Worked Example: From Forest Plot to Clinical Conclusion

From a Forest Plot to a Clinical Claim

A forest plot can make several covariate effects look directly comparable: each row has an estimate and an interval, often on a common ratio scale. But a plotted number is not yet a clinical conclusion. The meaning depends on the model, the contrast represented by that row, the reference values or categories, and the range of effects considered clinically important. A continuous covariate’s effect also depends on the size of the change being considered; a per-unit estimate may be numerically small while a realistic contrast is not.

This tutorial uses a hypothetical population pharmacokinetic analysis in which covariates affect typical clearance. The forest plot reports ratios of typical clearance for stated contrasts. All numerical values below are invented for teaching. The clinical relevance range used in the examples is also hypothetical: a clearance ratio from 0.80 to 1.25 is treated as a range of effects not considered large enough, by itself, to be clinically important for the decision at hand. Such a range must be justified for the particular drug and clinical context; it is not a universal rule.

Read the Contrast Before Reading the Number

Suppose typical clearance is described by a log-linear covariate model, \(CL_{\mathrm{typ}}(x)=CL_{\mathrm{ref}}\exp\{\beta(x-x_{\mathrm{ref}})\}\), for a continuous covariate \(x\). Here \(CL_{\mathrm{ref}}\) is typical clearance at the reference value \(x_{\mathrm{ref}}\), and \(\beta\) is the change in log typical clearance per unit of \(x\). Comparing values \(x_1\) and \(x_0\) gives a ratio, not a difference:

Thus the coefficient alone does not specify a clinically useful effect until a contrast \(x_1-x_0\) is chosen. For a categorical covariate with a reference category, a coefficient \(\gamma\) on the log-clearance scale gives the ratio \(\exp(\gamma)\) for the named category compared with that reference. The forest-plot row should make the comparison direction explicit. “Female versus male,” for example, is not interchangeable with “male versus female”; reversing the contrast takes the reciprocal of the ratio.

A forest plot may place these estimates and intervals on a logarithmic horizontal axis. On a ratio scale, the null line is at 1. A confidence interval entirely on one side of 1 is evidence, under the model and interval procedure, that the ratio differs from 1 at the corresponding confidence level. That fact alone does not say whether the effect is clinically important. Clinical relevance requires comparing the effect’s magnitude and uncertainty with an appropriate clinical benchmark.

Calculate a Clinically Relevant Continuous Contrast

For the weight row, suppose the fitted log-linear model estimates \(\hat{\beta}_W=0.00600\) per kg, with a 95% confidence interval from 0.00120 to 0.0108 per kg. The chosen comparison is 90 kg versus 70 kg, so the contrast is 20 kg. Exponentiating the coefficient multiplied by the contrast gives the estimated ratio:

The corresponding percentage difference is calculated from the ratio, not by treating the ratio itself as a percentage:

Under this model, the estimated typical clearance at 90 kg is about 12.75% higher than at 70 kg. The result is an adjusted model contrast, conditional on the other covariates and model structure, not a causal claim that increasing a person’s weight would cause clearance to rise.

If the coefficient interval is \([0.00120,0.0108]\), the contrast interval follows by multiplying both endpoints by 20 and exponentiating. Because the exponential function is increasing, the order of the endpoints is preserved:

The ratio interval corresponds to an estimated range from about 2.43% higher to 24.11% higher typical clearance for 90 kg versus 70 kg. This transformation is valid when the coefficient interval is for the same model coefficient and the contrast is treated as fixed. It does not add uncertainty about the model form, covariate measurement, or whether the selected contrast represents the intended clinical population.

Compare Magnitude and Uncertainty With Clinical Relevance

For illustration, suppose ratios from 0.80 to 1.25 have been specified in advance as the range not considered clinically important for this hypothetical decision. This benchmark is centered approximately symmetrically on a log scale: 0.80 is the reciprocal of 1.25. A useful assessment considers both the point estimate and the confidence interval against this range. The benchmark should be scientifically justified and chosen for the decision, rather than selected after seeing which boundary makes an estimate appear important.

For weight, the point estimate 1.1275 is within the 0.80-to-1.25 range, and the full 95% interval, 1.0243 to 1.2411, is also inside it. The estimate suggests a difference from the null ratio, but the entire interval remains within the illustrative range of effects not considered clinically important. Under this benchmark, the plot does not suggest a large weight-associated clearance contrast for 90 versus 70 kg, even though the interval excludes 1.

For eGFR, the estimate is 0.820, or an estimated 18.0% lower typical clearance at 60 than at 90 mL/min/1.73 m². The interval from 0.680 to 0.990 excludes 1, but it extends below 0.80. The data are compatible both with a relatively small decrease near 1 and with a decrease exceeding the illustrative clinical boundary. The point estimate alone is just inside that boundary; uncertainty prevents a confident claim that the contrast is clinically important under the stated criterion.

For female versus male, the point estimate is 0.910, an estimated 9.0% lower typical clearance in females. The interval, 0.780 to 1.060, includes 1 and extends below 0.80. It therefore includes no difference, modest differences in either direction, and a potentially larger decrease. It would be inaccurate to conclude that the categories have identical clearance, just as it would be inaccurate to claim a precisely established clinically important decrease. The interval indicates that the effect is uncertain.

Diagnostics, Alternatives, and Limits

Before interpreting a plot, check that each row is on the same effect scale, that the direction of every contrast is explicit, and that each interval is paired with the right estimate. For continuous covariates, verify that the displayed contrast is clinically interpretable and that the model’s assumed functional form is appropriate over that range. For categorical covariates, check the stated reference category and whether the model includes relevant interactions. A single ratio may not summarize the effect if the covariate relationship changes across another covariate.

Look for signs that the plot is being asked to support more than the analysis can: a wide interval, an estimate near a clinical boundary, sparse representation of one category, or a contrast extending beyond the covariate values well represented in the data. Such issues can make an apparently precise plotted summary misleading. The plot also does not show all model uncertainty, such as uncertainty in the chosen structural model, residual error specification, or covariate functional form.

There are several reasonable ways to express an effect. A continuous coefficient per unit is useful for describing the fitted relationship, but a selected contrast is usually easier to interpret clinically. A ratio is natural for multiplicative effects; a difference in clearance units may be preferable when absolute clearance changes matter to the decision. Either representation should be derived from the same fitted model and should state its contrast and units. On a ratio scale, a log-scale plot can display reciprocal changes symmetrically; the clinical interpretation, however, still needs to be stated in words.

Finally, a confidence interval is not a probability statement about the fixed parameter for this particular interval. In repeated applications of the same procedure under its assumptions, 95% of the intervals constructed this way would contain the parameter. The interval’s width describes estimation uncertainty under the fitted model and interval method; it does not guarantee that the model is correct or that every clinically relevant source of uncertainty has been represented.

Common Mistakes

• Calling a confidence interval that excludes 1 clinically important. Exclusion of 1 concerns evidence for a non-null ratio; clinical importance depends on a clinical benchmark and the magnitude of the effect.
• Reporting a continuous coefficient without its contrast. A coefficient per kg does not state the effect from 70 to 90 kg. Multiply by the contrast on the log scale, then exponentiate.
• Reversing a categorical comparison without changing the ratio. Female versus male and male versus female have reciprocal ratios and different verbal interpretations.
• Treating an interval that includes 1 as proof of no difference. It indicates that no difference is among the values compatible with the interval; the interval may also include effects that matter clinically.
• Rounding intermediate values too early. Carry sufficient precision through exponentiation and percentage calculations, then round the final result. For example, \(\exp(0.120)=1.12749685\ldots\), which rounds to 1.1275 at four decimal places or 1.127 to three decimal places.