Scaling Markers Before Estimating Association Effects

Why Scale Matters in a Multimarker Hazard

A joint model may link several longitudinal markers to an event hazard. One marker might be measured in concentration units and another in a much larger numerical range; their association coefficients then have different units and magnitudes. Comparing those raw coefficients directly is generally unhelpful. Rescaling can put marker effects on interpretable, comparable scales, but it changes coefficient values. The key is to transform the coefficients consistently: then the hazard contribution need not change at all.

This tutorial concerns that change of scale, not the choice of which markers belong in a model. Assume a fitted multimarker joint model supplies a latent trajectory for each marker, and that the hazard is linked to specified features of those trajectories. The trajectories and event data are not altered by a change of units. A coefficient describes the effect of a particular feature in its stated units, so changing those units requires a corresponding coefficient transformation.

Definition: For marker $k$, let $m_{ik}(t)$ be subject $i$'s latent marker trajectory at time $t$. For a chosen reference center $c_k$ and positive scale $d_k$, define the transformed trajectory $$ m_{ik}^{*}(t)=\frac{m_{ik}(t)-c_k}{d_k}. $$ The center and scale are common to subjects and time points. A common choice is a reference-population mean and standard deviation, but other clinically meaningful centers and scales are possible.

The transformation is affine: it subtracts a constant and divides by a positive constant. A positive scale preserves the ordering of marker values. If $d_k$ is the standard deviation in the chosen reference population, one unit of $m_{ik}^{*}(t)$ is one reference-population standard deviation of marker $k$. The center determines what transformed value zero means; the scale determines the unit of change.

Rescaling Current-Value Associations

For a current-value association with $K$ markers, write the hazard as

h_i(t)=h_0(t)\exp\left\{\boldsymbol{\gamma}^{\mathsf T}\mathbf{w}_i+\sum_{k=1}^{K}\alpha_k m_{ik}(t)\right\}.

Here $h_0(t)$ is the baseline hazard, $\mathbf{w}_i$ contains survival covariates with coefficient vector $\boldsymbol{\gamma}$, and $\alpha_k$ is the association coefficient for marker $k$. The longitudinal model supplies $m_{ik}(t)$; the survival part uses its value at time $t$. The association coefficient is conditional on the other terms included in this specified model. It is not, by itself, a causal effect.

Because $m_{ik}(t)=c_k+d_km_{ik}^{*}(t)$, the marker contribution can be rewritten as

\sum_{k=1}^{K}\alpha_k m_{ik}(t) = \sum_{k=1}^{K}\alpha_k c_k+ \sum_{k=1}^{K}(\alpha_k d_k)m_{ik}^{*}(t).

Thus define $\alpha_k^{*}=\alpha_k d_k$. The constant $\sum_k\alpha_k c_k$ is common to all subjects and times. In a model with an unspecified baseline hazard, it can be absorbed into the baseline hazard by defining $h_0^{*}(t)=h_0(t)\exp\{\sum_k\alpha_kc_k\}$. The resulting hazard is exactly the same for every subject and time:

h_i(t)=h_0^{*}(t)\exp\left\{\boldsymbol{\gamma}^{\mathsf T}\mathbf{w}_i+ \sum_{k=1}^{K}\alpha_k^{*}m_{ik}^{*}(t)\right\}.

This is an algebraic reparameterization, not a different fitted association. With a parameterized baseline hazard, the constant must instead be represented consistently in its intercept or other baseline parameters. If centering constants or scales vary across subjects, the added term is no longer a common baseline shift and this invariance argument does not apply.

Worked Example: Two Current-Value Markers on Different Scales

Suppose a hypothetical joint model has two latent marker values: marker A is measured in units with $c_A=10$ and $d_A=2$, and marker B in units with $c_B=100$ and $d_B=20$. Its fitted hazard is $$ h_i(t)=h_0(t)\exp\{0.30m_{iA}(t)+0.010m_{iB}(t)\}. $$

The transformed values are $m_{iA}^{*}=(m_{iA}-10)/2$ and $m_{iB}^{*}=(m_{iB}-100)/20$. The transformed coefficients are $$ \alpha_A^{*}=0.30(2)=0.60,\qquad \alpha_B^{*}=0.010(20)=0.20. $$ The centering constant in the exponent is $0.30(10)+0.010(100)=4$, so the equivalent model has $h_0^{*}(t)=h_0(t)e^4$ and marker contribution $0.60m_{iA}^{*}(t)+0.20m_{iB}^{*}(t)$.

For a numerical check, suppose a subject has $m_{iA}=12$ and $m_{iB}=80$. The original marker contribution is $0.30(12)+0.010(80)=4.40$. The transformed values are $m_{iA}^{*}=1$ and $m_{iB}^{*}=-1$; the transformed baseline shift plus marker contribution is $4+0.60(1)+0.20(-1)=4.40$. The hazard is unchanged. The transformed coefficients describe a one-reference-SD increase in each marker, holding the other modeled terms fixed; their different magnitudes now correspond to these chosen increments.

Slopes and Cumulative Features

The same principle applies when the hazard uses a feature other than the current marker value. For a linear trajectory feature, differentiating the transformation with respect to time gives $$ \frac{d}{dt}m_{ik}^{*}(t)=\frac{1}{d_k}\frac{d}{dt}m_{ik}(t). $$

If the original hazard includes a slope term $\xi_k\,dm_{ik}(t)/dt$, its equivalent transformed-scale coefficient is $\xi_k^{*}=\xi_kd_k$. Unlike a current-value term, subtracting a constant center has no effect on a slope. The slope's units change from marker-units per time to reference-SD units per time.

For a cumulative feature $A_{ik}(t)=\int_0^t m_{ik}(u)\,du$, transformation gives $$ A_{ik}(t)=c_kt+d_k A_{ik}^{*}(t),\qquad A_{ik}^{*}(t)=\int_0^t m_{ik}^{*}(u)\,du. $$

Consequently, an original term $\rho_k A_{ik}(t)$ becomes $\rho_kd_k A_{ik}^{*}(t)+\rho_kc_kt$. The scaled coefficient is $\rho_k^{*}=\rho_kd_k$, but the centering contributes a time-dependent term $\rho_kc_kt$, not merely a constant. It cannot in general be absorbed into a time-constant multiplier of the baseline hazard. To preserve the exact hazard, that term must be retained or represented by a compatible time-dependent baseline component. This distinction matters: for current values, centering adds a constant; for cumulative exposure from time zero, it adds a term proportional to time.

Worked Example: Rescaling a Slope Association

Suppose a hypothetical hazard contains the marker A slope in units per year, with coefficient $\xi_A=0.40$ per marker-unit per year. Marker A is transformed as $m_A^{*}(t)=(m_A(t)-10)/2$. A subject's original slope is $0.50$ marker-units per year.

The transformed slope is $0.50/2=0.25$ reference SDs per year. To preserve the exponent, use $\xi_A^{*}=\xi_A d_A=0.40(2)=0.80$ per reference SD per year. Substitution gives $0.40(0.50)=0.20$ in the original model and $0.80(0.25)=0.20$ in the transformed model. The slope contribution, and therefore the hazard contribution from this term, is identical.

The center $10$ does not enter this calculation: differentiating a constant gives zero. This conclusion assumes the same differentiable trajectory and time unit in both parameterizations. If time itself is rescaled, the slope coefficient must also be adjusted for that change of time unit.

A General Justification and Practical Workflow

The invariance is an immediate consequence of substituting an invertible transformation into the linear predictor. More generally, let $q_{ik}(t)$ be the feature of marker $k$ entering the hazard, and suppose its transformation can be written $q_{ik}(t)=a_k+b_kq_{ik}^{*}(t)$, where $a_k$ is common across subjects. Then

\sum_k\theta_kq_{ik}(t) = \sum_k\theta_ka_k+\sum_k(\theta_kb_k)q_{ik}^{*}(t).

The feature coefficient becomes $\theta_k^{*}=\theta_kb_k$; the remaining term is handled by the baseline or explicitly retained, as appropriate to its dependence on time. This derivation also shows why a coefficient cannot be rescaled in isolation while leaving the feature unchanged: the product in the hazard exponent must be preserved.

A careful analysis follows a consistent sequence:

Specify the feature.
State whether each marker enters through its current value, slope, cumulative exposure, or another defined function of its trajectory.

Choose a common reference.
For each marker, record the center $c_k$, positive scale $d_k$, source population, and units. Apply these same constants across subjects and time.

Transform the feature and coefficient together.
Substitute the transformed feature into the original linear predictor, identify any added term, and verify that the exponent is preserved.

Report the resulting scale.
Describe each association per one stated increment, such as one reference-population SD, and clarify whether coefficients are from a refitted model or are algebraically converted.

Worked Example: Interpreting a One-SD Hazard Ratio

Consider a hypothetical current-value association for a marker with original coefficient $\alpha=0.25$ per unit and reference standard deviation $d=4$ units. Its standardized coefficient is $\alpha^{*}=0.25(4)=1.00$ per SD. The hazard ratio for a one-SD higher current marker value, with the other modeled terms fixed, is $$ \exp(\alpha^{*})=\exp(1.00)=2.718. $$

Under this model, two otherwise identical modeled marker profiles differing by one reference SD in the current marker have an instantaneous hazard ratio of about 2.72 at that time. This is a model-based association, not a statement that increasing the marker would cause an individual's hazard to rise by that amount. A 0.5-SD contrast has hazard ratio $\exp(1.00(0.5))=\exp(0.50)=1.649$. The scale and contrast must be stated; “the coefficient is 1” alone has no useful interpretation.

Choosing a Scale, Diagnostics, and Alternatives

Standardization is useful when markers have incomparable units or when a numerical algorithm is affected by very different predictor scales. It also gives a coefficient per reference SD, which can support a limited comparison of association magnitudes. Such comparisons remain dependent on the selected reference population and its variability: a one-SD change in one marker need not be clinically comparable to a one-SD change in another.

A standard deviation calculated from repeated observations can give disproportionate influence to subjects with more visits. If the intended reference is the between-subject distribution, define it from an appropriate subject-level or model-based reference rather than silently treating all repeated measurements as independent draws. One may instead choose a clinically meaningful increment, such as a specified concentration change. The algebra is identical: use that increment as $d_k$, and explain it clearly.

A practical diagnostic is to recompute the linear predictor for selected subjects and times under both parameterizations. For a current-value model, verify equality after including the baseline shift; for slope or cumulative terms, verify all added terms as well. A mismatch usually indicates an incorrect multiplier, a missed center term, or inconsistent units. Also check that every $d_k>0$ and is nonzero. A constant marker has no estimable standard deviation, so it cannot be standardized this way or provide a varying association feature.

Standardizing is not the same as refitting a different model. If the transformation is invertible and applied consistently, it is a reparameterization: fitted hazards and likelihood are unchanged in principle, although numerical results can differ slightly because of optimization tolerances. By contrast, changing the reference constants, changing the marker feature, or changing the marker data can alter the model being fitted. Reporting raw-unit coefficients remains appropriate when those units are clinically meaningful; per-SD coefficients are an alternative scale, not a universally superior one.

Common Pitfalls and Limitations

Comparing raw coefficient magnitudes. A numerically small coefficient can reflect a large unit of measurement. Compare specified contrasts or transformed coefficients, not the bare numbers.
Scaling the marker but not its coefficient. This changes the linear predictor and therefore changes the hazard. For a divisor $d_k$, the coefficient multiplying the standardized feature is the original coefficient times $d_k$.
Forgetting what centering does. In a current-value model it adds a constant to the exponent; in a cumulative model it adds a time-varying term. The latter generally needs explicit treatment.
Standardizing separately for each subject or visit. This removes a common interpretation and can destroy the shared-scale transformation used in the invariance argument.
Treating per-SD effects as directly comparable clinical effects. The SD is population-dependent. Different reference distributions can yield different per-SD coefficients even when raw-unit coefficients agree.
Reading an association as causal. Rescaling changes units, not the assumptions required for causal interpretation.

Standardization also does not resolve poor measurement, highly correlated markers, weakly determined coefficients, or an unsuitable longitudinal or survival model. Those issues require separate assessment; a convenient scale cannot supply information absent from the data or repair a misspecified model.

Check Your Understanding

Answer each question, then compare with the brief solutions.

A current-value marker has coefficient $0.08$ per unit and is divided by $5$ to form its standardized value. What coefficient multiplies the standardized marker?
Why does centering a current-value marker produce a constant shift in the hazard exponent?
A slope coefficient is $0.6$ per unit per year, and the marker is divided by $3$. What is the transformed coefficient per SD per year?
For cumulative exposure from time zero, what extra term results from centering at $c$?
Does a per-SD coefficient establish that changing the marker causes the corresponding hazard change? Explain briefly.

Solutions. (1) $0.08(5)=0.40$ per standardized unit. (2) Substitution $m=c+dm^{*}$ adds $\alpha c$, which is common across subjects and times for a current-value term. (3) $0.6(3)=1.8$ per SD per year; the marker slope itself is divided by $3$. (4) The cumulative feature includes $ct$, so the exponent gains $\rho ct$ for cumulative coefficient $\rho$. (5) No. It is a model-based association on a stated scale; standardization does not establish causality.