Reporting Continuous Outcomes in Clinical Units

Why Clinical Units Matter

A posterior treatment contrast is most useful when readers can connect its numerical value to the clinical endpoint. “A posterior mean difference of \(-0.42\)” is incomplete if the reader does not know whether this means 0.42 points on a symptom scale, 0.42 percentage points in a laboratory measure, or something else. A report should identify the endpoint, the contrast direction, the units, and the uncertainty.

The goal is not simply to attach a unit label to a number. The chosen scale and summary must preserve what the endpoint measures. A difference in millimetres of mercury is directly interpretable as a difference in blood pressure; a standardized difference measured in standard deviations is less directly tied to a clinical decision. The original clinical scale is generally the clearest one when a difference on that scale has a meaningful interpretation.

The estimand determines what is being contrasted: for example, the mean endpoint value at week 12, or the mean change from baseline to week 12. These are not interchangeable descriptions. Report the endpoint definition, analysis time, population, and whether the contrast concerns endpoint values or changes. If the fitted model adjusts for baseline or other prespecified variables, say so; the resulting contrast is still reported in the endpoint’s units, but its precise interpretation is conditional on the model and estimand.

What to Report

A concise report gives the treatment contrast, a posterior summary, and an uncertainty interval. State whether the summary is a posterior mean or median, identify the interval as a credible interval, give its level, and include units. In Bayesian reporting, a 95% credible interval has 95% posterior probability of containing the parameter, conditional on the stated model, prior, and observed data. It is not a guarantee that the interval contains the true value.

For posterior draws \(\Delta^{(1)},\ldots,\Delta^{(M)}\), calculate a draw-wise contrast, then summarize that contrast distribution. For example, the posterior mean is the average of the contrast draws, and an equal-tailed 95% credible interval is bounded by their 2.5th and 97.5th percentiles. The interval should be calculated from the posterior distribution of the contrast itself. In general, subtracting the endpoints of separate treatment- and control-group intervals does not produce the correct interval for their difference because it ignores their joint posterior distribution.

The subtraction must use the declared direction consistently. If lower values are better, a negative treatment-minus-control contrast may favor treatment. If higher values are better, a positive contrast may favor treatment. State the direction in words rather than expecting readers to infer it from a sign.

Preserving the Endpoint Scale

Some models use a transformed outcome or a parameterization that is not itself in the clinical units. When possible, report the posterior contrast after transforming it to the original scale, and explain the transformation. The arithmetic depends on the transformation. For a linear change of measurement units, let \(Y^*=a+bY\), where \(a\) is an offset and \(b\) is a conversion factor. The offset cancels in a difference:

Thus a contrast and its interval endpoints are multiplied by \(b\) when \(b\) is positive. If \(b\) is negative, the order of the interval endpoints reverses. This is a unit conversion, not a change in the underlying clinical effect.

A standardized mean difference instead expresses a contrast in standard deviation units. It can assist comparison when different studies use different instruments for the same construct, but it is not usually the best primary scale for communicating an individual trial’s clinical effect. Readers cannot directly translate, for example, \(-0.30\) standard deviations into kilograms or symptom-score points without additional information. If a standardized effect is needed for a specific purpose, report it as an additional summary rather than silently replacing the clinical-unit contrast.

For nonlinear transformations, transforming a posterior mean or interval endpoint is not generally equivalent to summarizing the transformed posterior draws. For instance, the mean of transformed draws need not equal the transformation of the mean. When the target clinical quantity is on the original scale, apply the transformation to each posterior draw of the relevant quantity and summarize those transformed draws. State clearly what the transformed contrast represents.

Clinical Meaning and Thresholds

A numerical difference becomes clinically interpretable when it is tied to the instrument, time point, and a relevant benchmark. A minimally important difference or other clinical threshold can provide context, but it should be defined for the endpoint and population at issue. Do not describe a contrast as clinically important merely because its interval excludes zero: statistical direction and clinical importance are different questions.

Suppose a clinically important reduction is defined as at least 0.30 percentage points in HbA1c. Because the contrast is treatment minus control, a reduction of at least that size corresponds to \(\Delta\leq -0.30\) percentage points. The estimate and interval can then be compared with this threshold. If the interval includes values on both sides of the threshold, it communicates uncertainty about whether the effect reaches that benchmark. The interval alone does not give the exact posterior probability of exceeding the threshold; that probability should be computed from posterior draws if it is to be reported.

Always describe the clinical direction. HbA1c is generally interpreted so that a lower value is favorable, but that contextual fact does not change the mathematical definition of treatment minus control. The sign, unit, visit, and benchmark should all agree in the report.

Assumptions, Diagnostics, and Common Pitfalls

Reporting cannot repair a poorly defined contrast or an unsuitable analysis. Before presenting a result, verify that the endpoint and time point are unambiguous, the target contrast matches the estimand, and the reported summary comes from the intended posterior quantity. Confirm that the posterior computation is reliable enough for the reported digits and interval endpoints. If the model is on a transformed scale, confirm that the back-transformation corresponds to the stated clinical quantity. These checks support accurate reporting; they do not replace the substantive assumptions of the model.

A useful final check is to read the result aloud in context: “The treatment group’s mean [endpoint] at [time] was estimated to be [amount] [units] higher/lower than the control group’s.” If the sentence is ambiguous about which group is higher, the contrast direction has not been communicated clearly.

• Omitting units: A value such as \(-5.4\) cannot be clinically interpreted without knowing the endpoint scale.
• Mixing up difference and percent change: A difference of \(-0.35\) percentage points is not a relative reduction of 35% or 0.35%.
• Reversing direction: State whether the contrast is treatment minus control or control minus treatment, and keep that choice consistent.
• Calling a group contrast an individual effect: A difference in group means does not imply that every treated participant changes by that amount.
• Overreading an interval against a clinical threshold: An interval crossing a threshold does not provide the exact posterior probability of crossing it.

The main alternatives serve different purposes. A standardized effect supports comparisons across differently scaled instruments but sacrifices direct clinical units. A ratio or relative change may be appropriate when the clinical question is inherently multiplicative, but it answers a different question than an additive mean difference. Reporting only treatment and control summaries can help readers see the group levels, but it does not replace a posterior contrast and its uncertainty. Choose the primary scale to match the clinical question, and label any additional scale explicitly.

A further limitation is that a single mean contrast may hide variation in response or fail to represent a skewed outcome distribution. It should not be presented as a complete description of all patient outcomes. The report should remain faithful to the prespecified endpoint and estimand; additional distributional summaries may be warranted when the mean alone is not informative.