Choi and colleagues [1] use data from the AIDS Clinical Trials Group (ACTG) Protocol 019 Trial to investigate the degree to which zidovudine's effect on the risk for clinical disease acts via its effect on the CD4⁺ lymphocyte count (CD4⁺ percent). They fit the CD4⁺ lymphocyte count (CD4⁺ percent) as a time-dependent covariate (that is, the current value) in a Cox proportional-hazards model of the effect of zidovudine on the risk for progression to AIDS. The small degree of attenuation in the placebo-zidovudine relative risk estimate resulting from such adjustment led to the conclusion that only a small proportion of the effect of zidovudine on progression to AIDS was statistically explained by its effect on CD4⁺ lymphocyte levels (CD4⁺ percent). The range of possible values for the proportion is given as 0% to 37% (although we calculated this value to be 45% from the results presented). The former estimate relates to use of the CD4⁺ lymphocyte count and the latter to the use of the CD4⁺ percent.

This range of values, however, is not a confidence interval and therefore does not reflect the potentially large uncertainty inherent in their results. For example, the estimated proportions are based on the assumption that the unadjusted placebo-zidovudine relative risk estimate (2.08) is free from error. The lower 95% CI of this relative risk estimate is 1.14. Thus, the true relative risk could be appreciably smaller than assumed. If so, the degree to which the relative risk estimate decreases toward 1.0 after adjustment for the current CD4⁺ lymphocyte count (CD4⁺ percent) could be somewhat larger than that reported. Although one would expect a proportion of any random excess risk in the placebo arm that has arisen because of chance alone to act via the CD4⁺ lymphocyte count, an appreciable part of it will not.

Further, as the authors point out, the inability to measure CD4⁺ lymphocyte counts accurately tends to underestimate its association with the risk for AIDS and therefore its ability to explain treatment differences in subsequent clinical progression.

Clearly, not all of zidovudine's effect on the progression to AIDS will ever be statistically explained by its effect on the CD4⁺ lymphocyte count (or CD4⁺ percent) alone. Nonetheless, the possibility remains that the proportion of the potentially explainable effect is greater than the upper value of 37% provided by Choi and colleagues.