Statistics and probability is the largest topic in IB Mathematics: Applications and Interpretation HL: the IB suggests 52 of the 240 teaching hours, more than any other AI HL topic. A good part of the HL extension is statistical inference: using a sample to estimate a population value, and testing a claim with a clear, numerical decision rule.
This guide covers the HL inference content: the Poisson distribution, unbiased estimates, confidence intervals for a mean, hypothesis tests based on the binomial, Poisson and normal distributions and on correlation, and Type I and Type II errors. It assumes you know the SL basics; if not, start with our AI SL statistics and hypothesis tests guide, which covers chi-squared and t-tests, so we do not repeat them here. For regular help with the whole course, see our IB Maths AI HL tutoring.
What AI HL adds to SL statistics
- The Poisson distribution, X ~ Po(m), for events happening randomly at a constant average rate. Its mean and variance are both m, and the sum of independent Poisson variables is also Poisson.
- Unbiased estimates. The sample mean x̄ estimates the population mean μ. The unbiased estimate of the population variance is s²ₙ₋₁ = (n / (n − 1)) s²ₙ, where s²ₙ divides by n.
- Confidence intervals for the mean of a normal population, using the GDC.
- Further hypothesis tests: for a proportion using the binomial distribution, for a mean rate using the Poisson distribution, for the mean of a normal population, and for whether the population correlation coefficient ρ is 0.
- Critical values and critical regions, and Type I and Type II errors, including calculating their probabilities.
A note on 2029. In the new AI course (first teaching August 2027, first exams May 2029), the Poisson distribution and some hypothesis tests leave AI HL. Students sitting exams in 2027 and 2028 are on the current course and still need everything in this guide. Our 2027–2029 syllabus guide has the details.
Worked example 1: unbiased estimates and a confidence interval
A battery maker tests a random sample of 10 batteries. Their lifetimes, in hours, are:
12.4, 11.8, 13.1, 12.9, 12.2, 11.5, 12.7, 13.4, 12.0, 12.6
Assume lifetimes are normally distributed.
(a) Unbiased estimate of μ.
x̄ = 124.6 / 10 = 12.46 hours.
(b) Unbiased estimate of σ².
From the GDC's one-variable statistics, sₙ = 0.56604..., so s²ₙ = 0.3204.
s²ₙ₋₁ = (10/9)(0.3204) = 0.356. (The GDC also shows sₙ₋₁ = 0.59666... directly; 0.59666² = 0.356.)
(c) A 95% confidence interval for μ.
σ is unknown, so use the t-distribution with n − 1 = 9 degrees of freedom (TInterval on the GDC).
The interval is 12.0 < μ < 12.9 (3 s.f.), or 12.03 to 12.89 to 4 s.f.
(d) Interpret. "We are 95% confident that the interval from 12.03 to 12.89 hours contains the true mean lifetime." A claim that the mean is 13 hours lies outside the interval, so this sample casts doubt on it.
The interval is centred on x̄ (12.46). Its width depends on the spread, the sample size and the confidence level: a 99% interval would be wider.
Worked example 2: a Poisson test with Type I and Type II errors
A help desk receives calls at an average rate of 6 per hour. After a new online FAQ page is launched, the manager records the number of calls X in one randomly chosen hour, and wants to test at the 5% level whether the rate has fallen.
Hypotheses. H₀: m = 6. H₁: m < 6 (one-tailed).
Under H₀, X ~ Po(6).
(a) Critical region.
P(X ≤ 1) = 0.0174 and P(X ≤ 2) = 0.0620.
The largest region with probability at most 0.05 is X ≤ 1.
(b) The hour has 2 calls. Conclusion?
p-value = P(X ≤ 2) = 0.0620 > 0.05, so 2 is not in the critical region. Do not reject H₀: there is not enough evidence at the 5% level that the call rate has fallen.
(c) Probability of a Type I error.
A Type I error means rejecting H₀ when it is true: P(X ≤ 1 | m = 6) = 0.0174. For a discrete distribution, this is less than the 5% significance level.
(d) Suppose the true rate is now 3 per hour. Find the probability of a Type II error.
A Type II error means not rejecting H₀ when it is false: P(X ≥ 2 | m = 3) = 1 − P(X ≤ 1) = 1 − 4e⁻³ = 0.801.
So with only one hour of data, this test would usually miss a halving of the call rate. That is a typical exam interpretation point: one observation gives a weak test.
Worked example 3: tests for a proportion and a mean
Binomial test. A spinner is claimed to land on red with probability 0.25. In 40 spins it lands on red 16 times. Test at the 5% level whether red is more likely than claimed.
- H₀: p = 0.25. H₁: p > 0.25. Under H₀, X ~ B(40, 0.25).
- p-value = P(X ≥ 16) = 0.0262 < 0.05.
- Reject H₀. There is sufficient evidence at the 5% level that the spinner lands on red more often than 25% of the time.
- Check with critical values: P(X ≥ 15) = 0.0544 and P(X ≥ 16) = 0.0262, so the critical region is X ≥ 16, and 16 is in it.
Normal mean test (σ known). A machine fills bottles with volumes that are normally distributed with standard deviation 4 ml. The target mean is 500 ml. A random sample of 16 bottles has mean 497.8 ml. Test at the 5% level whether the mean differs from 500 ml.
- H₀: μ = 500. H₁: μ ≠ 500 (two-tailed).
- Test statistic: z = (497.8 − 500) / (4 / √16) = −2.2 / 1 = −2.2.
- p-value = 2 × P(Z < −2.2) = 0.0278 (Z-Test on the GDC).
- 0.0278 < 0.05, so reject H₀. There is sufficient evidence at the 5% level that the mean fill differs from 500 ml.
If σ were unknown, you would use the sample's sₙ₋₁ and a t-test instead.
Correlation test. To test H₀: ρ = 0 against H₁: ρ > 0 (or ρ ≠ 0) for bivariate normal data, enter the data in two lists and use the linear regression t-test on the GDC. Read the p-value, compare it with the significance level, and conclude in context about the population correlation.
GDC workflow that saves time
- Write the model first: "Let X be... Under H₀, X ~ Po(6)." This earns method marks before any button is pressed.
- Choose the tool: cumulative binomial or Poisson for discrete tests; Z-Test or T-Test for a mean; TInterval or ZInterval for confidence intervals; the linear regression t-test for correlation. Menu names differ between calculator brands, so practise on your own model. Our GDC skills guide covers the common ones.
- Record what the GDC used: test statistic, p-value, degrees of freedom where relevant.
- Compare and conclude in context, in one sentence.
Common mistakes
- Using sₙ instead of sₙ₋₁ for the unbiased variance estimate. Read which value your GDC labels as which.
- Writing hypotheses about the sample. H₀ and H₁ are about the population parameter (μ, p, m or ρ).
- Wrong tail. "Has fallen" is one-tailed (<); "differs" is two-tailed (≠). For a two-tailed test, double the one-tail probability or compare with half the significance level in each tail.
- Critical regions for discrete tests. Pick the region whose probability is closest to the level *without exceeding it*, and give the actual significance level if asked.
- Mixing up Type I and Type II. Type I: rejecting a true H₀. Type II: failing to reject a false H₀. The Type II calculation uses the *new* parameter value.
- Misreading a confidence interval. It is a statement about the method, not "a 95% chance that μ is in this particular interval".
- Saying "H₀ is true". A test only finds evidence, or not enough evidence, against H₀.
Exam technique: Papers 1, 2 and 3
- A GDC is required on all three AI HL papers, so every question on this page can appear in any of them.
- Paper 1 (short questions): expect a single test or interval with a clear context. Model, p-value, conclusion; no long explanations.
- Paper 2 (longer questions): tests often sit inside a multi-part problem, with a Type II error or "comment on the sample size" at the end.
- Paper 3 (two extended questions, 60 minutes): statistics can be the setting for an investigation, such as building a decision rule step by step or comparing two models. Our HL Paper 3 guide explains how to read the structure.
- Probability models themselves (binomial, normal, Poisson) are covered in our probability distributions guide. Regression with non-linear models fits naturally with our sibling post on logistic models and linearising data.
How Maths Bodhi helps
Our tutors teach AI HL inference as a repeatable routine: model, tool, numbers, conclusion. We practise on your own calculator model, check written conclusions against what examiners reward, and work on the areas where you lose marks. If you are still choosing a course, our sibling post on AA SL vs AI HL may help, and our statistics tutoring page covers wider support.
Home tuition is available in Gurgaon (Gurugram). Students elsewhere in India and abroad learn with us online. You can book a free demo class.
Frequently asked questions
- Which hypothesis tests are in IB maths AI HL?
- As well as the SL chi-squared and t-tests, AI HL includes tests for a proportion using the binomial distribution, for a mean rate using the Poisson distribution, for the mean of a normal population, and for whether the population correlation coefficient is 0, with critical regions and Type I and Type II errors.
- What is an unbiased estimate of variance in IB maths?
- It is s²ₙ₋₁ = (n / (n − 1)) s²ₙ, where s²ₙ divides the sum of squared deviations by n. The sample mean is the unbiased estimate of the population mean.
- How do I find a confidence interval for the mean on the GDC?
- Use TInterval when the population standard deviation is unknown and ZInterval when it is known. Enter the data or the summary statistics and the confidence level, then state the interval and interpret it in context.
- What is the difference between a Type I and a Type II error?
- A Type I error is rejecting H₀ when it is true; its probability is the actual significance level of the test. A Type II error is not rejecting H₀ when it is false; you calculate it using the true parameter value.
- Is the Poisson distribution still in IB maths AI HL?
- Yes, for students sitting exams in 2027 and 2028. It leaves AI HL in the new course, which is first examined in May 2029.
- Do I need to know chi-squared tests for AI HL?
- Yes. AI HL includes all the AI SL content, so the chi-squared tests and the t-test from SL can also be examined at HL.
