How the parameter differs
- , or
Which is a correct pair of hypotheses?
Critical region
The values of the test statistic for which is rejected.
The test statistic is calculated from the sample, such as the number of heads ; under its distribution is known, such as . The critical region is the set of values for which is rejected, with edge the critical value. The significance level is the probability of falling in it when is true.
Because the statistic is discrete, use the largest region with probability at most the level. The actual level is usually a little smaller.
. What is the lower-tail critical region at the level?
One-tailed and two-tailed
A one-tailed test has all the level in one tail; a two-tailed test splits it.
In a one-tailed test the alternative has a single direction, or , used when the claim suggests a direction such as "the rate has fallen". In a two-tailed test the alternative is , and a test has in each tail.
For two-tailed at : but , so the region is or , and the actual level is .
One-tailed
- or
- All the level in one tail
- "fallen", "biased towards"
Two-tailed
- Level split: 2.5% each tail
- "has changed"
A die is tested to see whether it gives fewer sixes than expected, with . This is a:
Concluding in context
Reject if the result is in the critical region, then say what it means in context.
If the observed value is in the critical region, reject ; otherwise do not reject it. Then write a conclusion in context that is not over-confident.
For example: "There is sufficient evidence, at the level, that the coin is biased towards heads." or "There is insufficient evidence that the pupil is not just guessing." A larger significance level makes rejection easier.
- 1
Compare
is the observed value in the critical region?
- 2
Decide
in the region: reject H₀; otherwise do not reject H₀
- 3
Conclude in context
say what it means, without being over-confident
In the critical region
- Reject
- Sufficient evidence for
Not in the critical region
- Do not reject
- Insufficient evidence for
A result lies in the critical region at the level. What is the correct conclusion?
Refining models
A model is a claim about a parameter, so new data can test it, and it is refined if rejected.
A model such as "late arrivals are Poisson with mean " is a claim about a parameter, so it can be tested with new data. If the result is in the critical region, the model no longer fits and should be refined, for example by changing the mean.
If is not rejected, there is no evidence against the model and it may be kept, although it has not been proved correct.
- 1
Propose a model
A claim about a parameter
- 2
Collect data
New observations
- 3
Test H₀
At a significance level
- 4
Refine if rejected
Change the parameter or model
Then back to step 1
The test data do not reject for a Poisson model with mean . What can you say about the model?
Try an exam question
A coin is thought to be biased towards heads. It is tossed times and lands heads times. Test, at the significance level, whether the coin is biased towards heads. Use and for .
[5 marks]
- [1], , where is the probability of heads.
- [1]Under , .
- [1].
- [1], so reject .
- [1]There is sufficient evidence, at the level, that the coin is biased towards heads.
That's the notes covered.
Carry on to the next subtopic.
Try a question
Which pair of hypotheses should be used to test the suspicion?
1 mark, marked against the mark scheme
Which pair of hypotheses should be used to test the suspicion?
1 mark, marked against the mark scheme