Keiding, Niels^a; Slama, Rémy^b

An important aim of reproductive epidemiology is to document environmental or occupational exposures or lifestyle factors that affect human fecundity—that is, the ability to obtain a pregnancy. Fecundity is often approached by fecundability (defined as the probability that a couple with unprotected intercourse will conceive in a given menstrual cycle¹,²), as assessed in time-to-pregnancy surveys.

Ideally, such surveys should contain information on sexual activity in each menstrual cycle, in order to exclude cycles with no activity and therefore no possibility of conception. In practice, this information is not always available, particularly in a retrospective setting.

In their paper in this issue of the journal, Mumford et al³ discuss as their main example a new aim of time-to-pregnancy surveys. They propose estimating the distribution of the actual time it will take for a couple to get pregnant (ideally measured in menstrual cycles, with origin at the date they decided to start the attempt, termed “initiation” below), under realistic, rather than optimal, timing of their sexual practice. It is claimed that such a distribution would be useful in advising couples who want to get pregnant, many of whom do not have optimal behaviors in terms of timing of intercourse. Thus, the target is changed from (1) time of unprotected intercourse until pregnancy to (2) actual time-to-pregnancy in the real world. In practice, many retrospective time-to-pregnancy studies, in which collecting cycle-specific information on intercourse is challenging, already follow this approach.

Rather than a change in target, Mumford et al seem to consider sexual activity as the treatment and the event of getting pregnant as the outcome. This yields an analogy to the intention-to-treat (ITT) approach to analysis of clinical trials, where subjects are classified according to assignment of treatment, independent of compliance. Mumford et al emphasize that, just as is the case for other ITT studies,⁴ “the portability of this ITT/effectiveness estimate depends on the assumption that the noncompliance in the target population will be the same as seen in the trial population,” which is a standard claim for external validity.

This emphasis on the concrete circumstances in the target population deviates from the view going back to Miettinen,⁵ and still heralded by influential epidemiologists,⁶ that epidemiology is a “science” for which “… the generalization from the actual study experience is not made to a population of which the study experience is a sample in a technical sense of probability sampling … In science the generalization is from the actual study experience to the abstract, with no referent in place or time.”⁵

Mumford et al also mention the possibility that in an analytic/etiologic study, exposure may be associated with “time at risk” (here, timing of intercourse), and they claim that sometimes “exclusion [of time at risk] may lead to bias, and … it may be advantageous to include person-time not-at-risk.”³

With no examples, this point is hard to judge. However, our preference would be that if an exposure is suspected of acting through time-at-risk/pattern of intercourse, this should be studied separately in an analysis of exposure effects on intercourse pattern.

It is worthwhile to reflect on how to choose a good design for this new target.

The couple interested in becoming pregnant will of course be looking to the future and has no guarantee that their attempt will succeed; thus, advice to them should consider the possibility that it will not succeed. This rules out survey designs such as the most commonly used retrospective pregnancy-based design, conducted by interviewing couples who are already pregnant about how long time it took them to reach that goal. The target of this design is the conditional distribution of time-to-pregnancy given that pregnancy has happened, and so it has no way of estimating the probability that pregnancy never happens. Furthermore, as clearly explained by Basso et al⁷ in 2000 (although overlooked by most later authors, as noted by Keiding et al.⁸), the conditioning implies that pregnancy has to happen before the couple gives up trying; a “standard” analysis (ie, ignoring competing risks) will erroneously interpret any differential persistence-in-trying associated with exposure as differential fecundability between exposed groups.⁷,⁸

Mumford et al illustrate their commentary with data from the early time-to-pregnancy survey by Baird and Wilcox.⁹ This survey had the drawbacks of not only being pregnancy based but also explicitly self-selected (ie, with no sampling frame); therefore, it was not in line with the above requirements for the ideal design for the new purpose. (Baird and Wilcox, too, were concerned about the self-selection issue. They wrote “Of primary concern is any source of bias that might result in finding an association in our study population even if no true association exists in the general population.”⁹ They went on to perform an admirable sensitivity analysis on the question of differential occurrence of accidental pregnancies among smokers and nonsmokers, long before sensitivity analyses were routine.) The elaborate calculations that Mumford et al build on top of data from the Baird and Wilcox study do not address the prospective distribution of time-to-pregnancy relevant for advising couples wishing a pregnancy, but similar calculations in other designs might be informative.

In view of the need for information about the unconditional distribution of time-to-pregnancy, an obvious design choice would be the prospective design based on following a cohort of couples trying to conceive forward in time from initiation of their attempt at conception. Primarily because it is difficult to recruit couples at initiation of their pregnancy attempt, such studies are rare (see, Buck Louis et al.¹⁰ for a recent survey), and it is almost impossible to assess the representativeness of such a cohort within its study base, which would be required to provide a description of the probability of pregnancy according to time to a couple’s starting a pregnancy attempt.

An interesting approach to the recruitment problem for prospective time-to-pregnancy studies is to use the Internet, as exemplified by a current time-to-pregnancy survey in Denmark,¹¹ with more than 8500 women recruited so far (personal communication, 2014). Women are recruited shortly after initiation of attempt at conception, and then followed until pregnancy, giving up trying, or 12 cycles after initiation—whichever occurs first. Using appropriate delayed-entry survival analysis, this study would, in principle, deliver the desired prospective estimates of time-to-pregnancy, if we could believe that the sexual behavior (and other characteristics) of participants reflected the general Danish pregnancy-seeking population. That has not been a central issue so far for the study, for which the primary aims were etiologic. Publications from this project use arguments such as “Differences between study participants and non-participants do not affect the validity of internal comparisons within a cohort study of volunteers, which is the main concern. Given internal validity, the only problems with studying Internet users would occur if the biologic relations that we are studying differed between Internet users and non-users, a possibility that seems unlikely.”¹¹

These points bypass the self-selection issue. (It is unclear that all Internet users have the same selection probability and that probability of selection is independent of exposures and fecundity. This approach also disregards alternative targets, such as the one proposed by Mumford et al.) It is therefore encouraging that researchers with this study have now (personal communication, 2014) decided to check representativeness of main parameters against register data for all births in Denmark.

More generally, the prevalent cohort setting,⁸ which allows prospective collection of information on intercourse (again, with proper delayed-entry analysis), appears as a relevant one to consider for the approach by Mumford et al.

A final possibility might be to use the current-duration design,⁸ based on asking a cross-sectional sample of women (or men) from couples trying to conceive about the time they have spent so far trying. This has proved feasible in France¹² and the U.S.¹³,¹⁴ This design also allows for the fact that pregnancy does not always happen; the target is the distribution of the time of pregnancy attempt (ie, the minimum of time-to-pregnancy and the time to giving up the attempt),⁸ which, from a viewpoint of practical relevance as advocated by Mumford et al, might be considered relevant. Current-duration surveys may be somewhat easier to carry out than prospective time-to-pregnancy surveys (although the eligibility rate remains low), and their representativeness may be assessed. For example, in a nationwide implementation¹² of this design, a participation rate (63%) could be assessed and selection bias assessed and corrected for.

The new approach by Mumford et al. will strongly rely on the notion of planning a pregnancy. This is challenging because there is probably a lot of between-country and between-subject heterogeneity in this notion. For example, in our experience, some couples may consider that, after some months of unsuccessful trying, they are subfertile and not trying to become pregnant anymore (possibly because they think they cannot). The classical approach, consisting of defining cycles at risk on the basis of the occurrence of unprotected intercourse in a given period, may suffer from less between-couple and between-culture variability.

Mumford et al have formulated a new target for time-to-pregnancy studies, and, as was to be expected, this new target requires new design considerations. The new concept should not be used as a way for time-to-pregnancy studies with poor data on intercourse and months of separation of the couples to claim that they focus on a more meaningful outcome from a public health perspective than if information on intercourse had been taken into account. On the contrary, it should be used to call for more thorough collection of data on intercourse, to allow both analyses with and without cycles not at risk, and analyses of any effects of exposure on intercourse pattern. Mumford et al rightly point out that, to avoid immortal-time bias in other fields of epidemiology, the analysis excluding person-time (or cycles, in the analogy with time-to-pregnancy) is considered to be the one less prone to bias, and we believe that this also generally applies to time-to-pregnancy studies. In our view, the most crucial aspect of the examples of Mumford et al is their greater dependence on concrete facts of the sampling frame, highlighting the differences in targets for various time-to-pregnancy designs and the complex notion of trying for a pregnancy, and emphasizing that concrete attention to external validity is still necessary.

ABOUT THE AUTHORS

NIELS KEIDING is a professor of biostatistics at the University of Copenhagen, Denmark. His methodological research is mostly in survival and event history analysis and in epidemiologic methodology. For the last 10 years, he has been engaged in designing and analyzing a large time-to-pregnancy study in France. RÉMY SLAMA is an environmental epidemiologist and a Senior Investigator at Inserm (the main public institute in the field of biomedical research in France), currently based in Grenoble, where he created the team of environmental epidemiology applied to reproductive and respiratory health.

REFERENCES