Invited Commentary: Influenza, Influenza Immunization, and Pregnancy—It's About Time

Abstract

Immunization of pregnant women against influenza has the potential to reduce adverse fetal outcomes by reducing prenatal exposure to influenza illness. However, as touched on by Fell et al. (Am J Epidemiol. 2016;184(3):163–175) and Vazquez-Benitez et al. (Am J Epidemiol. 2016;184(3):176–186) in this issue of the Journal, observational studies in which the causal effect of maternal influenza illness and influenza immunization on fetal health are evaluated are prone to bias because of the complex temporal nature of influenza illness seasonality, influenza immunization schedules, and gestation itself. Immortal time bias is introduced by an “anytime-in-pregnancy” exposure definition because the shortened pregnancy duration associated with many adverse fetal outcomes limits the opportunity to become exposed, whereas including follow-up time during which pregnancies are no longer at risk of an adverse outcome (e.g., gestational time after 37 weeks in studies of preterm birth) can lead to overestimation of any true benefits of immunization (or harms from influenza illness). We present a framework to avoid time-related biases in the study of influenza illness and immunization in pregnancy and advise that investigations of fetal benefit from maternal influenza immunization should only be undertaken when information is available on the calendar time of influenza virus circulation and the gestational age at which maternal influenza immunization occurred.

Numerous public health organizations recommend that pregnant women be given high priority for seasonal influenza vaccination (1, 2). These recommendations are based on clear evidence that vaccination reduces the risk of maternal influenza illness and improves neonatal immunity (3, 4). However, uptake of maternal influenza vaccination programs and investment in vaccine delivery programs for pregnant women have been limited, particularly in low-income countries. Stimulated by increased interest after the 2009–2010 H1N1 pandemic, there have been several observational studies in which investigators reported a reduced risk of adverse pregnancy outcomes (e.g., stillbirth, preterm birth, and fetal growth restriction) associated with maternal influenza immunization (5–7). These findings have generated substantial enthusiasm in the public health community, with growing advocacy for low-income countries to invest in maternal influenza vaccination programs as a strategy to reduce fetal morbidity (8). However, the cost-effectiveness of such investments is predicated in part on the assumptions that influenza illness is harmful to pregnancy and that immunization will therefore lead to improved pregnancy outcomes by preventing maternal influenza illness.

In this issue of the Journal, research by Fell et al. (9) and Vazquez-Benitez et al. (10) calls into question the scenario of reduced risk of adverse fetal outcomes from maternal influenza immunization. Using a time-series analysis, Fell et al. observed no association between weekly levels of circulating influenza illness and the adverse pregnancy outcomes of preterm birth, stillbirth, and perinatal mortality, whereas Vazquez-Benitez et al. demonstrated that an association between maternal influenza immunization and preterm birth is greatly attenuated by the use of appropriate statistical methodology.

These studies highlight the major methodological challenges that complicate epidemiologic research on the effects of influenza illness and immunization on pregnancy outcome. The potential for confounding due to unmeasured differences in the health profiles of women who do or do not receive influenza immunization or contract influenza illness is well recognized, as noted in studies of influenza vaccination among the elderly (11). Additional issues, such as limited statistical power due to the modest fraction of women who develop influenza and variable vaccine efficacy across influenza seasons, have also been noted (12). More complex and less well described is a challenge that was touched on by Vasquez-Benitez et al. (10) in particular but is also pertinent to the time-series study by Fell et al. (9): the temporal nature of seasonal influenza infection, vaccine availability, and pregnancy. In the present commentary, we elucidate the importance of carefully considering calendar time and gestational time to provide unbiased estimates of the effects of seasonal influenza illness and immunization on fetal health.

SEASONALITY OF INFLUENZA ILLNESS AND IMMUNIZATION

In temperate climates, the risk of influenza infection is seasonal, with outbreaks in the northern hemisphere typically occurring between October and May. Within an influenza season, peak influenza activity is variable from region to region and year to year. Assuming that fetal benefits of maternal influenza immunization are achieved through prevention of influenza illness, immunization will only be effective during periods when influenza illness is circulating. Yet, not all studies in which the association between influenza immunization and fetal outcome have been examined account for the timing of local influenza activity, so that any increased rates of adverse perinatal outcome associated with lack of influenza immunity will be diluted by the lack of impact of vaccination when influenza illness is not circulating. In response to the seasonality of influenza, influenza immunization is also seasonal, with the vaccine being available in the pre-influenza season (fall in the northern hemisphere).

TEMPORAL CONSIDERATIONS IN PREGNANCY

Most pregnancy outcomes potentially affected by influenza illness and immunization are dependent on gestational age. Preterm birth and miscarriage are defined by a shortened gestational duration, whereas infants who are stillborn or die during infancy have systematically younger gestational ages at delivery than do babies that survive infancy. In studies of preterm birth, for example, pregnancies cease to be at risk of preterm birth once 37 weeks' gestation is reached. If there is a true protective effect of immunization against preterm birth, pregnancies in the immunized group will be systematically more likely to have follow-up time that was accumulated after 37 weeks (no longer at risk of preterm birth). Including follow-up time after 37 weeks will therefore add more time to the denominator of the exposed group compared with the unexposed group, leading to an overestimation of the beneficial effects of vaccination. The reverse effect, an overestimation of any excess preterm birth risk due to influenza, would be expected from failing to censor pregnancies at 37 weeks in studies of influenza illness, with exposed pregnancies more likely to end early (preterm) than unexposed pregnancies. A similar bias will occur in the study of miscarriage, which will not occur once the dividing line for classification as a stillbirth is reached at 20 weeks' gestation. Although the gestational age–dependent natures of stillbirth and infant death do not require censoring of follow-up time, they creates a differential opportunity for exposure during pregnancy, which can introduce bias, as outlined below.

The time at which pregnancies begin to be at risk of experiencing an adverse event is also dependent on gestational age. Preterm birth and infant death require an infant to be born alive, which is rarely documented before 16–20 weeks. Likewise, stillbirths are typically only recorded after the gestational age at which legal registration of stillbirths is required. Including follow-up time that occurred before pregnancies are at risk of experiencing an adverse event should not bias relative measures of effect because the added follow-up time in the denominators is unlikely to be differential by exposure status. However, artificially increasing the follow-up time in the denominators will bias absolute measures of effect such as the rate difference.

Women's influenza illness and immunization statuses change over the course of pregnancy depending on the calendar time during which influenza illness is circulating and immunization is offered. In observational studies with time-varying exposures such as influenza and immunization, inappropriate accounting of follow-up time through use of a time-fixed analysis may result in immortal time bias (13). In pregnancy, immortal time bias arises when entry into the exposed group depends to some extent on a woman remaining pregnant and free of adverse outcomes (such as stillbirth) long enough to have the opportunity for exposure (14–17). The time before exposure is considered “immortal” because any adverse event that happens before a participant had the opportunity to become exposed would result in the event being assigned to the unexposed group. In other words, until she is exposed, she cannot possibly become an exposed case of disease because all participants start as unexposed. This apparent survival advantage in the exposed group creates a spurious appearance of a protective effect of exposure. For example, previously observed protective effects of decongestant use in the third trimester on preterm birth have been shown to reflect instead that women who remains pregnant longer (i.e., who do not deliver preterm) have more opportunity to use decongestants (17).

In the study of influenza immunization, the potential for immortal time bias arises because the opportunity for immunization increases the longer a woman remains pregnant. Consider a woman who is 34 weeks' pregnant on October 1, the first day of vaccine availability for the upcoming influenza season in her region. If she remains pregnant until her next scheduled antenatal visit at 36 weeks' gestation on October 14, she may be vaccinated by her primary care provider (or receive a recommendation to obtain immunization through a public health unit, pharmacy, or workplace). However, if she goes into spontaneous labor on October 2nd, her preterm birth will be assigned to the “unvaccinated” group. In a time-fixed analysis, the time between the start of the study follow-up and vaccination becomes immortal, because remaining pregnant and free of adverse events during this period is necessary to become vaccinated. The immortal follow-up time is incorrectly assigned to the “vaccinated” group, creating a spurious negative association between vaccination and preterm birth by artificially inflating the denominator in the preterm birth rate for vaccinated women. A similar concern exists in the study of influenza illness, in which opportunity to contract influenza increases the longer a woman remains pregnant, which tends to lead to underestimation of any true adverse effect.

The findings of Vazquez-Benitez et al. (10) suggest that immortal time bias may be responsible for a comparable or greater degree of bias than confounding in observational studies of influenza immunization and preterm birth. In their study, accounting for immortal time attenuated the naïve risk ratio of 0.79 (95% confidence interval: 0.74, 0.85) to 0.88 (95% confidence interval: 0.82, 0.94), whereas further adjustment for confounding (by covariates routinely available in a large clinical database) only attenuated the estimate from 0.91 (95% confidence interval: 0.84, 0.98) to 0.92 (95% confidence interval: 0.84, 1.00). The impact of immortal time bias was greatest in the third trimester (risk ratio = 0.63, 95% confidence interval 0.55, 0.72), when the likelihood of women going into preterm labor before an opportunity to be vaccinated is greatest (because the probability of delivery increases markedly with advancing gestation). We speculate that outcomes that occur at systematically younger gestational ages than preterm births (such as stillbirth and infant death) will have an even greater degree of bias because of immortal time. In contrast, the outcome of small-for-gestational-age birth showed no bias due to immortal time. By definition, small-for-gestational-age birth (weight <10th percentile for gestational age) is independent of gestational age; thus, prolonged duration of pregnancy should not provide a survival advantage against small-for-gestational-age birth.

Finally, pregnancy is not a homogeneous period at risk. Gestational age–specific effects are obvious for such outcomes as congenital anomalies and miscarriage and are thought to be relevant for later pregnancy outcomes as well. For example, the effect of smoking on birth weight appears to be specific to the late-pregnancy period (18). As described by Hertz-Picciotto et al. (19), the overall prevalence of influenza illness at any point in pregnancy is considerably higher than its prevalence at a specific point in pregnancy. Exposures such as this are particularly prone to misclassification when an “anytime-during-pregnancy” exposure definition is used with an agent that only exerts its effects in a limited period of development (19), adding inconsequential exposure to exposure in the time periods that are relevant to disease etiology. This tends to bias effect estimates towards the null for gestational age–dependent outcomes such as miscarriage, preterm birth, stillbirth, and infant death. Although it is unknown whether there is a critical time window during which influenza illness exerts its effect on fetal health, an analysis plan that fails to allow for narrower time windows of effect reduces the potential to detect harm from influenza illness or benefits from immunization. For example, Fell et al. isolated the first month of gestation as a potentially sensitive time window for exposure to influenza illness (9). They considered additional exposure time windows in their study, although those exposure windows were isolated relative to time of delivery (i.e., week or month before delivery) rather than the gestational age window.

RECOMMENDED APPROACH TO ADDRESSING TEMPORAL FACTORS IN STUDYING INFLUENZA ILLNESS, INFLUENZA IMMUNIZATION, AND PREGNANCY OUTCOME

The need to account for change in exposure status during the course of pregnancy emphasizes the importance of collecting information on date of vaccination or influenza illness instead of only noting whether exposure occurred at some (unknown) point during pregnancy. Information on local periods of influenza virus circulation is also critical. With these data, analytic plans can appropriately account for the influences of calendar time and gestational time. A summary of strategies to reduce bias is provided in the Appendix.

First, gestational days (or weeks), rather than pregnancies, should be used as the unit of analysis. Follow-up time should be restricted according to the time periods during which each adverse perinatal outcome can plausibly occur. For studies of preterm birth, all pregnancies should be censored at 37 weeks' gestation because illness and immunization status in the remaining weeks of pregnancy when the period at risk of the outcome has passed have no relevance. For studies of stillbirth, infant death, and preterm birth, follow-up time should only begin after the lower gestational age limit for stillbirth or birth registration.

Influenza illness or immunization status should be classified on a time-varying basis. Follow-up time before influenza illness or immunization should be classified as “unexposed.” Immortal time bias can then be prevented by conducting time-dependent analyses using Cox proportional hazards models or alternative approaches, such as the use of a time-matched, nested case-control analysis (13).

To account for the differential effect of immunization on adverse outcomes according to the presence of circulating influenza, calendar time of influenza illness activity should be accounted for as an effect measure modifier. Analyses can include a term for the interaction between immunization status and calendar time (or other regression-based approaches to account for effect measure modification), restrict follow-up time to the period during which influenza is circulating, or stratify by time period. Stratified analyses can use the estimated effect of immunization in periods without circulating influenza as a negative control to evaluate the potential for confounding because any benefit from vaccination in the absence of circulating influenza must be due to factors other than the prevention of influenza illness (setting aside the possibility of a nonspecific benefit from immunization itself). Controlling for influenza virus circulation as a time-dependent covariate is not appropriate because doing so will produce an estimate of effect that is the weighted average of the vaccine's effects during the influenza season and noninfluenza season.

Finally, the temporal considerations regarding both influenza and pregnancy should be refined to ensure that the hypothesized effect is examined with as much precision and statistical power as possible. To the extent that intervals of pregnancy vary in their susceptibility to influenza-caused adverse outcomes, this time-dependent analysis should be applied within intervals. If it is hypothesized that the harm from influenza and the benefits of immunization preventing influenza are specific to the third trimester before 37 weeks' gestation, we would consider the period at risk to start at 26 weeks' gestation and end at 37 weeks' gestation, assigning each unit of exposure in that window as exposed or unexposed to immunization and changing a participant's status if she is vaccinated during that window. Analogously, the time periods of relevance for observing a potential harmful effect of acquiring influenza or being protected from such harm through immunization would only occur during the influenza season. Therefore, the interval with and without circulating influenza should be examined as an effect measure modifier.

APPENDIX

Accounting for Calendar and Gestational Time in Studies of Maternal Influenza Illness, Influenza Immunization, and Pregnancy Outcome

1. Use gestational days (or weeks) as the unit of analysis rather than pregnancies.

2. Restrict follow-up time according to the time periods during which each adverse perinatal outcome can plausibly occur. For studies of preterm birth, all pregnancies should be censored at 37 weeks' gestation because illness and immunization status in the remaining weeks of pregnancy when the period at risk of the outcome has passed have no relevance. For studies of stillbirths, infant deaths, and preterm births, follow-up time should only begin after the lower gestational age limit for stillbirth or birth registration (typically 20 weeks). For small-for-gestational-age birth, the gestational age limit at which growth chart reference values begin (e.g., 22 weeks in the chart by Oken et al. (20)) should be used.

3. Account for the differential effect of immunization on adverse outcomes according to the presence of circulating influenza by treating calendar time of influenza illness activity as an effect measure modifier. Controlling for influenza virus circulation as a time-dependent covariate is not appropriate because this will produce an estimate of effect that is the weighted average of the vaccine's effects during the influenza season and noninfluenza season.

4. Classify influenza illness or immunization status as a time-varying exposure. Follow-up time before influenza illness or immunization (as well as the 2-week period needed to achieve full immunity) should be classified as “unexposed.”

5. Conduct time-dependent analyses using approaches such as a Cox proportional hazards model or a time-matched, nested case-control analysis (13).

6. Refine the temporal considerations regarding both influenza and pregnancy to ensure that the hypothesized effect is examined with as much precision and statistical power as possible. To the extent that intervals of pregnancy vary in their susceptibility to influenza-caused adverse outcomes, this time-dependent analysis should be applied within intervals.