Scaling Continuous Covariates for Transition Effects

Why Scale a Continuous Predictor?

A regression coefficient depends on the units used to record its predictor. If a biomarker is measured in single units, its coefficient describes a one-unit difference; if it is recorded in tens of units, the coefficient describes a ten-unit difference. The fitted comparisons need not change, but their numerical expression and practical interpretation do. In a multi-state model, this choice matters separately for each transition whose hazard has a covariate effect.

For instance, a hazard ratio of 1.02 per one-unit increase may sound small even when a clinically relevant 10-unit increase corresponds to a much larger contrast. Conversely, reporting a hazard ratio per one year of age may be less useful than reporting it per decade. Centering answers a different question: it makes zero on the recoded scale correspond to a chosen reference value, often one with direct scientific meaning.

The transformations below are exact algebraic recodings, not new modeling assumptions. Their interpretations still depend on the underlying transition model being appropriate, including its proportional-hazards and linear-predictor assumptions.

Model and Transformation

Here, \(c\) has the same units as \(x\), and \(q\) is a positive number of \(x\)-units per one recoded unit. The recoding combines two operations: subtracting \(c\) centers the predictor, while dividing by \(q\) sets the size of one coded unit. If \(q=1\), only centering occurs. If \(c=0\), the predictor is rescaled without being centered.

Substituting \(x=c+qu\) into the model gives

Thus the coefficient changes to \(q\beta_r\). In a Cox model, the factor \(\exp(\beta_r c)\) can be absorbed into the baseline intensity for transition \(r\); the recoding leaves the fitted relative hazard comparisons unchanged. For any two predictor values \(x_1\) and \(x_2\), the hazard ratio is \(\exp\{\beta_r(x_2-x_1)\}\), so it depends on their difference, not on the chosen center.

This interpretation assumes that the specified model is meaningful over the predictor range being compared: in particular, the log transition intensity is linear in \(x\) and the coefficient does not vary with time under the proportional-hazards specification. Rescaling itself does not establish either assumption. It also does not make a large increment scientifically sensible if that increment lies mostly outside the observed data.

Transforming Estimates and Confidence Intervals

If the original coefficient estimate is \(\hat\beta_r\) with standard error \(SE(\hat\beta_r)\), then the estimate and standard error on the recoded scale are multiplied by \(q\). A Wald confidence interval is transformed by the same positive factor on the coefficient scale. Exponentiating the endpoints gives the interval for the hazard ratio corresponding to a \(q\)-unit increase in \(x\).

The factor 1.96 gives an approximate 95% Wald interval when the coefficient estimate is approximately normal. Exponentiation is increasing, so exponentiate the lower and upper coefficient limits in that order. The resulting interval describes uncertainty in the hazard ratio for the specified increment; it is not a probability statement about where the fixed parameter lies.

The same transformation applies to a variance-covariance matrix. If several coefficients are rescaled, multiply each coefficient by its own scale factor and transform the covariance matrix accordingly. For a single coefficient, its variance is multiplied by \(q^2\), which is why its standard error is multiplied by \(q\), not by \(q^2\). Centering alone does not alter the slope estimate or its standard error.

Transition-Specific Scales

A common recoding can be used for a continuous predictor across transitions even when its coefficients differ by transition. The same scale makes the resulting hazard ratios directly comparable in units of the predictor. It does not force the effects to be equal: each transition retains its own coefficient and uncertainty. If the goal is to report different scientifically meaningful increments for different transitions, each transformation must be stated explicitly so readers can identify what each hazard ratio compares.

Choosing a Scale and Checking the Result

A useful scale is one that gives a hazard ratio for a contrast readers can understand, such as five laboratory units, ten years, or a specified clinically meaningful increase. A useful center is a relevant reference value, such as a prespecified threshold or a typical value within the study population. If zero is already meaningful and the predictor is on a convenient scale, recoding may not improve interpretation.

Before reporting a transformed effect, check the following:

• Units: State both the original units and the increment represented by one recoded unit. “Per unit” is not informative if the unit is unclear.
• Transition: Name the transition to which the coefficient and hazard ratio apply. A transition-specific effect is not automatically a statement about every transition.
• Arithmetic: Multiply the coefficient and its standard error by the same positive \(q\); form the interval on the coefficient scale, then exponentiate both endpoints.
• Data support: Check that the reported increment is represented by enough observations in the relevant transition risk sets. A convenient scale cannot remedy sparse information.
• Model form: Scaling does not diagnose whether a straight-line effect on the log-hazard scale is adequate, nor whether proportional hazards is reasonable.

A second common way to report a continuous effect is a hazard ratio per one original measurement unit. This is mathematically valid, and it can be preferable when that unit is already interpretable. A standardized effect per one standard deviation may help express changes on a common scale across differently measured predictors, but it depends on the chosen standard deviation and may be less directly clinical. Centering is useful for defining the reference value, while rescaling sets the contrast size; neither choice should obscure the underlying units.

Common Mistakes and Limitations

• Changing the coefficient but not its uncertainty. If the predictor is rescaled by \(q\), its standard error and coefficient interval must also be scaled by \(q\). Otherwise the hazard-ratio interval is inconsistent with the reported estimate.
• Thinking centering changes the slope effect. Subtracting a constant changes the reference point, not the coefficient for a one-unit difference. Dividing by \(q\) changes the numerical coefficient because one coded unit now represents \(q\) original units.
• Interpreting the baseline as a population-average hazard. After centering, the baseline corresponds to the centered predictor value and the model’s reference values for other covariates. It is not generally an average hazard or an average person’s hazard.
• Reporting a confidence interval as certainty about the parameter. A 95% confidence procedure has 95% long-run coverage under its assumptions. The calculated interval does not assign a 95% probability to the fixed hazard ratio being inside it.
• Assuming scaling repairs misspecification. A neat per-decade or per-five-unit estimate still relies on the assumed linear log-hazard relationship and proportional hazards. If these assumptions are poor, the transformed coefficient remains an inadequate summary.
• Extrapolating a large increment. A hazard ratio for a 20-unit increase is algebraically available under a linear model, but it may be misleading if few or no individuals span that contrast within the relevant risk sets.

A further technical point is that a negative multiplier reverses the coefficient interval’s endpoint order. The conventions here use a positive increment \(q\), so the transformed interval endpoints remain in their usual order. For ordinary reporting, define the recoded predictor so that increasing values correspond to increasing values of the original predictor.