# Trial-level surrogacy

The point estimates of i?2_{rial} that were obtained using the different models ranged between 0.6423 and 0.8151, and their 95% confidence intervals largely overlapped (see Table 4.2). It can thus be concluded that the accuracy by which *вг* (i.e., the treatment effect on *T* = visual acuity after 52 weeks in center *i)* can be predicted based on а_{г} (i.e., the treatment effect on *S* = visual acuity after 24 weeks in center i) is moderate.

Trial-level surrogacy can be visualized using a scatter plot that shows the estimated treatment effects on T = visual acuity after 52 weeks (i.e., в_{г}) against the estimated treatment effects on S = visual acuity after 24 weeks (i.e., *а _{г})* in the N = 36 different centers. Such a plot can be made for each of the different models shown in Table 4.1. By way of illustration, Figure 4.1 shows the results based on the reduced bivariate fixed-effects model, with and without weighting for cluster size in Stage 2 of the analysis (left and right figures, respectively). Note that the sizes of the circles in the left figure are proportional to the number of patients in the particular center. These plots

FIGURE 4.2

*Age-Related Macular Degeneration Trial. Scatter plot of the estimated residuals for T (the treatment effect on the true endpoint, i.e., visual acuity after *52 *weeks) against the estimated residuals for S (the treatment effect on the surrogate endpoint, i.e., visual acuity after 24 weeks) based on a (weighted or unweighted) reduced bivariate fixed-effects model.*

confirm the earlier conclusion about the prediction accuracy. The plots also suggest that the relationship between and a is approximately linear (when this is not the case, a quadratic or cubic model may be considered in Stage 2 of the analysis), and that the constant RE assumption (see Section 3.4.3.2) appears to be reasonable in the ARMD trial (i.e., the regression lines approximately pass through zero).