Combining Redundant Measures Into a Defensible Score

Why Combine Overlapping Measures?

A data set may contain several measurements that largely describe the same underlying feature. For example, step count and minutes of moderate activity both provide information about physical activity. Including both as separate predictors can make their individual regression coefficients difficult to estimate precisely when the measures overlap. Combining them into one score can simplify the predictor set, but it also changes what the regression asks.

A composite score is not automatically meaningful just because its components are correlated. The measurements must have a defensible shared interpretation, their directions and scales must be handled appropriately, and the weighting rule must make sense for the intended use. A score can summarize a common construct, but it cannot preserve every distinction between its components.

Building a Score That Has a Clear Meaning

Start with the scientific meaning of each measure, not with a calculation. Ask whether they are indicators of the same construct, whether higher values point in the same meaningful direction, and whether a combined value would answer a useful question. Similarity in the data is relevant, but correlation alone does not establish that two measures represent one construct.

When measures use different units, standardizing them can put them on a common scale. For a measurement \(x_j\), its standardized value for individual \(i\) is

Here \(\bar{x}_j\) and \(s_j\) are the sample mean and sample standard deviation of measure \(j\) in a specified reference group. A standardized value says how many sample standard deviations an observation lies above or below that reference mean. Standardization does not make two different concepts interchangeable; it only changes their numerical scales.

If higher values on one measure indicate less of the construct, align its direction before combining. For example, if higher values on one measure mean worse outcomes but higher values on another mean better outcomes, one measure may need to be reversed. The appropriate direction is a substantive decision. It should be set before examining which coding produces a more appealing relationship with the outcome.

A simple equal-weight score for two direction-aligned standardized measures is

This score is the average of the two standardized values. It gives each component equal weight on the standardized scale. Its unit is not minutes, steps, or a raw questionnaire point: it is an average standardized unit relative to the reference group. Equal weighting is transparent, but it is not automatically the scientifically correct choice. Weights should reflect the purpose and meaning of the score, rather than being selected solely because they produce a stronger association with the outcome in the same data.

What Changes in the Regression Model?

Suppose the response is \(Y_i\), and the composite score is used as the only predictor in a linear regression model:

Here \(\beta_0\) is the model intercept, \(\gamma\) is the slope for the composite score, and \(\varepsilon_i\) is the error for individual \(i\). This is an associational model: it describes how the conditional mean response changes across values of the score, not necessarily what would happen if activity were experimentally changed.

For the usual regression-based inference about this model, the conditional mean should be adequately represented by a straight-line relationship with \(S_i\); errors should be independent with mean zero and constant variance; and errors should be approximately normal when normal-theory small-sample inference is used. These conditions concern the score model and its errors. Creating a score does not guarantee that they hold, so the fitted relationship and residuals still need assessment.

The coefficient \(\gamma\) answers a different question from the separate coefficients in a model containing both measures. The distinction can be shown directly. With two standardized components and the equal-weight score \(S_i=(z_{i1}+z_{i2})/2\), a score model has conditional mean

Thus the score model imposes equal slopes for the two components on their standardized scales. It describes the response association with their shared, equally weighted direction. A model with separate coefficients, \(\beta_1\) and \(\beta_2\), allows the two associations to differ. Replacing the components with a score therefore reduces the number of predictor coefficients, but it also removes the ability to estimate those two distinct associations from that model.

Check Whether the Composite Hides Important Information

A composite compresses a profile of measurements into one number. Different profiles can produce the same score, so ask whether the information discarded is important to the scientific question. A score may be useful for a broad summary while being unsuitable for decisions that depend on which component is high or low.

A Practical Evaluation

Before using a composite as a regression predictor, make the scoring rule explicit and assess whether it fits the intended question. The following sequence separates the scientific decision from the arithmetic:

A related pitfall is choosing or changing the weights after repeatedly checking which version gives the largest fitted association with the response. That makes the definition of the predictor depend on the outcome in the same data, so the reported association no longer has the simple interpretation of a pre-specified score. If data-driven weighting is used, the method and the resulting uncertainty need to be addressed; it should not be presented as an ordinary, fixed composite without qualification.

Common Mistakes

• Combining measures because they are correlated. Correlation indicates overlap, not necessarily a common scientific meaning. Explain why the components belong in one score.
• Adding raw values with incompatible units. A sum of minutes and steps has no sensible unit. Standardize or use a justified scoring rule before combining.
• Ignoring opposite directions. If a higher value means more of the construct on one measure and less on another, align the coding before averaging.
• Reading the score slope as a separate component effect. It is the association with the specified composite. In an equal-weight score model, the components are constrained to contribute equally on the chosen scale.
• Treating a near-average score as an ordinary profile. Opposing component values can cancel. Inspect component-level patterns when those distinctions matter.