Dr. Eisner contends that despite the significant tests for trend in Table 3 of our paper, the data "clearly indicate a nonlinear risk pattern" on the basis of the similarity of the degree of risk elevation in the middle three quintiles of the homocysteine distribution. He expresses concern that imprecise modeling of the association has consequences for assessing causality and for public health intervention.

The mortality hazard ratios (adjusting for all confounders) in successive quintiles were 1.00, 1.42, 1.28, 1.53, and 1.97 for the total study sample; we referred to the association as nonmonotonic. But reliance on these figures alone, without taking their confidence intervals into account, may be misleading. Although the findings might suggest a nonlinear association, they are clearly not inconsistent with a linear association when the confidence intervals displayed in Table 3 are considered (0.92 to 2.18, 0.84 to 1.96, 1.01 to 2.33, and 1.31 to 2.98 for quintiles 2 to 5, respectively). If, instead of presenting the hazard ratios for quintiles, we had used homocysteine categories as modified from Nygard and colleagues (1)—less than 9.0, 9.0 to 11.9, 12 to 14.9, 15.0 to 19.9, and 20 µg/dL or more—the association would have been monotonic. The successive multivariate-adjusted mortality hazard ratios would then become 1.0, 1.16, 1.28, 1.54, and 1.94 (Wald statistic, 11.06; P = 0.026).

We should not lose sight of the wood for the trees. Both analyses indicate that the mortality risk tends to be lowest when the homocysteine level is low, highest when the level is high, and intermediate when the homocysteine level is in between. We pointed out that our study had insufficient power to firmly characterize the dose-response relation.

The public health implications of the findings (other than elevated homocysteine level serving as a risk marker for mortality) depend on establishing causality. Homocysteine-lowering trials, now under way (2), are a necessary step in this process. If efficacy is established, the homocysteine level at which intervention should begin becomes of substantive interest. Our estimates of the population attributable risk percentage, a measure of the potential public health benefits of intervention at levels of 13 or 14 µg/dL and higher (values that included nearly one third and one quarter of the study sample, respectively), pointed to a substantial preventable fraction. In our opinion, resolution of the central issue of causality will not now be advanced by alternate statistical modeling of associations but rather by performing intervention trials and clarifying biological mechanisms.