Intervals That Exclude Zero but Include Trivial Effects

Direction and Importance Are Different Questions

A regression confidence interval can lie entirely above zero while still including positive effects too small to matter in practice. The interval then supports a positive direction under the model, but it does not establish that the effect is large enough to be important. The same distinction applies to an interval entirely below zero: its direction is negative, yet some values it contains may be too small in magnitude to matter.

To assess this situation, compare the interval with two benchmarks: zero, which separates positive from negative effects, and a practical-importance threshold, which marks effects large enough to be worth attention. The effect and threshold must refer to the same predictor change and use the same outcome units. The threshold should be chosen for the application, not inferred from the confidence interval.

What the Confidence Interval Requires

Consider a simple linear regression model for an outcome \(Y_i\) and predictor \(x_i\):

For the usual \(t\)-based interval, assume the mean outcome is linear in the predictor, the errors have mean zero and a common variance, and observations are independent. For the exact \(t\) procedure, also assume the errors are normally distributed conditional on the predictor values. In this model, \(Y_i\) may be written as \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), with \(\varepsilon_i\sim N(0,\sigma^2)\) under those conditions.

The slope \(\beta_1\) describes the mean outcome change per one-unit increase in \(x\). Often, a one-unit change is not the contrast of interest. For a specified change of \(d\) predictor units, define the effect as \(\theta=d\beta_1\). Its estimate is \(\hat{\theta}=db_1\), where \(b_1\) is the fitted slope; if \(d>0\), its standard error is \(SE(\hat{\theta})=d\,SE(b_1)\). A two-sided \(100(1-\alpha)\%\) confidence interval is

Here \(t^*\) is the positive critical value from a \(t\) distribution with the regression residual degrees of freedom, chosen for confidence level \(1-\alpha\). The interval is for the effect over the stated predictor change, so its units are the outcome units. A 95% procedure produces intervals that cover the fixed effect in about 95% of repeated samples when the model assumptions hold; it does not assign a 95% probability to the effect being inside this particular interval.

Why an Interval Can Exclude Zero but Include Trivial Effects

Suppose the confidence interval is \([L,U]\) with \(0<L<U\), and the practical threshold is \(\delta>0\). The interval excludes zero because its smallest value \(L\) is positive. But if \(L<\delta\), the interval also contains positive values smaller than the threshold. If \(U\geq\delta\), it contains values that meet or exceed the threshold as well. Thus the data can support a positive direction without resolving whether the effect is practically important.

For an entirely negative interval, the corresponding conditions are \(L<U<0\). It excludes zero, but if \(U>-\delta\), values near zero—and therefore negative effects smaller than the threshold in magnitude—remain in the interval. When \(L\leq-\delta\), the interval also includes effects that meet or exceed the threshold in the negative direction.

Worked Examples on the Outcome Scale

A Short Justification for the Directional Reading

The \(t\)-based interval has endpoints \(\hat{\theta}-t^*SE(\hat{\theta})\) and \(\hat{\theta}+t^*SE(\hat{\theta})\). Its lower endpoint is positive exactly when

That condition means the entire interval lies above zero. If the interval is a two-sided 95% interval for the same effect under the same assumptions, excluding zero also corresponds to rejecting the two-sided null hypothesis \(H_0:\theta=0\) at \(\alpha=0.05\). This is evidence against a zero effect in the specified direction; it is not a comparison with the practical threshold. For the interval also to include trivial positive effects, it is enough that its lower endpoint is below \(\delta\). Those conditions can both hold:

The analogous reasoning applies below zero: an interval can have a negative upper endpoint that is still greater than \(-\delta\). In either direction, excluding zero and overlapping the trivial range are compatible statements because zero and the practical threshold answer different questions.

How to Report the Result

A useful report gives the predictor contrast, the estimated effect, the interval, the threshold, and the conclusion about each benchmark. For example: “For a 10-hour increase in study time, the estimated score change was 1.20 points (95% CI 0.493 to 1.907). The interval is entirely positive but includes effects below the 1.5-point practical threshold, so the direction is supported while practical importance remains uncertain.” This wording does not imply causation; a regression association alone does not establish that changing study time would cause the outcome to change.

Keep the threshold in the same units and for the same contrast as the interval. If the interval describes the effect of a 10-unit predictor change, a threshold for a one-unit change cannot be compared directly without converting one of them. Also report enough precision to make clear whether the endpoints cross zero or the threshold; rounding should not create a misleading boundary comparison.

Common Mistake

• Wrong reasoning: “The 95% interval is entirely above zero, so the effect is practically important.” Excluding zero addresses whether zero is compatible with the interval, not whether the effect exceeds a practical threshold.
• Correction: Compare the full interval with both zero and the threshold. If it excludes zero but contains effects with magnitude below \(\delta\), report evidence about direction and uncertainty about importance separately.
• Another wrong reasoning: “Because the interval includes trivial effects, the true effect is trivial.” A confidence interval describes values compatible with the data and assumptions; it does not identify which particular value is true.