Editorial Comment—Antibiotics and Stroke
It is sobering to consider that despite its prevalence and prominence, the cause of atherosclerosis remains unknown. In this issue of Stroke, Brassard et al have addressed its intriguing association with inflammation. If infection plays a role in the genesis of vascular events such as stroke, might antibiotics reduce their incidence?
Briefly, Brassard et al examined retrospectively a large data set of elderly hypertensive patients beginning therapy and, after adjustment for confounders, found that antibiotic use, particularly penicillin, was negatively associated with stroke. They noted a “protective association.”
Enthusiasm for a conclusion supporting a popular and intriguing hypothesis should be tempered by consideration of exactly what was done and what was found. Two principles are helpful in assessing the strength of this evidence:
First, in science, an answer cannot be more specific than the question it addresses.
Second, causation cannot be deduced from association.
With regard to the first issue, the hypothesis of the present study was that “subjects treated with antibiotics may be at lower risk of developing clinical manifestations such as cerebrovascular disease.” Of the 6 defined antibiotic groups, 3 were associated with fewer strokes, 3 with more. (Patients who had received penicillins, macrolides, and fluoroquinolones experienced somewhat fewer strokes; those receiving tetracyclines, cephalosporins, and other antibiotics experienced somewhat more.) Overall the odds ratio was 0.99 for current use, 0.98 for recent, and 0.99 for past use. Adjustment for confounders resulted in odds ratios of 0.80, 0.81, and 0.87, respectively, and a “statistically significant” negative association for penicillin administration.
Since it was statistical adjustment that resulted in the association, statistical adjustment might be responsible. A problem with retrospective studies is that one can adjust only for factors for which data are available. An overall association shown to be robust when adjusted for confounders is a good deal more persuasive than an association that appears only after statistical manipulation, particularly when the calculations are limited by the availability of data collected for another purpose.
It is unfortunate and confusing (and unnecessary) that contemporary statistical methods use the same calculations and units for predictions confirmed as for observations noticed. Confirmed predictions carry “predictive value”; they suggest that a similar maneuver, in similar patients, will have similar results. Not so with observations. All observations will be “statistically significant” given only adequate sample size. The effects are beyond the plausible range of the play of chance, or they would not have been noticed. The “statistical significance” of an observation does not reflect its predictive value.
The association of antibiotic type with stroke incidence varied a good deal, and of course one association must always be strongest. That past penicillin administration was associated with the least stroke incidence constitutes only an unexpected observation. Of course, the explanation for the association might be that penicillin protects from stroke, but in the context of this investigation, it provides only a hypothesis that might be assessed using other data. Of course, confirming the association would not establish that it was causal.
The second issue is that causality cannot be deduced from association. Although causality usually produces association, association does not of itself imply causality. More than 20 observational studies showed that women taking hormone replacement therapy (HRT) had fewer heart attacks, yet all of the randomized trials (from which causality can be deduced) showed that HRT was associated with an increase in heart attacks. The explanation appeared to be due to either or both survivor effect (high-risk women given HRT had events prior to recruitment to the observational studies, leaving low-risk women to be observed) and association of HRT use with socioeconomic status. It is puzzling that HRT was never prescribed to raise socioeconomic status, despite a similar association and identical logic.
If a negative association between penicillin use and stroke incidence were to be confirmed with other data, causality would still not have been established. This means that “protective association” overstates the strength of the evidence. A negative association no more means that antibiotics prevent stroke than that HRT prevents heart attacks.
Notwithstanding these issues, the investigators are to be congratulated on having squeezed the available data to shed light on an important issue. But the jury remains out.