Why the Direction of a Claim Matters
A research question about a regression coefficient can be directional. For example, a researcher may ask whether an additional hour of training is associated with an increase in average test score, or whether a treatment group’s adjusted mean is lower than a comparison group’s. These are not the same question as whether the coefficient differs from a benchmark in either direction. The direction determines which results count as evidence against the null hypothesis—and therefore which tail of a test distribution determines the p-value.
A one-sided test is appropriate only when the direction is justified by the research question and chosen before examining the estimated coefficient. Choosing whichever direction makes the observed result look more significant changes the test and makes its reported evidence misleading.
Model, Hypotheses, and Test Statistic
Consider a regression model with a coefficient \(\beta_j\), whose value describes the relationship represented by model term \(j\), holding the other included predictors fixed. Let \(\hat{\beta}_j\) be its estimated value and \(SE(\hat{\beta}_j)\) its standard error. A directional question can compare \(\beta_j\) with a benchmark \(c\). The benchmark may be zero or a meaningful nonzero value in the coefficient’s units.
For either test, the observed statistic measures how many estimated standard errors separate the estimated coefficient from the benchmark:
The statistic is positive when the estimate exceeds \(c\) and negative when it falls below \(c\). For the conventional regression t test, assume the linear model is correctly specified for the coefficient being tested, observations have independent errors with constant variance, and the errors are normally distributed conditional on the predictors. Under these conditions, at the null boundary \(\beta_j=c\), the statistic follows a t distribution with the regression residual degrees of freedom, denoted \(df\). The examples below assume these conditions have been checked or are stipulated for the hypothetical analyses.
The p-value uses the tail named by the alternative. For an upper-tailed test it is \(P(T_{df}\geq t)\); for a lower-tailed test it is \(P(T_{df}\leq t)\), where \(T_{df}\) denotes a random variable with a t distribution having \(df\) degrees of freedom. These expressions use the null-boundary distribution. Under the rest of the composite null, the probability of an outcome at least as extreme in the alternative’s direction is no larger than at the boundary.
Why the Tail Must Match the Alternative
The tail rule follows directly from what would count as evidence for the alternative. If \(H_a:\beta_j>c\), an estimate far above \(c\) produces a large positive \(t\), which is unusual under the boundary null. The relevant null probability is therefore the area to the right of the observed statistic. If \(H_a:\beta_j<c\), an estimate far below \(c\) produces a large negative \(t\), so the relevant area is to the left.
This is a direct consequence of the meaning of a p-value: it is the probability, assuming the null boundary is true, of a test statistic at least as far into the alternative’s direction as the observed statistic. It is not a probability that the null hypothesis is true. A small p-value indicates that the observed result would be unusual under the null in the direction specified by \(H_a\).
Worked Example: Testing Whether a Slope Is Positive
In a hypothetical analysis of test scores, a regression coefficient estimates the change in mean score points associated with one additional hour of training, holding the other included predictors fixed. Suppose \(\hat{\beta}_j=1.20\) points per hour, \(SE(\hat{\beta}_j)=0.45\) points per hour, and \(df=24\). The research question, specified before examining the estimate, is whether the slope is positive.
Hypotheses. The benchmark is \(c=0\) points per hour. The hypotheses are \(H_0:\beta_j\leq 0\) versus \(H_a:\beta_j>0\). The hypothetical regression analysis is assumed to satisfy the model conditions stated above.
Statistic.
P-value. This is an upper-tailed test, so the p-value is the probability to the right of \(2.6667\) under a t distribution with 24 degrees of freedom:
At a prespecified significance level of \(\alpha=0.05\), \(0.006749<0.05\), so the result is statistically significant for this upper-tailed test. The data provide evidence that the slope is greater than zero; in this model, the estimated direction is an increase of mean score with additional training hours. This is an association described by the regression model, not by itself evidence that training causes scores to increase.
Worked Example: Testing Whether a Slope Is Below a Benchmark
A hypothetical analyst studies the association between a pollutant concentration and an outcome. The coefficient is measured in outcome units per concentration unit. Before fitting the model, the question is framed as whether the slope is less than \(-0.20\) outcome units per concentration unit. Suppose the fitted coefficient is \(-0.75\), its standard error is \(0.30\), and \(df=9\). Assume the regression conditions are satisfied.
Worked Example: A Lower-Tailed Test Against a Nonzero Benchmark
Hypotheses. Here \(c=-0.20\), so the hypotheses are \(H_0:\beta_j\geq -0.20\) versus \(H_a:\beta_j<-0.20\). The alternative, not the sign of the estimate alone, identifies the tail.
Statistic. Subtract the benchmark from the estimate before dividing by its standard error:
P-value. Since the alternative is lower-tailed, use the area to the left:
At \(\alpha=0.05\), this p-value is approximately equal to, and slightly below, 0.05 for the displayed inputs (the unrounded value is about \(0.04998\)). Thus, the test is just significant at the 5% level. The evidence is marginal and supports the directional claim that the coefficient is below \(-0.20\) outcome units per concentration unit; it does not establish that the coefficient has that value exactly.
Direction Is a Design Choice, Not a Result
A one-sided test makes a focused claim. Its benefit is that evidence in the prespecified direction is assessed using the corresponding single tail. Its limitation is equally important: a result in the opposite direction is not treated as evidence for the stated alternative. The alternative must therefore reflect the question that mattered before the estimate was known, not a choice made after seeing which sign appeared.
Worked Example: A Positive Estimate Does Not Support a Lower-Tailed Claim
Return to the hypothetical coefficient estimate \(1.20\), standard error \(0.45\), and \(df=24\) from the test-score illustration. Suppose, instead, that the research question had genuinely been whether the slope was negative, so \(H_0:\beta_j\geq 0\) and \(H_a:\beta_j<0\). The regression assumptions are again stipulated to hold.
The statistic is unchanged because the estimate, benchmark, and standard error are unchanged:
But the alternative is lower-tailed, so its p-value is the area to the left of this positive statistic. By symmetry of the t distribution, it is the complement of the upper-tail probability reported above:
This result does not support the claim that the slope is negative. It also does not justify changing the alternative to positive after seeing the estimate and reporting the smaller p-value as though that direction had been planned. If the scientific question was whether the slope differs from zero in either direction, the appropriate alternative is two-sided, \(H_a:\beta_j\neq 0\), rather than a post hoc one-sided test.
Decision and Interpretation
A test decision compares the p-value with the significance level \(\alpha\), selected before analyzing the data. If \(p\leq\alpha\), reject \(H_0\) in favor of the specified directional alternative. If \(p>\alpha\), do not reject \(H_0\). The latter outcome means the data did not provide sufficiently strong evidence for that alternative at the chosen level; it does not prove the null or establish the opposite direction.
A useful interpretation includes the coefficient, its units, the benchmark, and the direction tested. For example: “The test provides evidence at the 5% level that the slope is below \(-0.20\) outcome units per concentration unit.” Avoid saying that the p-value is the probability the null is true, or that a nonsignificant result shows the coefficient is zero. The test addresses a specified model coefficient and benchmark; it does not include every possible source of uncertainty or establish causation.
For a symmetric t distribution, a one-sided p-value is half of a two-sided p-value only when the observed statistic points in the direction of the one-sided alternative. When it points the other way, the one-sided p-value is large, as in the third example. The tail must be chosen from the alternative hypothesis, not inferred from a two-sided p-value or selected to obtain a smaller number.
Common Mistakes
- Choosing the tail after seeing the sign. “The estimate is positive, so I will test for a positive effect” is not a valid justification if the direction was not specified in advance. State the scientific direction before examining the results.
- Using the wrong tail. A positive \(t\) does not always call for an upper-tail p-value. The alternative hypothesis determines the tail.
- Writing equality as the entire null. For a one-sided test, the null includes equality and values on the opposite side: \(H_0:\beta_j\leq c\) or \(H_0:\beta_j\geq c\).
- Treating failure to reject as proof of the reverse claim. A large p-value for an upper-tailed test does not establish that \(\beta_j\leq c\); it indicates insufficient evidence for \(\beta_j>c\).
- Calling a directional association causal. A regression coefficient describes a conditional relationship in the model. A one-sided test does not, on its own, show that changing a predictor causes an outcome to change.
Check Your Understanding
For each question, identify the relevant direction before interpreting the result.
- A study asks whether a regression slope exceeds 2 score points per hour. Write \(H_0\) and \(H_a\), and state which tail supplies the p-value.
- A lower-tailed test has \(t=-2.1\) and \(df=15\). Should its p-value be \(P(T_{15}\leq -2.1)\) or \(P(T_{15}\geq -2.1)\)? Explain.
- Why is it misleading to choose an upper-tailed alternative only after observing a positive coefficient estimate?
- A test of \(H_a:\beta_j<0\) gives \(p=0.31\). What can and cannot be concluded at \(\alpha=0.05\)?
- For a test of \(H_a:\beta_j>0\), the observed statistic is negative. Should the p-value be small or large relative to 0.5? Explain using the tail definition.