Centering and Scaling With Incomplete Predictors

Why a Common Reference Matters

Centering and scaling can make a regression model easier to interpret or fit. But when a predictor contains missing values, its completed values differ across imputed data sets. If the centering or scaling constants are recalculated from each completed data set, the resulting transformed variables use different reference points. That choice can change the meaning of coefficients being compared across imputations.

The key distinction is between values that vary because the missing predictor was imputed and constants that define the transformation. A fixed transformation uses the same constants in every imputation, even though the transformed values for missing cases vary with their imputed predictor values. Recomputing a passive term after an imputed draw is still essential; what remains fixed is the reference used in that calculation.

Definition: Let $X_i^{(d)}$ denote the value of predictor $X$ for case $i$ in completed data set $d$, where $d=1,\ldots,M$. A fixed centering constant $c$ and positive scaling constant $s$ define the transformed predictor $$ Z_i^{(d)}=\frac{X_i^{(d)}-c}{s}. $$ The same $c$ and $s$ are used for every case and every completed data set. The transformed values for cases with missing $X$ can differ across imputations because $X_i^{(d)}$ differs.

The constants may be specified using subject-matter units, chosen in advance, or calculated once from the observed values of $X$. If they are calculated from observed values, they are still fixed across the completed data sets for the analysis at hand. A scale must be positive and finite. A value such as $s=1$ performs centering without changing units; a clinically meaningful increment can also be used as the scale.

This convention does not claim that a data-derived constant is known without uncertainty. It means that the analyst chooses one reference transformation and uses it consistently across imputations. The immediate goal is a common parameterization, not a new method for accounting for uncertainty in the chosen constants.

How the Transformation Fits Into Chained Equations

For each imputation, the incomplete predictor is completed first, and its passive transformation is then calculated from that completed value using the chosen constants. For example, if $X$ has a missing value, the value $X_i^{(d)}$ is drawn in each imputation and $Z_i^{(d)}$ is set to $(X_i^{(d)}-c)/s$. It is not separately imputed as though it were an unrelated variable.

This maintains the deterministic identity between $X$ and $Z$ in every completed data set. The constants must also be used consistently anywhere this transformation enters the analysis. If a centered-and-scaled predictor is used in a model, its units and reference point should be recorded so that the coefficient has an unambiguous interpretation.

Choose the reference.
Specify $c$ and $s$ before fitting the analysis model. State whether they are externally chosen or calculated once from observed predictor values.

Complete the predictor.
In each imputation, retain observed $X$ values and use the imputed draw for each missing value.

Recalculate passive terms.
Apply the same transformation to the completed $X$ values in that imputation, including any product or other derived term that depends on $Z$.

Fit and interpret consistently.
Use the same variable definition in every analysis and interpret coefficients using the common $c$ and $s$.

Worked Example: Fixed Constants Versus Recomputed Means

Suppose a hypothetical study records an incomplete predictor $X$, measured in years. Two observed values are 4 and 8 years; the other two are missing. For illustration, use the fixed constants $c=6$ years and $s=2$ years. Two imputations produce completed values $(4,8,6,10)$ and $(4,8,4,14)$.

Using the fixed constants: In the first completed data set, the transformed values are $ (4-6)/2=-1$, $ (8-6)/2=1$, $ (6-6)/2=0$, and $ (10-6)/2=2$. In the second, they are $-1, 1, -1,$ and $4$, respectively. Both transformed data sets use the same reference: $X=6$ years maps to $Z=0$, and a two-year increase maps to an increase of one unit in $Z$.

Recomputing the mean and sample standard deviation separately: The first completed data set has mean $\bar{x}=(4+8+6+10)/4=7$ years. Its deviations are $-3,1,-1,3$, so its sample standard deviation is $s=\sqrt{(9+1+1+9)/(4-1)}=\sqrt{20/3}=2.582$ years. The second has mean $(4+8+4+14)/4=7.5$ years and sample standard deviation $\sqrt{(12.25+0.25+12.25+42.25)/3}=\sqrt{67/3}=4.726$ years.

With within-imputation standardization, the same case with $X=4$ becomes $(4-7)/2.582=-1.162$ in the first data set but $(4-7.5)/4.726=-0.7406$ in the second. These values are not errors: they reflect different means and scales. But they do not share the fixed interpretation “two years above or below 6 years.” The fixed transformation preserves that interpretation across imputations.

Why Coefficients Depend on the Reference

The difference matters because coefficients belong to a model parameterization. Consider a quadratic regression model, written before transformation as $Y=\beta_0+\beta_1X+\beta_2X^2+\varepsilon$. If $X=c+sZ$, substitution gives $Y=(\beta_0+\beta_1c+\beta_2c^2)+(\beta_1s+2\beta_2cs)Z+\beta_2s^2Z^2+\varepsilon$. Thus a change in centering changes the intercept and the linear coefficient, while a change in scale changes the coefficients’ units and magnitudes. The fitted curve can be the same even when its coefficient values differ.

\begin{aligned} \beta_0+\beta_1X+\beta_2X^2 &=\beta_0+\beta_1(c+sZ)+\beta_2(c+sZ)^2\\ &=(\beta_0+\beta_1c+\beta_2c^2) +(\beta_1s+2\beta_2cs)Z+\beta_2s^2Z^2. \end{aligned}

The same point appears in a linear model: if $Y=\beta_0+\beta_1X+\varepsilon$ and $X=c+sZ$, then the coefficient of $Z$ is $s\beta_1$, and the intercept is $\beta_0+c\beta_1$. If $s$ differs by imputation, the coefficient on $Z$ is expressed per a different number of units of $X$ in each data set. If $c$ differs, the intercept refers to a different value of $X$. Consequently, pooling those transformed-scale coefficients directly does not summarize one common parameter unless the target and parameterization have been made the same.

Worked Example: Re-expressing a Quadratic Model

Consider the hypothetical fitted relationship $Y=2+3X+0.5X^2$, where $X$ is measured in years. Use $c=4$ years and $s=2$ years, so $X=4+2Z$. Substituting gives

\begin{aligned} Y&=2+3(4+2Z)+0.5(4+2Z)^2\\ &=2+12+6Z+0.5(16+16Z+4Z^2)\\ &=22+14Z+2Z^2. \end{aligned}

At the reference point $X=4$, or $Z=0$, the predicted value is 22 units of $Y$. The coefficient 14 is the linear coefficient in the model expressed in $Z$; it is not the original coefficient of 3 per year of $X$.

Now keep the same original curve but choose $c=6$ and $s=2$, so $X=6+2Z$. The transformed model is $Y=38+18Z+2Z^2$. The coefficients differ, but substituting $Z=(X-6)/2$ gives the original curve. If the two choices of constants were made separately in two imputations, their transformed coefficients would not have identical reference-point interpretations. Using one choice in all imputations avoids this variation in definition.

With products and other derived terms, the same principle applies. Centering can change the coefficients attached to lower-order terms while preserving the represented relationship when the model includes the required terms. The identity should be checked for the model actually being fitted; a transformation does not guarantee that an incomplete or restricted model retains the same flexibility.

Worked Example: Updating a Centered Interaction

Suppose two incomplete predictors are $X$, in kilograms, and $W$, in hours. Set fixed constants $c_X=10$ kilograms, $s_X=4$ kilograms, $c_W=2$ hours, and $s_W=1$ hour. Define $Z_X=(X-10)/4$, $Z_W=W-2$, and the passive interaction $Q=Z_XZ_W$.

In one imputation, a missing $X$ is drawn as 12 kilograms and $W=3$ hours is observed. Then $Z_X=(12-10)/4=0.5$, $Z_W=3-2=1$, and $Q=0.5(1)=0.5$.

In another imputation, the draw for that same missing $X$ is 8 kilograms. The transformed values are $Z_X=(8-10)/4=-0.5$, $Z_W=1$, and $Q=(-0.5)(1)=-0.5$. The interaction changes because the imputed value changes, while both transformations retain the same reference and units. Keeping the first imputation’s interaction value in the second data set would be wrong: it would violate $Q=Z_XZ_W$ for the second completed values.

Assumptions, Diagnostics, and Alternatives

A fixed transformation is straightforward when the constants are well defined and the scale is strictly positive. If a scale is zero, the division is undefined; if it is extremely small relative to the predictor’s measurement precision, transformed values can be needlessly large. A centering value should also be meaningful for the intended interpretation. These are choices about model parameterization, not assumptions that make missingness ignorable or make an imputation model valid.

In practice, check the transformation in every completed data set. Confirm that observed predictor values remain unchanged, missing values use their imputed draws, and passive terms are recalculated from those values. Check that the same constants are recorded and applied throughout. If a transformed predictor is included in an analysis model, verify that coefficient interpretation uses the declared units and reference point.

Stale passive values: a derived term was calculated before an updated draw of its input and not refreshed. Recompute it after the input changes.
Accidental re-standardization: means or standard deviations were recalculated separately without intending to change the reference. Use the fixed constants, or explicitly plan to transform estimates to a common target before comparing them.
Unclear scale: a coefficient is reported per “one standardized unit” without saying how many original units that represents. State $c$, $s$, and the original measurement units.
Changing model meaning: coefficients from different parameterizations are compared as though they represented an identical quantity. Establish a common parameterization, or express results on a common scale before interpreting or pooling.

There are reasonable alternatives. Keeping $X$ on its original scale avoids choosing a centering or scaling constant, though the intercept may then refer to $X=0$ and large numerical magnitudes can be inconvenient. Standardizing separately within each completed data set can be useful when the goal is a within-data-set descriptive scale, provided its changing reference is understood. For a pooled regression interpretation, however, a common fixed transformation makes the coefficient definition transparent. Whichever approach is chosen, recalculating a passive transformation from its current inputs is distinct from recalculating the constants themselves.

Common mistake: “The predictor has different imputed values, so its mean and standard deviation must be recalculated in every imputation.” The first clause is correct; the conclusion does not follow. The passive transformed values should be recalculated because their inputs changed. But constants can remain fixed, giving each completed data set the same reference and scale. Within-imputation constants are a different modeling choice, not a requirement of passive imputation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

For $Z=(X-50)/10$, state what $Z=0$ means and what a one-unit increase in $Z$ represents.
A predictor’s imputed value is 30 in one completed data set and 40 in another. With fixed $c=35$ and $s=5$, calculate both transformed values.
Why can coefficients on standardized predictors be difficult to compare if each imputation uses its own standard deviation?
In a completed data set, $Z_X=0.4$ and $Z_W=-2$. What is the passive product $Q=Z_XZ_W$, and what must be done if either input is later updated?
Is recalculating a mean and standard deviation separately within every imputation mathematically invalid? Explain the more precise concern.

Solutions. (1) $Z=0$ corresponds to $X=50$, and an increase of one unit in $Z$ corresponds to an increase of 10 units in $X$. (2) The values are $(30-35)/5=-1$ and $(40-35)/5=1$. (3) Each coefficient is then expressed relative to a different original-unit increment, so the coefficient values need not estimate the same parameterization. (4) $Q=0.4(-2)=-0.8$; if either input changes, recompute $Q$ from the updated inputs. (5) No. It can be a deliberate within-imputation standardization, but it changes the reference or scale across data sets, complicating direct comparison and pooled interpretation unless results are put on a common target scale.

Centering and Scaling With Incomplete Predictors

What You'll Learn