To the Editor:
Penelope Howards1 is to be congratulated for revisiting to the problems we face when we analyze more than one pregnancy per woman in reproductive epidemiology. Most reproductive epidemiologists have probably been met with requests to include robust standard errors in their analyses, taking account of the dependency among pregnancies—a request that may be valid in animal experiments where some adjustments for litter size are needed but makes little sense in human populations. The problem is that pregnancies are intensively surveyed and risk factors are modified between pregnancies. These modifications are decided by the woman herself or by her many advisors, and are based upon past pregnancy experience that may be difficult to capture accurately by directed acyclic graphs (DAGs). Howards presents a number of scenarios that are probably all valid for subsets of a larger study in reproductive health. She did not, however, include a situation that is of frequent concern.2 The problem is that many exposures are consider to be causes for reproductive failures, although they need not be, or at least not for the pregnancy in question. We quite simply do not know.
A woman may, for example, have a preterm birth while working night shifts. She may think that night shifts caused the preterm birth, and so when she wants to become pregnant again she gives up shift work. If the shift work was not the cause of the preterm birth and the unknown actual cause is carried over to the next pregnancy (as in DAGs 2–5 in Howards’ paper except without an arch between E0 and O0), it may become difficult to obtain comparability between shift workers and daytime workers in this second pregnancy. The comparison will be confounded by the unknown causes of preterm birth. These may now be unequally distributed between shift workers and nonshift workers in the second pregnancy, but not necessarily in the first.
Previous advice2 has been to restrict the analyses to first pregnancies. This avoids modifications of preventive actions driven by a previous reproductive experience and induced by health care workers or the women themselves, and thus avoids a distortion of comparability because of past experience. I am not sure we should give up this advice. It may be too much to hope some statistical model can adjust for self-selection by the mother’s stopping reproduction or changing modifiable perceived risk factors when trying having another child.
DAGs have been a useful tool to identify colliders and to display a number of possible analytical strategies. Howards presents 5 out of many possible scenarios, and she reminds us that there is no single “right strategy” except for a subset of the population that we often cannot identify. We are still left with the task of trying out different analytic strategies—not to find a publishable result, but to see if we can make spurious associations go away.
Department of Epidemiology
School of Public Health
Los Angeles, CA
1. Howards PP, Schisterman EF, Heagerty PJ. Potential confounding by exposure history and prior outcome. Epidemiology
2. Olsen J. Options in making use of pregnancy history in planning and analyzing studies of reproductive failure. J Epidemiol Community Health