Critical Regions and Error Types in Hypothesis Testing

Every time a researcher runs a test, they are making a decision under uncertainty. They cannot study the entire population, so they rely on a sample. And because a sample is only a slice of reality, the conclusion can sometimes be wrong even when the maths is done perfectly. Hypothesis testing builds a clear set of rules to manage this uncertainty. Two of the most important ideas in those rules are the critical region (the zone that tells you when to reject your assumption) and the two types of errors that can sneak into your conclusion. Understanding both is what separates someone who merely calculates a test statistic from someone who actually understands what their result means.

The setup before any test begins 🔗

Before getting to critical regions, it helps to fix two ideas that everything else depends on. In any hypothesis test you start with a null hypothesis (written as H₀), which is a default position assuming there is no effect, no difference, or no relationship in the population. Against it you set the alternative hypothesis (H₁), which claims the opposite, that a real effect or difference does exist. The entire test is a structured way of deciding whether the sample data gives you enough reason to reject H₀ in favour of H₁. You never “prove” H₁ directly. You only check whether the evidence against H₀ is strong enough to abandon it.

This framing matters because the conclusions of a test are always phrased carefully. You either reject the null hypothesis or you fail to reject it. You do not “accept” it. That small wording difference is the root of one of the errors we will discuss later.

Understanding critical regions 🔗

The critical region is the heart of the decision rule. A critical region, also known as the rejection region, is the set of values of the test statistic for which the null hypothesis is rejected. In other words, it is a specific zone marked out on a probability distribution. If your calculated test statistic lands inside that zone, you reject H₀. If it lands outside, in what is called the acceptance region, you fail to reject H₀.

The logic behind this is simple once you see it. When the null hypothesis is true, the test statistic should usually land near the centre of the distribution. Extreme values, far out in the tails, are unlikely under H₀. So if your result is extreme enough to land in the tails, that is treated as strong evidence that the null hypothesis is probably not true. The critical region is exactly that tail area we agree to treat as “too unlikely to ignore.”

How the level of significance defines the region 🔗

This is where the level of significance, written as alpha (α), enters. The level of significance is the probability we set in advance for rejecting a true null hypothesis. It is also described as the size of the critical region. The most common values chosen are 0.05, 0.01, and 0.10. When you pick α = 0.05, you are deciding that the critical region will cover 5% of the area under the curve. Land in that 5%, and you reject H₀.

The boundary between the acceptance region and the critical region has its own name: the critical value. The critical value depends on the level of significance chosen and on whether the alternative hypothesis points in one direction or two. For example, in a standard normal distribution with α = 0.05, the z-score that marks off 5% of the area is about 1.645 on one side, or 1.96 when the 5% is split across both tails. Critical values are read off standard statistical tables such as the z-table, t-table, or chi-square table.

One-tailed and two-tailed regions 🔗

The critical region does not always sit in the same place. Its position is decided by the alternative hypothesis. If H₁ claims a parameter is simply “different” from a value, the test is two-tailed and the critical region is split across both tails of the distribution. If H₁ claims the parameter is specifically “greater than” or “less than” a value, the test is one-tailed and the entire critical region sits in a single tail. Choosing the right type before collecting data is important, because it changes where the critical value falls and therefore changes your decision.

Why errors are unavoidable 🔗

Here is the uncomfortable truth about hypothesis testing: because decisions are based on probability, mistakes can happen even when every step is followed correctly. The sample might just happen to be unusual. This is not a failure of the researcher; it is a built-in feature of working with samples instead of entire populations. Statisticians classify these mistakes into two clean categories, and the beauty of the framework is that both can be measured and controlled.

The cleanest way to see this is a simple decision table. The reality is either that H₀ is true or false. Your decision is either to reject H₀ or not. That gives four combinations: two correct outcomes and two errors.

Type I error: the false positive 🔗

A Type I error occurs when you reject a null hypothesis that is actually true. It is the error of incorrectly concluding that a significant effect or difference exists when in reality there is none. This is why a Type I error is also called a false positive. You raised an alarm, but there was no fire.

The probability of committing a Type I error is exactly the level of significance, alpha. This connection is direct and deliberate. When you set α = 0.05, you are accepting a 5% chance of wrongly rejecting a true null hypothesis. That is why the critical region and the Type I error are two sides of the same coin: the critical region is the slice of outcomes where you will reject H₀, and if H₀ is actually true, landing in that slice is precisely the mistake.

What a false positive looks like in practice 🔗

Consider a pharmaceutical trial. The null hypothesis says a new drug has no effect. A Type I error here means concluding that the drug improved symptoms when in fact it did not, with the apparent improvement coming from random chance or measurement noise. The consequences can be serious: an ineffective treatment might be approved, money is spent, and patients receive something that does not work. In a courtroom analogy that statisticians often use, a Type I error is like convicting an innocent defendant.

Type II error: the false negative 🔗

A Type II error is the opposite mistake. It occurs when you fail to reject a null hypothesis that is actually false. In short, it is a failure to reject a null hypothesis that is false. Because the test missed a real effect, this is called a false negative. The fire was real, but the alarm stayed silent.

The probability of making a Type II error is denoted by beta (β). Notice that we say “fail to reject” rather than “accept.” A Type II error is not the same as accepting the null hypothesis, because a hypothesis test can only tell you whether to reject it. It means failing to conclude there was an effect when there actually was one.

What a false negative looks like in practice 🔗

Return to the drug trial. If the test fails to reject the null hypothesis and concludes the drug is not effective, when in fact it does work, that is a Type II error. A genuinely useful treatment gets shelved. Importantly, a Type II error is only possible when the null hypothesis is actually false; if the null hypothesis is true, only a Type I error can occur. The two errors live in different worlds, defined by what reality actually is.

The trade-off between the two errors 🔗

Students often ask why we cannot simply make both errors as small as possible. The answer reveals one of the most important relationships in statistics. For a fixed sample size, the chances of these two errors are inversely related: decreasing the Type I error rate increases the Type II error rate, and vice versa.

The reason is the critical region itself. If you tighten alpha, say from 0.05 down to 0.01, the critical region shrinks. A smaller rejection zone makes it harder to reject the null hypothesis, which protects you against false positives. But that same caution makes you more likely to miss a real effect, increasing beta. Loosen alpha and the opposite happens. You cannot squeeze both down at once by adjusting alpha alone.

Statistical power and the role of sample size 🔗

Closely tied to the Type II error is the idea of statistical power. Power is the probability of correctly rejecting the null hypothesis when it is actually false, and it equals 1 minus beta. A test with low power struggles to detect real effects, which means even a well-designed study can waste time and resources if power was never considered.

So how do you reduce both errors together? The key lever is sample size. If you want both alpha and beta to decrease at the same time, you increase the sample size. A larger sample gives more precise estimates, which sharpens the test’s ability to tell a true null from a false one. This is exactly why sample size calculations are done at the design stage of a study rather than after. A common target is to set alpha at 0.05 and aim for a power of 0.80, meaning beta is held at 0.20. This convention reflects a deliberate balancing of the two risks.

Choosing which error matters more 🔗

Since you cannot eliminate both risks cheaply, you must decide which error is more damaging in your specific situation. This is a judgement call, not a formula. In a medical screening test for a serious disease, a false negative (Type II) could mean a sick patient is sent home untreated, so you might tolerate more false positives to avoid missing real cases. In a criminal trial, the legal system deliberately keeps the Type I error low, preferring to risk acquitting a guilty person rather than convicting an innocent one. As one widely cited statement puts it, good research design assesses the relative risk of each type of error and then strikes an appropriate balance between them. The right balance depends entirely on the cost of being wrong in each direction.

Bringing it together 🔗

The three ideas in this post are really one connected system. The critical region is the decision zone, its size is fixed by the level of significance (alpha), and alpha is itself the probability of a Type I error. The Type II error (beta) is the flip side, controlled largely by sample size and study design. Once you see how the critical region directly sets your false positive rate, and how that rate trades off against your false negative rate, hypothesis testing stops being a mechanical procedure and becomes a thoughtful exercise in managing risk. Every test is a bet, and these concepts tell you exactly what you are betting on.

What do you think? If you were designing a study to test whether a low-cost water purifier removes harmful bacteria, would you set a stricter alpha to avoid false alarms, or accept a higher Type I error rate to make sure you never miss a purifier that actually works? And in fields you care about, can you think of a real decision where a false negative would be far more costly than a false positive?

References

1. https://www.mas.ncl.ac.uk/ask/numeracy-maths-statistics/statistics/hypothesis-testing/critical-region-and-confidence-interval.html
2. https://medium.com/@nandiniverma78988/understanding-the-critical-region-and-significance-level-in-hypothesis-testing-31c8eb9a0f85
3. https://www.jove.com/science-education/v/13610/critical-region-critical-values-and-significance-level
4. https://stats.libretexts.org/Bookshelves/Applied_Statistics/An_Introduction_to_Psychological_Statistics_(Foster_et_al.)/07:__Introduction_to_Hypothesis_Testing/7.05:_Critical_values_p-values_and_significance_level
5. https://www.geeksforgeeks.org/maths/type-i-and-type-ii-errors/
6. https://www.scribbr.com/statistics/type-i-and-type-ii-errors/
7. https://en.wikipedia.org/wiki/Type_I_and_type_II_errors
8. https://pmc.ncbi.nlm.nih.gov/articles/PMC2850991/
9. https://www.simplypsychology.org/type_i_and_type_ii_errors.html
10. https://stats.libretexts.org/Bookshelves/Introductory_Statistics/Inferential_Statistics_and_Probability_-_A_Holistic_Approach_(Geraghty)/10:_One_Population_Hypothesis_Testing/10.07:_Type_II_Error_and_Statistical_Power
11. https://online.stat.psu.edu/stat200/lesson/6/6.5
12. https://denvirlab.marshall.edu/MS2/error.html
13. https://cxl.com/blog/statistical-power/