Choosing One-Sided Alternatives Before Seeing the Data

Why the Direction Must Come First

A regression analysis can ask whether the population slope differs from a specified value, or whether it is specifically greater or less than that value. The choice changes the hypothesis test: a one-sided test counts evidence in one direction, while a two-sided test counts evidence in either direction. That choice must follow the research question and be made before looking at the estimated slope or test statistic.

Suppose a research question predicts that a higher value of a predictor is associated with a higher mean response. A positive-slope alternative may fit that question. If the question is simply whether there is a linear association, with no justified direction, the alternative should allow either a positive or negative slope. Choosing whichever direction happens to match the observed estimate makes a nominal one-sided p-value too small too often under the null hypothesis.

Model and Test Logic

For slope inference, the simple linear regression model is \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), where \(Y_i\) is the response for observation \(i\), \(x_i\) is its predictor value, and \(\beta_1\) is the population slope. The slope describes the change in mean response associated with a one-unit increase in the predictor. It does not, by itself, establish a causal effect.

The usual exact t test assumes a linear mean relationship, independent errors with mean zero and a common variance, and normally distributed errors conditional on the predictor values. The observations and sampling or study design must also support the intended inference. Under these conditions, when \(H_0:\beta_1=\beta_{1,0}\) holds, the statistic

has a t distribution with \(df=n-2\). Here \(b_1\) is the estimated slope and \(SE(b_1)\) is its standard error. For a right-sided alternative, the p-value is the probability under the null model of a t statistic at least as large as the observed value. For a left-sided alternative, it is the probability of a statistic at least as small. For a two-sided alternative, it is the probability of a statistic at least as far from zero in either direction.

The p-value is calculated using the alternative specified in advance. At significance level \(\alpha\), the test rejects \(H_0\) when the p-value is no greater than \(\alpha\). A small p-value is evidence against the null in the direction or directions included in \(H_a\); it is not the probability that the null hypothesis is true.

Matching the Alternative to the Question

The alternative hypothesis expresses what the analysis is designed to detect. If the substantive question only predicts an increase relative to a specified slope, a right-sided alternative may be suitable. If it predicts a decrease, a left-sided alternative may be suitable. If either direction would matter, or if there is no defensible directional prediction, use a two-sided alternative.

A one-sided alternative is not simply a way to obtain a smaller p-value. It makes a narrower claim and deliberately treats evidence in the opposite direction as not supporting the stated alternative. For example, a test of \(H_a:\beta_1>0\) does not count a strongly negative slope as evidence for the research claim, even though that result may be important and should be reported. Choosing a direction should be justified by the question and relevant prior scientific reasoning, not by which direction the sample happens to show.

Why Choosing a Direction Afterward Fails

A t statistic can be positive or negative. Under the null hypothesis, its t distribution is symmetric around zero. If a researcher first looks at the statistic and then chooses the tail that makes the result appear more extreme, the resulting procedure has effectively searched in both directions, even though it reports only a one-sided p-value.

The inflation can be shown directly. Let \(c\) be the positive cutoff whose upper-tail probability under the null t distribution is \(\alpha\), so \(P(T\ge c)=\alpha\). If the direction was fixed in advance as right-sided, the chance under \(H_0\) of crossing that cutoff is \(\alpha\). But if the analyst selects the direction to match the sign of the observed statistic, the analyst declares significance whenever \(T\ge c\) or \(T\le -c\). By symmetry, each event has probability \(\alpha\), and the events do not overlap. Therefore,

For example, at a nominal \(\alpha=0.05\), selecting whichever one-sided result looks more favorable after seeing the sign gives a false-positive probability of \(0.10\) in this simple setting, not \(0.05\). More generally, the selected one-sided p-value is the smaller tail probability, and under the null it is twice as likely to fall below a given cutoff up to \(0.5\). The reported p-value no longer has the advertised calibration for a prespecified one-sided test.

This argument assumes the only data-dependent choice is which tail to use. In a real analysis, other unplanned choices can also affect error rates. The central principle remains: a test’s p-value is meaningful only relative to the analysis procedure that was actually planned and carried out.

A Practical Decision Process

Before fitting or inspecting the regression results, write down the population slope being tested, the null value, and the alternative. Then check that the direction in the alternative genuinely matches the scientific question. The following sequence keeps the decision explicit.

Common Mistake and Key Takeaway

Wrong reasoning: “The estimated slope is positive, so I will use a right-sided test. Its p-value is below 0.05, so the association is significant.” This reasoning uses the same data both to choose the direction and to evaluate evidence in that direction. It ignores the chance of seeing an equally extreme result in the negative direction and then selecting that tail instead.

Correction: Choose the alternative from the research question before examining the slope estimate. When the direction was not specified in advance or either sign matters, use a two-sided alternative. If a direction was genuinely specified in advance but the estimate points the other way, do not reverse the test; report that the result does not support the prespecified directional claim and describe the estimate accurately.