Coding and Scaling Covariates for Allocation Models

Why Coding Matters in an Allocation Model

An adaptive outcome model uses baseline information to describe expected outcomes for treatment and patient profiles. A model’s coding determines how its coefficients are expressed: a coefficient may compare a category with a reference group, or describe the outcome change for a one-unit covariate increase. These choices affect interpretation and numerical stability. They do not necessarily change the fitted predictions used by an allocation procedure.

The key distinction is between changing the representation of the same model and changing the model itself. A one-to-one change of coordinates, with coefficients transformed accordingly, preserves the model’s possible predictions. Omitting a category indicator, changing a covariate’s functional form, or leaving old coefficients in place after changing the coding can alter predictions. Since allocation probabilities may depend on model-based predictions, this distinction matters operationally as well as statistically.

Reference and Indicator Coding

For a binary variable, coding one category as 1 and the other as 0 is indicator coding; the 0 category serves as the reference. For a variable with more than two categories, the same idea uses one indicator for every nonreference category. This is often called treatment or dummy coding. The word “treatment” in that name refers to the contrast coding, not necessarily to a clinical treatment.

With an intercept, including indicators for all \(K\) categories creates exact linear dependence: for every participant, the \(K\) indicators sum to 1, which is also the intercept column. A model cannot uniquely estimate all \(K\) category coefficients and the intercept from this redundant representation. Use \(K-1\) indicators with an intercept, or use all \(K\) indicators without an intercept when that parameterization is appropriate.

Centering and Scaling Continuous Covariates

A continuous baseline covariate can be used in its original units or transformed before it enters the model. Centering subtracts a chosen value \(c\), commonly the sample mean or a clinically meaningful value. Scaling divides by a positive constant \(s_x\), such as a sample standard deviation or a meaningful unit increment. These operations change coefficient units and often make the intercept easier to interpret.

In a linear predictor \(\beta_0+\beta_x x_i\), the intercept is the predicted value at \(x_i=0\), which might be outside the relevant range. Rewriting the same predictor in terms of \(z_i\) gives

Thus the transformed coefficients are \(\beta_0^*=\beta_0+c\beta_x\) and \(\beta_z=s_x\beta_x\). The predictions are unchanged, while the transformed intercept describes \(x_i=c\), and the transformed slope describes the predictor change for a \(s_x\)-unit increase in the original covariate. In a generalized linear model, the same identity applies to the linear predictor; applying the inverse link then gives identical predicted means or probabilities.

What Changes Predictions—and What Does Not

A useful way to check a coding change is to write down the linear predictor before and after the change and substitute the new variables. If the expressions are algebraically identical for every possible covariate profile, then the model’s fitted predictions are identical when coefficients are transformed consistently. This is a statement about the model’s representation, not a claim that every estimation procedure or allocation algorithm will behave identically under all implementations.

For ordinary unpenalized regression with the same observations and a full-rank, equivalent design matrix, changing from one full set of coordinates to another leaves the fitted values unchanged. The coefficient values and interpretations can differ. The justification is direct: each predictor value represented by the original model is also represented by the transformed model, and vice versa. Therefore both models minimize the same residual criterion over the same set of possible fitted-value vectors. For a generalized linear model, the same equivalence holds when the linear predictors attainable under the two parameterizations are the same.

Interactions and Practical Model Checks

When a model includes treatment-by-covariate interactions, transformations must be applied to the interaction terms as well as to the main effects. For example, with treatment indicator \(A_i\) and centered covariate \(z_i=(x_i-c)/s_x\), a linear predictor containing \(\beta_0+\beta_A A_i+\beta_x x_i+\beta_{Ax}A_ix_i\) can be rewritten using \(z_i\). Since \(x_i=c+s_xz_i\), the treatment coefficient becomes \(\beta_A+c\beta_{Ax}\), and the coefficient of \(A_iz_i\) becomes \(s_x\beta_{Ax}\). Retaining the old interaction coefficient after scaling would generally change predictions.

A common convention is to include the lower-order terms whenever an interaction is included. With that structure, centering changes the coefficient interpretations but preserves the same fitted relationships when all corresponding coefficients are transformed. If a model omits lower-order terms or imposes other constraints, equivalence must be checked from the actual model formula rather than assumed.

In practice, check the following before using a transformed baseline variable in an adaptive outcome model:

• Category map: confirm that each recorded category maps to exactly one intended indicator pattern and that the reference category is explicit.
• Redundancy: with an intercept, do not include every category indicator unless a deliberate identifying constraint is used.
• Units and center: record the original units, center, and scale so coefficients and predictions can be interpreted and reproduced.
• Model terms: transform interactions and any polynomial terms consistently; transforming only the main effect may change the model.
• Prediction checks: calculate predictions for several profiles using both parameterizations. Differences beyond rounding suggest a coding or coefficient-conversion error.
• Data support: verify that the center and scale are sensible and that categories or covariate ranges used for prediction are represented adequately. Scaling cannot remedy sparse categories or unsupported extrapolation.

Centering and scaling are often preferred to raw measurements when the raw zero is scientifically meaningless, units are very large or small, or numerical computation is poorly conditioned. They do not create information, correct a misspecified outcome model, or make an unsupported patient profile reliable. Standardizing a covariate also changes the meaning of its coefficient: it is now tied to the chosen scale, so report that scale.

Choosing a Coding Convention

Reference coding is especially useful when comparisons with a particular baseline category are central to interpretation. Alternative contrast systems can instead express comparisons with an overall mean or other specified combinations of categories. With the same model space, these are alternative coordinate systems: coefficient interpretations change, but fitted values do not. The choice should make the model easiest to interpret and communicate, while avoiding redundant columns.

For continuous variables, retaining original units makes a slope easy to explain per year, kilogram, or other natural unit. Centering can make the intercept meaningful, and scaling can make coefficient magnitudes easier to compare or improve numerical conditioning. In penalized models, models with prior distributions on coefficients, or algorithms that use coefficient-dependent constraints, rescaling can affect the fitted result unless the penalty, prior, or constraint is transformed too. Thus the simple prediction-invariance argument applies directly to equivalent unpenalized model parameterizations, not automatically to every estimation procedure.