Asymmetric Thresholds for Benefit and Harm

Why Benefit and Harm May Need Different Thresholds

A change in an outcome can be desirable in one direction and undesirable in the other, but the two directions need not have equal practical importance. For example, a modest improvement in a health score might matter to patients, while a decline of the same size might be especially concerning—or the reverse might be true for a different decision. A single symmetric cutoff around zero can conceal this distinction.

Regression estimates describe changes in an outcome associated with a predictor or predictor contrast. To judge practical importance when benefit and harm have different thresholds, compare the estimate and its confidence interval with two distinct bounds. This comparison is about the size and direction of the estimated effect, not by itself about whether the effect is statistically significant or whether a predictor causes an outcome.

The word negligible here does not mean exactly zero, nor does it mean that an effect has no consequences under every possible decision. It means only that the effect falls between the agreed benefit and harm thresholds. The thresholds should be specified in the outcome’s units and justified for the decision at hand. They may be unequal: for instance, a benefit threshold of 3 points and a harm threshold of 1 point create a much narrower negligible region on the harmful side.

Orient the Effect Before Comparing It

A regression coefficient’s sign has meaning only after the outcome’s direction is clear. If higher outcome values are better, a positive coefficient for an increase in the predictor may indicate benefit. If lower outcome values are better, a negative coefficient may indicate benefit instead. Before applying the thresholds, define \(\theta\) so positive means benefit and negative means harm. Transform both the estimate and the confidence interval consistently if the raw coefficient uses the opposite sign.

For example, in a simple linear regression model \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), the coefficient \(\beta_1\) describes the modeled change in the mean outcome per one-unit increase in \(x\). The model-based interpretation assumes the linear mean relationship is appropriate over the predictor range being discussed, observations are independent, and errors have mean zero and constant variance; the usual exact \(t\)-based interval additionally assumes normally distributed errors. In an applied analysis, the model and its assumptions must be assessed before interpreting its estimate or interval. The examples below take a regression estimate, standard error, and degrees of freedom as given and explicitly assume those conditions are adequately supported for the stated analysis.

The practical thresholds and the interval must refer to the same effect, on the same outcome scale and in the same units. A cutoff stated for a 10-unit predictor change cannot be compared directly with a coefficient stated per one-unit change unless both are put on a common scale. Here, each example supplies an estimate and interval for the particular contrast under consideration; calculating uncertainty for a rescaled contrast is a separate task.

What a Confidence Interval Can Establish

A confidence interval summarizes uncertainty about the effect. Under the interval procedure’s assumptions, a 95% confidence procedure produces intervals that cover the fixed effect in 95% of repeated samples. A particular 95% interval is not a 95% probability statement about where the fixed effect lies. To assess practical importance, inspect the entire interval against both thresholds, rather than checking only whether it contains zero.

The boundary cases follow directly from the definitions: an effect exactly equal to \(B\) meets the benefit threshold, and an effect exactly equal to \(-H\) meets the harm threshold. Therefore, an interval with \(L=B\) is entirely at or beyond the benefit threshold; an interval with \(U=-H\) is entirely at or beyond the harm threshold. By contrast, an interval is wholly in the negligible region only when both endpoints are strictly inside its bounds.

Worked Comparisons

Why the Classification Rule Works

The classification is a direct comparison of the set of values in the interval with the set of values meeting each practical criterion. If the interval is \([L,U]\), every value it contains is at least \(L\) and at most \(U\). Thus, when \(L\geq B\), every included value is at least \(B\), so every included value meets the benefit definition. Similarly, when \(U\leq-H\), every included value is at most \(-H\), so every included value meets the harm definition.

For the interval to be wholly inside the negligible region, its smallest value must be greater than \(-H\), and its largest value must be less than \(B\). Those are exactly the conditions \(L>-H\) and \(U<B\). If any of these comparisons fails, the interval includes a value at or beyond a practical threshold, or spans more than one region. That failure does not prove an important effect exists; it means the interval does not confine the effect to the other region.

Common Mistakes and Practical Reporting

• Assuming equal thresholds by default. A benefit cutoff of 2 points and a harm cutoff of 2 points are not interchangeable with cutoffs of 2 points and 1 point. Use the bounds justified for the decision rather than imposing symmetry.
• Reading the raw coefficient sign without checking outcome direction. A negative coefficient can indicate benefit when lower outcomes are better. State what positive and negative values mean before labeling an estimate beneficial or harmful.
• Classifying from the point estimate alone. A point estimate beyond a practical threshold does not mean the confidence interval is entirely beyond it. Report the estimate and interval together.
• Treating threshold overlap as proof of importance. If an interval reaches the important-harm region, that region remains compatible with the data; the overlap does not establish that the true effect is important harm.
• Calling an interval within the negligible region proof of no effect. It supports only the conclusion that the interval lies between the specified bounds, conditional on the analysis. Smaller effects can still be present.

A useful report names the contrast, gives the estimate and confidence interval in outcome units, states the two thresholds, and describes how the interval compares with both. For example: “For this predictor contrast, the estimated change was \(-1.50\) points (95% CI \(-3.15\) to \(0.15\)); with important harm defined as \(-2\) points or lower and important benefit as 2 points or higher, the interval includes important harm but does not establish it.” This wording makes the direction, uncertainty, and practical criteria explicit without turning an association into a causal claim.