Why Stratified Summaries Can Help Set Covariate Seeds
A population model may relate a typical clearance to a covariate such as body weight, renal function, or age. Before estimation, the model needs initial values for its parameters. Stratified summaries can give a useful scale and direction for these starting values: if higher-weight groups have higher typical clearance, for example, that pattern can suggest a positive weight effect. The summaries provide a seed, not a substitute for fitting the model to individual-level data.
The conversion depends on the chosen functional form. A linear model estimates a change in clearance per unit of covariate. A power model estimates a dimensionless exponent that describes proportional change. In either case, specify the covariate reference value and the units before calculating. Without that information, an effect estimate can be numerically plausible but have the wrong interpretation.
Specify the Covariate Model and Reference
Let \(x\) denote a covariate, let \(x_{\mathrm{ref}}\) be its reference value, and let \(CL_{\mathrm{ref}}\) be typical clearance at that reference. A centered linear model is
Here \(\beta\) has units of clearance divided by covariate units, such as L/h per kg. The centering makes \(CL_{\mathrm{ref}}\) the clearance at the stated reference, rather than at \(x=0\). For two strata with summaries \(CL_1\) and \(CL_2\) at covariate values \(x_1\) and \(x_2\), a direct linear seed is
A common alternative is a power model:
The ratio \(x/x_{\mathrm{ref}}\) must be dimensionless, so \(x\) and \(x_{\mathrm{ref}}\) use the same units; \(\theta\) is dimensionless. This form requires positive covariate values. For two strata, if the summaries are positive, taking logarithms gives
This two-stratum calculation assumes the summaries are suitable representations of typical clearance in the groups and that the chosen functional form is a reasonable local description across the covariate range. It also requires \(x_2\ne x_1\). When \(x_2\) is close to \(x_1\), the denominator is small, so small summary differences or rounding can produce a highly unstable exponent.
Worked Example: Seeding a Linear Weight Effect
Suppose a hypothetical summary reports typical clearances of 4.80 L/h for a group with mean weight 60 kg and 5.60 L/h for a group with mean weight 80 kg. Use a centered linear model with \(x_{\mathrm{ref}}=60\) kg and \(CL_{\mathrm{ref}}=4.80\) L/h.
The slope seed is
Thus an initial linear model is \(CL(x)=4.80+0.0400(x-60)\), with \(x\) measured in kg and clearance in L/h. At 70 kg it predicts \(4.80+0.0400(10)=5.20\) L/h. At 90 kg it predicts \(4.80+0.0400(30)=6.00\) L/h. Those values are model-based starting predictions, not claims that the group summaries prove a causal weight effect.
The arithmetic is meaningful as a seed if the two summaries describe comparable groups and the linear form is plausible over the range. If the summaries have different dose, study, or measurement conditions, the difference may reflect those factors as well as weight.
Power Effects From Clearance or Exposure Ratios
For multiple strata, the power model can be written on the log scale relative to a reference stratum. Define \(z_j=\log(x_j/x_{\mathrm{ref}})\) and \(y_j=\log(CL_j/CL_{\mathrm{ref}})\). The power model implies \(y_j=\theta z_j\). If the reference summary is fixed and the non-reference strata are given weights \(w_j\), a weighted least-squares seed through the origin is
The weights should reflect a defensible precision or design choice; they are not automatically equal to group sample sizes. If all strata have similar precision, equal weights can be a transparent exploratory choice. This regression uses transformed group summaries, not individual observations. It should be understood as a way to initialize the model, not as an inferential procedure that accounts for individual variability or uncertainty in each summary.
Worked Example: A Power Effect From Three Weight Summaries
Consider hypothetical geometric-mean clearances of 5.00 L/h at 50 kg, 6.50 L/h at 70 kg, and 8.00 L/h at 90 kg. Set \(x_{\mathrm{ref}}=50\) kg and \(CL_{\mathrm{ref}}=5.00\) L/h. Use the two non-reference strata with equal weights. The log covariate ratios are approximately \(z_1=\log(70/50)=0.3365\) and \(z_2=\log(90/50)=0.5878\). The corresponding log clearance ratios are \(y_1=\log(6.50/5.00)=0.2624\) and \(y_2=\log(8.00/5.00)=0.4700\).
With equal weights, the weight cancels from the weighted-slope formula. Using the displayed rounded values for the intermediate products,
The reported slope is \(\theta_{\mathrm{seed}}\approx0.7947\), calculated from the unrounded logarithms; the displayed intermediate values are rounded. The resulting seed predicts \(CL(70)=5.00(70/50)^{0.7947}\approx6.53\) L/h and \(CL(90)=5.00(90/50)^{0.7947}\approx7.977\) L/h (about 7.98 L/h). The latter is close to the 8.00 L/h summary used to construct the seed. A close match is not guaranteed: with several strata, the fitted log-scale line generally compromises among them.
The reference matters: the model’s \(CL_{\mathrm{ref}}\) is clearance at 50 kg, and the ratios are relative to 50 kg. Changing the reference changes the value assigned to the reference clearance and the covariate ratios used in the calculation; it does not justify mixing a reference clearance from one group with ratios defined against another.
Dose-normalized exposure summaries can be used similarly, but the relationship to clearance must be stated. Under linear pharmacokinetics, for intravenous dosing, \(AUC/D=1/CL\). For oral dosing, \(AUC/D=F/CL\), where \(F\) is bioavailability; exposure differences then cannot be attributed to clearance alone unless bioavailability is known or reasonably constant across the groups.
Worked Example: Converting Dose-Normalized Exposure to a Power Seed
Suppose a hypothetical intravenous study reports \(AUC/D=0.250\) h/L at renal-function value 60 units and \(AUC/D=0.200\) h/L at value 90 units. Assume linear pharmacokinetics and use \(x_{\mathrm{ref}}=60\). Because \(CL=1/(AUC/D)\) for this setting, the implied summaries are
The power-effect seed is
The implied exposure ratio is \(0.200/0.250=0.8000\). Equivalently, the power exponent for \(AUC/D\) against renal function is \(\log(0.8000)/\log(1.50)\approx-0.5503\), the negative of the clearance exponent under the stated assumptions. Here the covariate units cancel in the ratio; report the renal-function unit when describing the reference value. If this were oral exposure and \(F\) varied by group, the same reciprocal conversion would not identify clearance differences without additional information.
Derivation, Diagnostics, and Failure Modes
The multi-stratum power formula follows directly from the log model. Since \(CL(x_j)=CL_{\mathrm{ref}}(x_j/x_{\mathrm{ref}})^\theta\), division by \(CL_{\mathrm{ref}}\) and taking logarithms gives \(y_j=\theta z_j\). Weighted least squares chooses \(\theta\) to minimize \(Q(\theta)=\sum_j w_j(y_j-\theta z_j)^2\). Differentiating and setting the derivative to zero yields
provided \(\sum_jw_jz_j^2>0\). This derivation explains why the fitted line is constrained through the reference point \((0,0)\). It also shows why strata close to the reference contribute little information about the slope: their \(z_j^2\) terms are small.
Before using a seed, check the summaries and the model’s implied predictions. Plotting group summaries against the covariate can reveal a trend that bends, reverses, or is driven by one extreme stratum. For a power model, inspect the log-ratio relationship; for a linear model, inspect the clearance change per covariate unit. Also verify that the covariate values and reference use the same units, that clearance units are consistent, and that the summaries are comparable in dose, route, population, and measurement method.
- Wrong direction after inversion. Higher \(AUC/D\) at a covariate value implies lower clearance under the intravenous linear-PK relationship. Invert the exposure ratio before assigning a clearance-effect direction.
- Uncentered linear interpretation. In \(CL(x)=CL_{\mathrm{ref}}+\beta(x-x_{\mathrm{ref}})\), \(CL_{\mathrm{ref}}\) is the value at \(x_{\mathrm{ref}}\), not necessarily at zero.
- Incompatible units. A slope calculated per kg cannot be used as a slope per g without conversion. The ratio in a power model must also compare covariate values expressed in the same units.
- Over-reading group summaries. A between-group contrast can reflect factors other than the covariate. It does not establish a causal effect or capture within-group variability.
- Unstable exponent. If covariate values are nearly equal, \(\log(x_2/x_1)\) is near zero. Small errors in the group summaries can then greatly alter the calculated exponent.
A two-stratum contrast is the simplest alternative to a multi-stratum log-scale fit. It is easy to audit but can be dominated by the selected pair. A multi-stratum fit uses more of the available pattern but remains a fit to summaries, and its result depends on the weighting and reference choices. In either case, individual-level model estimation is the appropriate stage for evaluating the covariate effect using the model and data; the summary-derived value is only an initial value. These seeds are also limited by the assumed functional form: a linear slope may extrapolate to implausible clearance values, while a power law may be unsuitable for zero or negative covariates and can misrepresent a nonlinear trend.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Two hypothetical groups have mean weights 40 kg and 60 kg and typical clearances 3.6 L/h and 4.4 L/h. What is the linear slope seed in L/h/kg?
- A power model uses \(x_{\mathrm{ref}}=50\) kg. What does \(CL_{\mathrm{ref}}\) represent, and what are the units of the power exponent?
- For intravenous linear pharmacokinetics, dose-normalized exposure rises across covariate groups. What direction of clearance change does that imply?
- Why can a power exponent from two strata be unstable when their covariate values are close?
- Why is a group-level covariate association not, by itself, evidence that the covariate causes a clearance change?
Solutions. (1) \(\beta=(4.4-3.6)/(60-40)=0.8/20=0.0400\) L/h/kg. (2) \(CL_{\mathrm{ref}}\) is typical clearance at exactly 50 kg; the exponent is dimensionless because it applies to a dimensionless covariate ratio. (3) Since \(AUC/D=1/CL\), higher dose-normalized exposure implies lower clearance, assuming the stated intravenous and linear-PK conditions hold. (4) The exponent divides a log clearance ratio by \(\log(x_2/x_1)\); when the covariate values are close, this denominator is near zero and small summary changes can cause a large change in the ratio. (5) Groups can differ in other characteristics or study conditions, so the observed contrast may be confounded and does not isolate a causal effect.