Reading Forest Plots of Covariate Effects

What a Forest Plot Shows

A forest plot places covariate-effect estimates and their confidence intervals on a shared horizontal scale. Each row usually represents a covariate contrast, a continuous-covariate change, or a model parameter. The point marks the estimated effect; the horizontal line shows an interval expressing uncertainty under a stated method. Vertical reference lines help readers judge the estimate against a no-effect value and, sometimes, one or more clinically relevant thresholds.

The graphic is compact, but it is not self-interpreting. Before reading whether a marker lies to the left or right, establish what parameter is being described, what comparison is made, and what scale is plotted. A value to the right of the reference may mean higher clearance, higher volume, or a higher parameter ratio; the clinical meaning depends on the row label and scale. A forest plot summarizes estimates from a fitted model. It does not by itself establish that a covariate causes a pharmacokinetic difference.

Start With the Comparison and Reference

Read the axis title, row labels, legend, and caption together. Identify whether the plot concerns clearance, volume, or another parameter; whether each row compares a group with a reference group or represents a specified change in a continuous covariate; and whether the estimate is a ratio, percentage change, or additive difference. The same point estimate can mean different things under different labels. For example, a ratio of 1.20 means 20% higher for the numerator group relative to the denominator group, while a ratio of 0.80 means 20% lower.

On a difference scale, no effect is at 0: an estimated difference of \(+0.4\) L/h is an increase of 0.4 L/h in the stated comparison. On a ratio scale, no effect is at 1: an estimated ratio of 1.20 is not a 1.20-unit increase. For a ratio \(R\), the percentage change relative to the reference is \(100(R-1)\%\). This formula also applies to the confidence limits, because the ratio scale is being expressed as percentage change relative to 1.

Direction is always relative to the specified order of comparison. If a row says “group B versus group A,” a ratio above 1 indicates a higher parameter value in group B. Reversing the comparison gives the reciprocal ratio, so the direction reverses too. For a continuous covariate, the plot must specify the increment, such as a 10-unit increase; the estimate is not an effect “per unit” unless that is what the axis or caption says.

What the Confidence Interval Tells You

A confidence interval communicates uncertainty in the estimated effect. Under the model and interval procedure used, a 95% confidence procedure has 95% coverage over repeated samples: if the procedure were repeated under its assumptions, 95% of the resulting intervals would contain the fixed underlying effect. It is not a statement that there is a 95% probability the particular effect lies inside the observed interval.

The interval’s position describes which effect values are compatible with the data and procedure at the stated confidence level; its width describes precision on the displayed scale. A narrow interval indicates greater precision than a wide interval on the same scale, but width comparisons across different scales or different covariate contrasts may not be meaningful. A point estimate near the reference can have a narrow interval or a wide interval. The estimate and its uncertainty must be read together.

For a two-sided 95% interval and a matching test of the no-effect value at \(\alpha=0.05\), an interval excluding the reference corresponds to statistical evidence against that no-effect value under the same inferential assumptions. This relationship does not say whether the effect matters clinically. Nor does an interval that includes the reference prove that the effect is exactly zero; it indicates that the interval includes the no-effect value.

Reading Clinical Thresholds

A clinical-threshold line answers a different question from the no-effect line. The no-effect reference marks zero change; a clinical threshold marks a magnitude selected for a particular clinical decision or interpretation. A threshold must be defined on the scale used for the comparison. For example, a threshold of a 20% change on a ratio-derived percentage scale corresponds to ratios of 0.80 and 1.20 if changes in either direction are considered relevant.

Compare the point estimate and the entire interval with the threshold. An estimate beyond the threshold suggests that the fitted point estimate reaches that prespecified magnitude. If the interval crosses the threshold, uncertainty includes effects on both sides of it. If the interval lies wholly beyond a threshold in one direction, the estimated effect is more consistently beyond that boundary, subject to the model and interval assumptions. These visual comparisons do not themselves determine a treatment decision; their meaning depends on whether the threshold is appropriate for the parameter, contrast, and clinical context.

A Short Justification for Ratio-Scale Intervals

Many covariate effects are estimated on a logarithmic scale and then displayed as ratios. If \(\beta\) is a log-ratio effect, the ratio is \(R=\exp(\beta)\); no effect is \(\beta=0\), which becomes \(R=1\). Suppose an approximate confidence interval for \(\beta\) is \([L,U]\). Since the exponential function is increasing, transforming both endpoints preserves their order and gives a ratio interval \([\exp(L),\exp(U)]\). For example, if the log-scale interval is \([-0.223,0.182]\), the ratio limits are approximately \([\exp(-0.223),\exp(0.182)]=[0.800,1.200]\). This is why a symmetric interval on the log scale generally appears asymmetric around the point estimate on the ratio scale. It is not an error: the scale has changed.

This transformation is justified when the interval is constructed for the log effect and then exponentiated. Other procedures, such as bootstrap intervals, may produce intervals by a different route. The plot caption should identify the interval method; do not assume that every displayed interval is a Wald interval or that every interval has the same construction.

Diagnostics, Comparisons, and Common Pitfalls

Before drawing a conclusion, check that the plot is readable and that the estimates are comparable. Confirm the axis scale and units, the no-effect line, the row-specific contrast, the confidence level, and whether any clinical thresholds are shown. On a logarithmic axis, equal distances correspond to equal multiplicative changes, not equal additive differences. If the axis is truncated, a small visual displacement can appear dramatic; read the tick marks rather than judging distance by eye. If intervals are clipped at plot boundaries or rows are labeled ambiguously, consult the accompanying table or model report.

A forest plot is often more efficient than a table for scanning direction and uncertainty across many covariates. A table, however, can report exact values, units, coding, and interval methods more precisely. A coefficient plot may display estimates on the model’s coefficient scale, while a forest plot may display transformed effects such as ratios. Neither display is inherently more valid; the choice should make the estimand and its uncertainty clear. Do not rank effects by the apparent lengths of bars when rows use different contrasts, parameters, or scales.

• Wrong reasoning: “The interval crosses 1, so the covariate has no effect.” Correction: The interval includes the no-effect value, but it may also include nonzero effects of meaningful size; its width matters.
• Wrong reasoning: “The point is to the right, so the covariate increases clearance.” Correction: First check whether the row concerns clearance and which group or covariate value is the numerator or reference.
• Wrong reasoning: “The confidence interval crosses the clinical threshold, so the effect is clinically irrelevant.” Correction: Crossing the boundary indicates uncertainty about whether the threshold is reached, not proof that it is not reached.
• Wrong reasoning: “Two rows have overlapping intervals, so their effects are not different.” Correction: Overlap alone is not a formal test of the difference between estimates; that comparison requires its own uncertainty calculation and covariance information.
• Wrong reasoning: “The plot shows a covariate causes a change in the PK parameter.” Correction: The plot reports model-based associations or contrasts; causal interpretation requires additional design and assumptions.

A forest plot is most useful as a disciplined summary, not as a substitute for the model specification. It makes direction, magnitude, and uncertainty visible, but interpretation still depends on a well-defined comparison, an appropriate scale, and clearly stated thresholds. In population pharmacokinetics, the model context also matters: a displayed covariate effect is conditional on the model and data used to estimate it.