Equality and Direction in Null Hypotheses

Why Equality and Direction Matter

A hypothesis test begins with a claim about a population parameter, not with a decision about what the sample appears to show. The wording matters: a claim that a rate is higher than a benchmark is not the same claim as one that it is lower, or simply different. Translating the claim into parameter statements fixes the direction of the question before the data are used.

The null hypothesis commonly includes equality because equality marks the boundary between parameter values that do and do not support a directional claim. For a claim that a parameter exceeds a benchmark, values at or below the benchmark belong to the null; values strictly above it belong to the alternative. Equality is not an arbitrary decoration: it determines which side of the comparison includes the boundary.

Definition: Let $\theta$ be the population parameter of interest, with possible values in a parameter space $\Theta$, and let $\theta_0$ be a specified benchmark. A null hypothesis $H_0$ and alternative hypothesis $H_A$ state competing conditions on $\theta$. The null and alternative parameter sets must be disjoint; together they describe the parameter values under consideration.

For a directional question, the usual forms are:

\begin{array}{lll} \text{Question of increase:} & H_0:\theta\leq\theta_0, & H_A:\theta>\theta_0,\\[2pt] \text{Question of decrease:} & H_0:\theta\geq\theta_0, & H_A:\theta<\theta_0. \end{array}

For a question asking whether the parameter differs in either direction, the usual statements are $H_0:\theta=\theta_0$ and $H_A:\theta\ne\theta_0$. Here the alternative contains values on both sides of the benchmark. The phrase “is different” does not specify a direction, so it should not be converted into only a greater-than or only a less-than alternative.

The Boundary and the Full Null Statement

There are two related ways equality appears in directional testing. The full one-sided null includes the equality boundary and all values on the null side. For example, if the question is whether $\theta$ is greater than $\theta_0$, the full null is $\theta\leq\theta_0$. In many test calculations, the equality value $\theta=\theta_0$ is used as a boundary reference. That calculation convention does not change the stated full null into $\theta=\theta_0$.

The distinction matters because a statement about all values at or below a benchmark is broader than a statement about equality alone. When writing hypotheses, give the full parameter sets that match the question. When a later calculation uses the boundary value as a reference, keep that role distinct from the original statement.

Key takeaway: For “greater than,” put equality and smaller values in $H_0$, and strict greater-than values in $H_A$. For “less than,” put equality and larger values in $H_0$, and strict less-than values in $H_A$. For “different,” use a two-sided alternative.

A Set-Based Justification

The direction rule follows by separating the parameter values that satisfy the research claim from those that do not. Suppose the claim is $\theta>\theta_0$. Every possible value of $\theta$ either satisfies that strict inequality or does not. Failure to satisfy it means $\theta\leq\theta_0$, including equality. Thus the two sets form a complete, nonoverlapping split:

\{\theta:\theta\leq\theta_0\}\cap\{\theta:\theta>\theta_0\}=\varnothing, \qquad \{\theta:\theta\leq\theta_0\}\cup\{\theta:\theta>\theta_0\}=\Theta

The same reasoning applies to a claim $\theta<\theta_0$: the values that fail that strict claim satisfy $\theta\geq\theta_0$. This explains why the equality sign goes into the null for a one-sided claim. If equality were omitted from both sets, the hypotheses would leave out the boundary value; if it were placed in both, the sets would overlap.

A two-sided claim has a different split. The values that are not different from $\theta_0$ are exactly equal to $\theta_0$, while values that are different are either smaller or larger. Therefore equality belongs to the null in that formulation as well, but the alternative extends in both directions.

Identify the parameter.
State what population quantity is being compared, such as a probability, a mean, or a difference between means.

Identify the benchmark.
Write the fixed comparison value as $\theta_0$, using the units and population definition in the problem.

Read the direction literally.
“Higher” means $>$, “lower” means $<$, and “different” means $\ne$.

State the complement in the null.
For a strict directional alternative, the null contains the boundary equality and values on the opposite side.

Worked Examples

Worked Example: A Higher Approval Probability

A hypothetical insurer reviews a new process for approving a specified class of applications. Let $p$ be the long-run probability that an eligible application in this class is approved under the new process. The business question, set before reviewing the sample, is whether the approval probability exceeds $0.72$. In a sample of 250 eligible applications, 198 were approved. State the hypotheses and calculate the observed sample proportion. Do not make a testing decision.

Model and method. Treat each sampled application’s outcome as a Bernoulli observation with approval probability $p$, and assume the 250 outcomes are independent and sampled under the same process. Use the parameter-and-complement method: express the research claim first, then put its complement in the null.

The claim is $p>0.72$, so the alternative is $H_A:p>0.72$. Its complement is $p\leq0.72$, giving the full null:

H_0:p\leq0.72, \qquad H_A:p>0.72.

The observed sample proportion is the number approved divided by the number reviewed:

\widehat{p}=\frac{198}{250}=0.792.

The arithmetic checks because $250(0.792)=198$. The sample proportion is above the benchmark, but that fact does not alter the hypotheses or by itself determine a testing decision. Common mistake avoided: writing $H_0:p\geq0.72$ would put the equality and the larger values on the null side, contradicting the stated question about an increase.

Worked Example: A Lower Mean Payment

A hypothetical benefit program is evaluating whether a revised review process reduces its mean payment per approved claim. Let $\mu$ be the population mean payment, in dollars, for approved claims under the revised process. The comparison benchmark is $\$1{,}250$. For a sample of 40 claims, the total payment is $\$47{,}040$. State the hypotheses and calculate the sample mean payment.

Model and method. Assume the 40 claim payments are an independent sample from the population defined in the question, and that the population has a finite mean $\mu$. Use the direction in the word “reduces”: the research claim is that the mean is less than the benchmark.

Thus $H_A:\mu<1250$. The values that do not satisfy this strict decrease claim are $\mu\geq1250$, so:

H_0:\mu\geq1250, \qquad H_A:\mu<1250.

The sample mean equals the total payment divided by the sample size:

\overline{x}=\frac{47{,}040}{40}=1{,}176.

Check: $1{,}176(40)=47{,}040$. The observed sample mean is $\$74$ below the benchmark because $1{,}250-1{,}176=74$. This is a descriptive comparison, not a conclusion about the population. Common mistake avoided: using $H_A:\mu>1250$ because the sample total or a payment amount is positive. Direction comes from the research question, not from whether the measurements themselves are positive.

Worked Example: A Difference With No Prespecified Direction

A hypothetical study compares the proportion of applications completed within one week under two independent processing arrangements. Let $p_A$ and $p_B$ be the corresponding population proportions. The question is whether the proportions differ, without specifying which arrangement should have the higher rate. Among 120 applications under arrangement A, 54 were completed within a week; among 120 under arrangement B, 42 were. State the hypotheses for the difference $p_A-p_B$, then calculate the two sample proportions and their observed difference.

Model and method. Assume outcomes within each arrangement are Bernoulli observations, the two samples are independent, and each sample represents its stated arrangement. Define the parameter of interest as $\theta=p_A-p_B$. Since “differ” allows either direction, use a two-sided alternative.

“No difference” means $p_A-p_B=0$, and “different” means $p_A-p_B\ne0$. Therefore:

H_0:p_A-p_B=0, \qquad H_A:p_A-p_B\ne0.

Calculate each sample proportion separately:

\widehat{p}_A=\frac{54}{120}=0.45, \qquad \widehat{p}_B=\frac{42}{120}=0.35.

The observed difference, in the stated order A minus B, is:

\widehat{p}_A-\widehat{p}_B=0.45-0.35=0.10.

The difference is positive, as a check against the two sample proportions: A’s observed proportion exceeds B’s by 0.10. The hypotheses remain two-sided because the question was posed without a direction. Common mistake avoided: changing the alternative to $p_A-p_B>0$ after noticing that the sample difference is positive would use the observed data to choose the direction.

Common Traps in Writing the Hypotheses

Putting equality in the wrong hypothesis. For a claim $\theta>\theta_0$, $H_0$ includes $\theta=\theta_0$; the alternative is strict. The matching statements are $H_0:\theta\leq\theta_0$ and $H_A:\theta>\theta_0$.
Reversing the direction. A claim that a loss mean is lower has $H_A:\mu<\mu_0$, not $H_A:\mu>\mu_0$. Keep the parameter and benchmark in the same order as the claim.
Treating “at least” like “greater than.” “At least $\theta_0$” means $\theta\geq\theta_0$, including equality. It is not the strict claim $\theta>\theta_0$.
Using equality-only wording without context. For a one-sided question, the full null includes values on one side of the benchmark as well as equality. Do not silently replace it with equality alone when stating the hypotheses.
Choosing the direction from the sample. The alternative follows the question being investigated. A sample estimate in one direction does not turn an originally two-sided question into a one-sided one.

Key takeaway: Translate the claim into an alternative first. Then state its complement as the null. For a strict greater-than or less-than alternative, the equality boundary belongs to the null; for a two-sided difference question, equality is the null value and the alternative includes both directions.

Check Your Understanding

For each problem, identify the population parameter, write the null and alternative hypotheses, and perform any requested arithmetic.

A hypothetical call center asks whether the fraction of service requests resolved on the first contact is greater than $0.68$. In a sample of 150 requests, 111 were resolved on the first contact. State the hypotheses and calculate the sample proportion. (Answer: $H_0:p\leq0.68,\ H_A:p>0.68;\ \widehat p=0.7400$.)
A hypothetical maintenance program asks whether its mean repair cost per incident is below $\$900$. A sample of 25 incidents has total repair cost $\$21{,}750$. State the hypotheses and calculate the sample mean. (Answer: $H_0:\mu\geq900,\ H_A:\mu<900;\ \overline{x}=\$870$.)
A study asks whether the probability of a late delivery differs between two independent hypothetical routing systems. Let the probabilities be $p_1$ and $p_2$. State the hypotheses for the difference in the order $p_1-p_2$. (Answer: $H_0:p_1-p_2=0,\ H_A:p_1-p_2\ne0$.)
A hypothetical insurer asks whether the probability of a specified policy event is at least $0.04$. Let $q$ be that probability. Express the claim and its strict opposite, and identify the null and alternative for a test of whether the probability is below $0.04$. (Answer: the “at least” claim is $q\geq0.04$; for testing whether it is below, $H_0:q\geq0.04,\ H_A:q<0.04$.)
A hypothetical audit asks whether the mean processing time under Method A is lower than under Method B. Let $\mu_A$ and $\mu_B$ be the population mean times. Define $\theta=\mu_A-\mu_B$, then state the hypotheses and explain where equality belongs. (Answer: $H_0:\mu_A-\mu_B\geq0,\ H_A:\mu_A-\mu_B<0$; equality belongs to the null.)

What You'll Learn