COMMENTARY Confounding and effect-measure modification in the evaluation of immunogenic agents Kenneth J. Rothman 1,2 & Barbara E. Mahon 1,3 1 Department of Epidemiology, Boston University School of Public Health, Boston, MA; 2 Division of Preventive Medicine, Department of Medicine, Boston University School of Medicine; 3 Department of Pediatrics, Boston University School of Medicine, MA, USA In this issue, Swift and Hunter [1] demonstrate some important pitfalls to avoid in studies of infectious diseases that give rise to protective immunity. Using several scenarios, they show that the causes of such disease may paradoxically appear to have small or even negative effects on disease occurrence in epi- demiologic studies that fail to take into account immunity from earlier exposure. They also show how this problem worsens with age and with the increas- ing prevalence of exposure. Swift and Hunter describe the effects of two causes of a disease, which they label A and B. Because exposure to either A or B provokes an immune re- sponse that eventually lowers risk of a specific disease, it is convenient to think of A and B as two different routes of exposure to the same causal agent. For example, exposure to Salmonella can produce salmo- nellosis either through contact with undercooked chicken or through the handling of reptiles. Although the routes of exposure are different, the causal paths converge at the more proximal cause, infection with Salmonella, which is the same for the two exposure routes. Biologically, then, the exposures A and B that Swift and Hunter describe are exposures to the same causal agent, though they may occur in different ways and at different times. The timing of exposure is important, because the first exposure to a cause does not necessarily have the same effect as a subsequent exposure. For some causal agents, repeated exposure can push the cumulative dose over a threshold and have a supra-linear effect. The carcinogenic effect of ionizing radiation or DES are possible examples of supra-linear effects (though the evidence is not defin- itive for either). For exposures that provoke an im- mune response, however, repeated exposures are likely to have a smaller effect than initial exposures. An immunogenic cause of disease has two distinct effects: one is to increase disease risk in the short term; the other is to decrease disease risk over the longer term by stimulating a protective immunity (Figure 1). The relation between the causal agent and these two effects is complicated. If exposure to the agent results in clinical disease, an immune response resulting in some protection against disease recur- rence is likely to follow. The immune response can occur, however, even if the initial exposure to the agent does not provoke clinical disease. If the immunity is long-lasting, repeated exposures to the agent itself or to various sources that are likely to contain the agent may have no clinical effect, but may serve to boost immunity, further decreasing the risk of developing disease. Because the effect of a second or subsequent exposure will differ markedly from that of the first exposure, ignoring the history of earlier exposure when assessing current exposure is likely to result in confounding of the effect of the current exposure by the effect of earlier exposure. The direction of this confounding is negative. If there is no control of past exposure in study design or analysis, the effect of a current exposure, compared with those who have never been exposed, will be a mixture of two effects: the increased risk for those for whom the current exposure is a first exposure, and the decreased risk for those for whom the current exposure is not the first exposure. With a strong and long-lasting immune response, and a strong correlation between first and subsequent exposure, the confounding effect can be large enough to negate or even reverse any positive effect of exposure. Without naming it as such, this confounding effect is what Swift and Hunter have described in their paper. They assume that exposure to A and to B occurs independently in a specific year, but in their first model, which illustrates the confounding effect, they assume that fixed subsets of the population can be exposed to either A alone, or to A and B, throughout life. They suggest that one can measure the effect of exposure B by comparing the risk of disease among those who can be exposed to both A and B with the corresponding risk among those who can be exposed only to A. (They do not explain why they chose this comparison in preference to compar- ing the exposed groups to an unexposed group.) Both A and B are routes of exposure to the same under- lying cause. In their models, disease can occur only among those with no previous disease, so the disease risk declines with age along with the size of the sus- ceptible pool. If the disease risk has been high enough to deplete the susceptible pool early in life, then late in life disease risk will be low. The more likely the exposure to A or B in their model, the lower the risk European Journal of Epidemiology 19: 205–207, 2004. Ó 2004 Kluwer Academic Publishers. Printed in the Netherlands.