Statistical Versus Practical Evidence About a Slope

Two Different Questions About a Slope

A regression analysis can provide strong evidence that a population slope is not zero while indicating that the association is too small to matter in a particular application. Conversely, an estimate may look large enough to matter, but be so imprecise that the data do not establish whether the population slope reaches that size. These are different questions, and neither a p-value nor a confidence interval answers both by itself.

The first question is statistical: is there evidence that the population slope differs from zero? The second is practical: are the plausible values of the slope large enough to matter for the intended use? Answering the second requires a meaningful-effect threshold chosen in context, such as the smallest change in an outcome per unit change in a predictor that would be consequential.

For example, if the response is quiz score in points and the predictor is study time in hours, the slope is measured in points per hour. A threshold of \(\delta=0.50\) points per hour says that a change of at least half a point in the expected score per additional study hour would count as practically important for the stated purpose. The threshold must be chosen before interpreting the evidence and should be defensible in the application.

What a Slope P-Value Establishes

For a two-sided test of a zero population slope, the hypotheses are \(H_0:\beta=0\) and \(H_a:\beta\ne 0\). Under the usual regression slope-inference model, the test statistic is

where \(b\) is the estimated sample slope and \(SE(b)\) is its standard error. Under \(H_0\), the statistic is compared with a \(t\) distribution with \(df=n-2\). The p-value is the probability, assuming \(H_0\) and the model, of obtaining a test statistic at least as far from zero as the observed statistic. It is not the probability that \(H_0\) is true.

A small p-value is evidence against a zero slope, but it does not establish that the slope is large enough to matter. The test is calibrated to distinguish zero from nonzero; it does not compare the slope with \(\delta\). With enough precision, even a very small slope can be statistically distinguishable from zero.

A confidence interval gives a range of slope values compatible with the data and model at a stated confidence level. For the usual two-sided \(t\) interval, the formula is

Here \(t^*\) is the positive critical value for the chosen confidence level and \(n-2\) degrees of freedom. The model assumptions needed for slope inference must be appropriate; the interpretations below presume that the regression inference conditions are met. A confidence interval makes it possible to compare the range of plausible slopes with a practical threshold, not just with zero.

Comparing a Confidence Interval With the Threshold

Write the confidence interval as \([L,U]\), where \(L\) is its lower endpoint and \(U\) its upper endpoint. The signs and magnitudes of both endpoints matter. In particular, a boundary-touching interval is different from one that extends strictly across a boundary.

• If \(L>\delta\), every slope in the interval is strictly greater than the positive practical threshold. The interval supports a positive slope that is practically important.
• If \(L=\delta\), every slope in the interval is at least \(\delta\), but the interval includes its endpoint \(\delta\). It supports a slope at least as large as the threshold; it does not establish that every value is strictly greater than \(\delta\).
• If \(U<-\delta\), every slope in the interval is strictly less than the negative practical threshold. The interval supports a negative slope whose magnitude is practically important.
• If \(U=-\delta\), every slope in the interval is at most \(-\delta\), including the boundary value. It does not establish that every value is strictly less than \(-\delta\).
• If \(-\delta<L\) and \(U<\delta\), the interval lies strictly inside the negligible range. It rules out slopes reaching either practical threshold at the stated confidence level.

Boundary equality matters because the definition above counts \(|\beta|=\delta\) as practically important. Thus, an interval contained in \([-\delta,\delta]\) rules out slopes larger in magnitude than the threshold, but if it touches either boundary it does not consist exclusively of practically negligible values. If instead the application defines importance using a strict inequality, the boundary convention should be stated explicitly.

An interval extending strictly across a boundary contains values on both sides of it. For example, if \(L<\delta<U\), it contains slopes below \(\delta\) as well as slopes above \(\delta\). By contrast, if \(L=\delta\), it contains no slopes below that boundary. Similar care applies at \(-\delta\). If an interval includes both negligible slopes and practically important slopes, the data leave practical importance uncertain, even if the interval also provides clear evidence against zero.

Worked Examples

Why the Test and Threshold Answer Different Questions

The distinction can also be seen by comparing the confidence interval with zero and with \(\delta\). For a two-sided test at significance level \(\alpha\), the matching \(100(1-\alpha)\%\) confidence interval excludes zero exactly when the test rejects \(H_0:\beta=0\). That agreement concerns the zero-slope question. It does not mean that excluding zero automatically excludes all practically negligible slopes.

For a positive practical threshold, the interval must be compared directly with \(\delta\). For instance, a confidence interval from \(0.02\) to \(0.15\) excludes zero but lies below a threshold of \(0.50\); a confidence interval from \(0.80\) to \(1.20\) excludes both zero and the practical boundary. Both intervals could correspond to small p-values for testing zero, but only the second places the full interval beyond the stated threshold.

The boundary comparison follows directly from the interval formula. If \(L=b-t^*SE(b)>\delta\), then the entire interval lies above the threshold. Equivalently,

The inequality says that the estimate is more than \(t^*\) standard errors above the practical boundary, not merely above zero. If \(L=\delta\), the corresponding statistic equals \(t^*\); the interval reaches the boundary but does not extend below it. This is why equality should not be described as crossing the threshold or as proving a strictly greater-than-threshold slope.

Common Mistake and Practical Importance

A common mistake is to write, “The slope is statistically significant, so the association is important.” The reasoning confuses evidence that the slope differs from zero with evidence that its magnitude reaches a meaningful level. A corrected conclusion reports both questions separately: for example, “The data provide evidence of a nonzero slope, but the 95% interval lies entirely below the prespecified threshold of 0.50 points per hour.”

The threshold is meaningful only in context. It should use the slope’s units and reflect consequences relevant to the application, rather than being selected because it makes a result appear important. Its interpretation also depends on the scale and range of the predictor: a change per hour may have different practical consequences when the observed predictor values span minutes rather than many hours.

Finally, avoid treating an interval that overlaps a practical boundary as a definite practical conclusion. If it contains both values below the threshold and values at or beyond it, the interval expresses uncertainty about practical importance. If it merely touches the boundary, say so precisely: it includes the threshold, but does not extend strictly across it. Statistical evidence, practical judgment, and uncertainty should be reported as distinct parts of the conclusion.