Why the Direction of a Question Matters
A hypothesis test begins with a claim about a population parameter. Sometimes the question is specifically whether a quantity is larger than a benchmark or smaller than it. In other settings, any departure from the benchmark matters, whether the quantity is larger or smaller. These are different research questions, so they call for different alternatives.
The choice cannot properly be made by looking at the sample and then picking the direction that best fits what happened to be observed. The alternative describes what the investigation set out to detect; it is not a label for the direction of the sample result. Decide it from the research question before examining the results.
The parameter and benchmark must be named in context. For example, if \(\mu\) is the population mean change in a measurement, the alternative \(\mu<0\) means the population mean change is below zero. If \(p\) is the population proportion of people experiencing an outcome, \(p>0.05\) means that proportion exceeds five percent. The sign in the alternative expresses the research question, not a conclusion about the data.
Matching the Alternative to the Research Question
Ask what findings would matter before choosing a direction. If only an increase would address the stated research aim, use a greater-than alternative. If only a decrease matters, use a less-than alternative. If a change in either direction could be important, use a not-equal alternative.
A two-sided question is not the same as a question that is vague or undecided. It can be highly specific: for instance, a process should meet a target, and both falling short and exceeding it are undesirable. Conversely, a one-sided question is not justified merely because researchers hope for an improvement. The direction should follow from the scientific or practical question, including which departures would count as relevant.
| Research question | Alternative form | Meaning |
|---|---|---|
| Is the population mean greater than the benchmark? | \(H_a:\mu>\mu_0\) | A directional increase |
| Is the population proportion below the benchmark? | \(H_a:p<p_0\) | A directional decrease |
| Does the population parameter differ from the benchmark? | \(H_a:\theta\ne\theta_0\) | A departure in either direction |
In each case, the null is written with equality to the benchmark: \(H_0:\theta=\theta_0\). The equality convention gives a specific reference value for formulating the test. The alternative then states which departures from that value are of interest. The question of how results are evaluated under the null is separate from choosing the alternative.
Worked Example: Testing for a Reduction in a Measurement
Suppose a hypothetical study asks whether a treatment reduces the mean measurement compared with no change. For each participant, define change as the measurement after treatment minus the measurement before treatment. Let \(\mu\) be the population mean change, in millimeters. What hypotheses match the question?
Step 1: Identify the benchmark. No average change corresponds to \(\mu_0=0\) millimeters.
Step 2: Translate “reduces.” With the stated after-minus-before definition, a reduction produces a negative change. The research question is specifically whether the population mean change is below zero, so the alternative is \(H_a:\mu<0\).
Step 3: State the null. Using the equality convention, \(H_0:\mu=0\). The hypotheses concern the population mean change, not the change for any one participant.
The direction depends on how change was defined. Had change been defined as before minus after, a reduction would correspond to a positive value instead. Defining the measurement and its direction before writing hypotheses prevents a sign error.
Worked Example: A Proportion Above a Benchmark
A hypothetical monitoring program asks whether the population proportion of devices that fail within one month is greater than the accepted benchmark of 5%. In a sample of 200 devices, 14 fail. Let \(p\) denote the population proportion that fail. Write the hypotheses and describe the sample result.
Step 1: State the benchmark and parameter. The benchmark is \(p_0=0.05\), and the parameter is the population failure proportion \(p\).
Step 2: Match the alternative to the question. The question asks whether the failure proportion is greater than 5%, so \(H_a:p>0.05\). The null, by the equality convention, is \(H_0:p=0.05\).
Step 3: Calculate the observed sample proportion.
The sample proportion is 0.07, or 7%. Its difference from the 0.05 benchmark is
Thus the sample proportion is 0.02, or 2 percentage points, above the benchmark. This describes the sample; it does not define the alternative. The greater-than direction was specified by the monitoring question and should have been set before the failure count was examined.
Two-Sided Questions and Meaningful Departures
A two-sided alternative is appropriate when departures in either direction could answer the research question. Suppose a filling process has a target mean of 500 milliliters. If too little product is a problem and too much product is also a problem, the question is whether the population mean differs from 500, not just whether it is below 500.
Worked Example: Checking for Departure in Either Direction
A hypothetical quality-control investigation asks whether the population mean fill volume differs from the 500-milliliter target. Underfilling and overfilling are both relevant concerns. A sample of 40 containers has a mean fill volume of 498.6 milliliters. Let \(\mu\) denote the population mean fill volume.
Step 1: Identify the benchmark and parameter. The target is \(\mu_0=500\) milliliters, and the parameter is the population mean fill volume.
Step 2: Translate the question. Because either lower or higher mean volume matters, the alternative is \(H_a:\mu\ne500\) milliliters. The null is \(H_0:\mu=500\) milliliters.
Step 3: Describe the observed sample without changing the hypotheses. The sample mean differs from the target by
The sample mean is 1.4 milliliters below the target, but the research question still allows departures in either direction. The negative sample difference does not turn the pre-specified two-sided alternative into a one-sided one.
A useful check is to ask whether a result in the opposite direction would still count as an answer to the original question. If both higher and lower values would matter, a one-sided alternative would ignore part of the question. If only one direction has a meaningful interpretation for the investigation, a two-sided alternative may not express the stated aim.
Why the Direction Must Be Chosen Before Seeing Results
If investigators inspect the result first, they may be tempted to choose the alternative pointing toward the observed difference. That choice makes the question depend on the data. It also creates more than one opportunity to find a result that looks unusually far from the benchmark, while pretending that only one direction had been considered. The resulting evidence can then appear stronger than a plan fixed in advance warrants.
A short probability argument illustrates the issue in a symmetric setting. Suppose a statistic \(Z\), under the null, has a continuous distribution symmetric about zero. Choose a positive cutoff \(c\) such that the probability of exceeding \(c\) is \(\alpha\). If the direction is fixed in advance as positive, the upper-tail event has probability \(\alpha\). If, after seeing the statistic, either a value above \(c\) or below \(-c\) is treated as evidence in the direction it happens to point, the probability of crossing one of those cutoffs is
The middle equality uses the fact that the two events do not overlap, and symmetry gives \(P(Z<-c)=P(Z>c)\). This simple argument shows why selecting the direction after looking is not equivalent to having specified one direction in advance. The precise behavior of a particular test depends on its setting, but the general lesson is the same: data-driven direction selection changes the testing plan.
This does not mean the observed direction should be hidden. It is important to report whether a sample result was higher or lower, and that direction helps describe the finding. The distinction is that the observed direction is a result, whereas the alternative direction is part of the question planned before the result was examined.
If a result suggests a new question, it can motivate a later investigation. It should not be used to retroactively present that question as if its direction had been chosen beforehand. A clear report distinguishes a planned analysis from a follow-up prompted by what was observed.
A Reliable Choice in Practice
Specify what population quantity is under study and the value against which it is compared.
Use the research question to decide whether only an increase, only a decrease, or either direction would be meaningful.
Use \(>\) for the specified upper direction, \(<\) for the specified lower direction, or \(\ne\) when both directions matter.
Write the parameter equal to its benchmark, and keep the hypotheses fixed when describing the observed sample.
Common Mistake
Wrong reasoning: “The sample mean was lower than the target, so I will use a lower-tailed alternative.” This confuses the observed sample direction with the research question. If the study was designed to detect any departure, the lower result does not justify discarding the possibility of a higher result.
Correction: Write the alternative from the question before inspecting the data. A pre-specified question about a reduction uses a lower-direction alternative; a question about any difference uses a two-sided alternative. Report the sample’s direction separately, and do not rewrite the hypotheses after seeing it.
Check Your Understanding
For each question, identify the parameter and write an alternative that matches the stated research aim. Keep the proposed alternative fixed even if a hypothetical sample result points the other way.
- A study asks whether the population mean recovery time is shorter than 10 days. What is the alternative, and what does its direction mean?
- A safety team asks whether the population proportion of products with a defect differs from 3%. Which alternative matches the question?
- A sample proportion is 0.08 when the benchmark is 0.10. Does this sample result alone establish that the alternative should be \(p<0.10\)? Explain.
- A process has a target mean of 250 grams, and both overfilling and underfilling matter. Write the null and alternative hypotheses.
- In a symmetric null setting, why can choosing either tail only after seeing the statistic produce a probability of crossing a one-tail cutoff of \(2\alpha\), rather than \(\alpha\)?