Dr. Aursnes and others question the use of a nonexperimental study design to evaluate drug effects. We are aware of the limitations of nonexperimental assessment of efficacy or side effects of medications, including the methods we applied to limit bias that resulted from lack of random allocation. We emphasize, however, that the case–control method has important advantages over randomized trials and may be the only way to study rare outcomes, such as sudden death, different antihypertensive medications, and the influence of dosage and duration of drug use [1]. Obviously, the published trials lacked the power to address these issues [2]. Although our finding of an increased risk for sudden death among recipients of ß-blockers was unexpected, we disagree that this result contradicts evidence from randomized trials. Although few trials in hypertensive patients compared ß-blockers with other antihypertensive medications, findings from the Medical Research Council trial in the elderly [3] suggest that ß-blockers, compared with potassium-sparing diuretics, may increase the risk for coronary heart disease. Moreover, our findings are similar to those reported in a previous case–control study [4].

The concern of Dr. Aursnes and of Drs. Cabell and Oddone, that multivariate analysis may not fully remove differences in prognostic characteristics between users of different antihypertensive medications, is legitimate. Given our choice of study population and reference exposure (antihypertensive medication without non-potassium-sparing diuretics and ß-blockers), these differences were small and could be adjusted for in the analysis. Although residual confounding can never be ruled out in a nonexperimental study, it cannot fully explain our findings. In contrast to Cabell and Oddone's criticism, we did not perform "multiple post hoc analyses." Multivariate analysis was necessary to prevent confounding by indication, and our data analysis (including identification of the major potential confounders) was planned in the design phase of the study. It makes no sense to adjust the level of statistical significance when doing a single multivariate analysis to control for confounding. The prerequisite that case–control studies should produce relative risks of at least 3.0 to be of value reflects a misunderstanding of the potential of case–control studies and of the distinction between validity and precision.

We agree with Dr. Horowitz that the difference in risk for sudden death seen between non-potassium-sparing diuretics and potassium-sparing diuretics could result from an antiarrhythmic effect of the latter. A randomized trial to assess whether potassium-sparing diuretics prolong life, preferably compared with non-potassium-sparing diuretics, could certainly be valuable, but the sample size required may be prohibitively large.

Dividing the ß-blocker group into ß₁ selective and nonselective drugs, as suggested by Dr. Buck, may help explain our finding of an increased risk for sudden death among users of ß-blockers. In addition, our data set may contribute to the ongoing debate on the potential harmful effect of calcium antagonists in hypertensive patients, as mentioned by Dr. Winkelmann. Both analyses will be finalized soon.

Although Drs. Maun and Concato correctly note that the crude relative risks were statistically insignificant, their statement that the findings are clinically insignificant and that such as "negative" study does not warrant further analyses is surprising. Obviously, relative risks as large as 1.7 (that is, a 70% increased risk) are clinically important given the wide use of diuretics and ß-blockers. Further, as mentioned previously, multivariate statistical techniques are necessary in nonexperimental studies to adjust for confounding that may result from nonrandom allocation to medications. Whether the crude analysis produced a statistically significant result is irrelevant in this context.

Case-patients and controls are bound to differ in several clinical characteristics. It is no surprise that persons at the highest risk are more likely to become a case-patient. The relative risk seen after adjustment of potential confounders, however, depends on the distribution of these confounders among the various categories of antihypertensive medication, including the reference exposure group. In fact, all risk estimates changed in the expected direction after we included a confounding variable. Thus, we do not share the skepticism of Drs. Maun and Concato.