Aligning Outcome Scales Before Aggregation

Why outcome scales must be aligned

A benefit-risk assessment may place a response rate, a symptom score, and an adverse-event risk side by side. These are not directly comparable numbers: a difference of 5 percentage points, a change of 5 symptom-score points, and a difference of 5 percentage points in risk have different meanings. Adding them as if they shared a scale would produce a result whose value depends on arbitrary choices of units.

Scale alignment makes outcome contrasts easier to interpret and, when appropriate, puts them on a common numerical reference scale. It does not by itself determine the relative importance of benefit and harm. That further step requires explicit judgments about trade-offs; a numerical rescaling must not be mistaken for those judgments.

Start by specifying the contrast for each outcome. For a favorable outcome, a positive treatment-minus-comparator contrast can represent benefit. For an adverse outcome, the same subtraction may represent harm; reverse its direction if the intended scale is harm avoided. For a symptom score, first check whether higher or lower values indicate improvement. Direction must be settled before aggregation, not inferred from the final score.

Three useful forms of alignment

The simplest alignment is a conversion to an interpretable natural unit. A difference in response proportions can be reported in percentage points or as additional responders per 100 patients. An absolute risk difference can similarly be expressed as additional or fewer events per 100 patients over a specified period. Such conversions improve interpretation, but they do not make a symptom-score point equivalent to a patient or an event.

A second option is to retain an outcome's natural units and orient its contrast so that positive values consistently indicate a favorable result. This is often the clearest approach when outcomes are reported separately. A third option is to transform contrasts to a dimensionless reference scale using stated anchors. For a benefit-oriented quantity \(x\), with lower and upper anchors \(L\) and \(U\), one linear transformation is:

Under this mapping, \(x=L\) corresponds to \(z=0\), and \(x=U\) corresponds to \(z=1\). Values between the anchors are linearly interpolated. If \(x\) is a harm measure for which lower values are better, reverse the orientation, for example by mapping \(U\) to 0 and \(L\) to 1. The scale and anchors must be labelled: a value of 0.5 means halfway between the chosen anchors under this linear mapping, not “half as important” or “half as beneficial” in a general clinical sense.

Anchors can be based on a clinically interpretable contrast, such as a prespecified range of response-rate differences or a specified symptom-score improvement. They should be justified for the decision at hand. If data fall outside the anchors, report that fact and state how it is handled; silently clipping values to 0 or 1 changes the analysis.

Worked examples

Why units and anchors matter before aggregation

Suppose an analyst adds the raw values 12 percentage points of additional response and 1.7 symptom-score points of improvement. The numerical total, 13.7, has no coherent unit: it combines percentage points with score points. Changing the symptom instrument from a 0-to-40 scale to a 0-to-100 rescaling could change the total without changing any patients' outcomes. This is why a raw sum of unlike outcomes is not a meaningful aggregate.

Even after transformation to a shared reference range, addition makes a further choice. Consider the two normalized contrasts from the examples: 0.567 for symptom improvement and 0.400 for risk reduction. An equal-weight average would be \((0.567+0.400)/2=0.4835\). That arithmetic is well-defined, but equal weighting is a substantive assumption, not a consequence of putting both numbers between 0 and 1. A different weighting could yield a different summary.

The mathematical reason is straightforward. If the unit for an outcome \(x\) is changed by a positive factor \(c\), then its raw contrast becomes \(c\Delta x\). A sum of raw contrasts therefore changes when the unit changes. With a fixed anchor transformation, the same contrast maps to \(\Delta z=\Delta x/(U-L)\); this removes dependence on the original unit only because the anchors define the denominator. Changing those anchors changes \(z\), so anchor selection remains consequential.

This derivation justifies using explicit anchors to standardize scale, but it does not justify any particular anchors or aggregation weights. The aligned outcomes are inputs to a decision model, not a complete decision rule.

A practical sequence and its assumptions

Before combining outcomes, check that each contrast refers to compatible treatment strategies, populations, and assessment periods. Otherwise the numbers may be individually interpretable but not suitable for a single comparison. The endpoint definition and denominator must also be clear: for example, a response proportion and an event risk should use the intended analysis population and handle intercurrent events consistently.

These steps assume that the chosen contrasts are defined consistently and that the transformation is appropriate for the intended interpretation. A linear map assumes that equal increments in the original quantity receive equal increments on the reference scale. That may be a convenient convention, but it is not automatically clinically realistic. A nonlinear transformation may be needed when the meaning of a change varies across the range; choosing or eliciting such functions is a separate modeling task.

Diagnostics, alternatives, and limitations

A useful diagnostic is to reverse-check every transformation: plug the lower and upper anchors into the formula and confirm that they map to 0 and 1, respectively. Also check the direction using a small example: if an adverse-event risk falls, does the harm-avoided scale increase? Compare the units in every reported contrast, and verify that a “per 100” interpretation uses an absolute difference and the stated time horizon.

Common alternatives answer different questions. Reporting outcomes separately in their natural units preserves direct clinical interpretation and avoids premature aggregation. Converting binary outcomes to counts per 100 improves communication but does not align them with continuous symptom changes. Standardizing by a statistical standard deviation can make outcomes dimensionless, but it depends on the variability in a particular population and does not establish clinical equivalence. Anchor-based scaling can be more interpretable, but its result depends on the selected anchors. None of these approaches alone supplies the trade-off weights needed for an overall benefit-risk judgment.

Further limitations matter in practice. Estimates have uncertainty, and a transformed point estimate does not remove it. If several outcomes are aggregated, their joint uncertainty and any dependence between them may affect uncertainty in the total. A transformation may also conceal meaningful differences if it compresses a wide range or relies on anchors unsupported by clinical context. Report the original-scale results alongside transformed values, and examine whether conclusions are robust to defensible alternatives.