A Prospective Cohort Study of Seasonal Variation in Spontaneous Abortion

Abstract

Background:

Up to 30% of pregnancies end in spontaneous abortion, yet few risk factors have been identified. Examining seasonal patterns in risk of spontaneous abortion can generate new hypotheses regarding environmental and lifestyle determinants.

Methods:

We used data from Pregnancy Study Online—a preconception cohort study of pregnancy planners from the United States and Canada—to examine seasonal variation in spontaneous abortion risk. We enrolled 12,197 women during 2013 to 2020, 6104 of whom reported a conception within 12 months of enrollment. On follow-up questionnaires, participants reported date of spontaneous abortion and weeks gestation at time of loss. We used periodic regression to estimate two aspects of seasonal occurrence: peak/low ratio—a measure of intensity of seasonal variation—and peak timing. We examined season at risk (from the date of each gestational week) in relation to spontaneous abortion; in a secondary analysis, we examined season of conception in relation to spontaneous abortion. We controlled for seasonal patterns in attempt initiation via month the pregnancy attempt began.

Results:

Almost 20% of women experienced spontaneous abortion. Risk was highest in late August, with a peak/low ratio of 1.3 (95% confidence interval [CI] = 1.1, 1.6). This seasonal pattern was evident almost exclusively for spontaneous abortion at <8 weeks since the last menstrual period date (peak/low ratio = 1.4; 95% CI = 1.2, 1.8), and associations were stronger among women living in the Southern and Midwestern United States.

Conclusions:

Environmental or lifestyle factors more prevalent in late summer may be associated with increased risk of early spontaneous abortion.

Up to 30% of detectable pregnancies end in spontaneous abortion, defined as a pregnancy loss before 20 weeks’ gestation.¹ Spontaneous abortion can cause substantial morbidity, including symptoms of posttraumatic stress, depression, and anxiety.² Chromosomal abnormalities, which occur either before pregnancy or during early pregnancy, account for around 50% of spontaneous abortions,³ and the remaining 50% are of unknown etiology. Few risk factors for spontaneous abortion have been identified. Therefore, identifying risk factors for spontaneous abortion is an important public health goal that could allow for the development of individual- and structural-level interventions to reduce the burden of this outcome.

The study of cyclic variations in health outcomes allows for the generation of hypotheses regarding causes, since cyclic patterns in an outcome imply cyclic patterns in its causes. For example, examination of seasonal variation in sudden cardiac death^4,5 led to the hypothesis that cold temperatures increase risk of this outcome.^6,7 Many environmental and lifestyle factors vary by season. Therefore, assessing seasonal variation in spontaneous abortion allows for the generation of more specific hypotheses regarding environmental or lifestyle causes of the outcome (Figure 1).

FIGURE 1.

Conceptual diagram depicting the various meteorological, medical, environmental, and lifestyle factors that are downstream consequences of season.

To our knowledge, there have been few studies on seasonal patterns in spontaneous abortion. Two hospital-based studies reported conflicting results. In a hospital-based registry from New South Wales, Australia, spontaneous abortion occurrence peaked in March.⁸ However, among women recruited from a fertility clinic in Kentucky, there was a bimodal annual pattern, with peaks in May and December.⁹ Studies that rely on case ascertainment in a clinical setting likely miss the majority of spontaneous abortions, which often occur early in gestation, sometimes before pregnancy detection.¹⁰ In addition, although fertility clinics closely monitor early reproductive events, the findings from fertility clinic data may not be applicable to couples not experiencing challenges conceiving, as the causes of spontaneous abortion differ in fertility treatment populations.¹¹

Women planning a pregnancy are more likely to use pregnancy tests early in gestation and may detect early losses more consistently than women not planning to conceive. Therefore, cohort studies that enroll pregnancy planners during preconception have distinct advantages in assessing spontaneous abortion risk. Nonetheless, only one preconception cohort study to our knowledge has examined seasonal variation in spontaneous abortion. Among 221 women from North Carolina, who had recently discontinued contraception, risk of early spontaneous abortion was highest for pregnancies conceived during early fall. The seasonal effect was strong: the proportion of pregnancies ending in early spontaneous abortion was nearly 40% when conceived in September and only 10% when conceived in May.¹²

In the present analysis, we use data from a large North American cohort of pregnancy planners to assess seasonal variation in spontaneous abortion risk.

METHODS

Study Design

Pregnancy Study Online (PRESTO) is a web-based preconception cohort study.¹³ Recruitment began in 2013 and takes place primarily online through social media advertising. Eligible women are of 21 to 45 years of age, residents of the United States or Canada, and are trying to conceive without use of fertility treatment. Participants complete an online baseline questionnaire that ascertains information on sociodemographics, lifestyle factors, anthropometries, and reproductive and medical histories. We send participants follow-up questionnaires every two months for up to 12 months to ascertain pregnancy status. Participants who conceive complete additional questionnaires in early and late pregnancy. This study was approved by the Institutional Review Board at the Boston University Medical Campus, and all participants provided online informed consent.

Exclusions

Between June 2013 and August 2020, 12,197 participants completed the baseline questionnaire. For our analysis, we included the 6104 participants who reported a pregnancy within 12 months of enrollment (eFigure 1; http://links.lww.com/EDE/B906). We analyzed each participant’s first reported pregnancy during the study period. Four percent of the cohort had not yet had the opportunity to complete a follow-up questionnaire, 24% were lost to follow-up before conceiving, and 21% completed 12 months of follow-up without reporting a conception.

Assessment of Spontaneous Abortion

We identified spontaneous abortion primarily using self-reported data. On follow-up questionnaires, we asked participants if they had experienced any pregnancy losses since their last questionnaire, including “miscarriage (including chemical pregnancy),” “induced abortion,” “ectopic/tubal pregnancy,” or “blighted ovum.” We also asked if they were currently pregnant and if so, the date of the first positive pregnancy test and the pregnancy due date. We identified spontaneous abortion that occurred after report of a current pregnancy from a late pregnancy questionnaire, sent at 32 weeks’ gestation. We asked if participants were still pregnant and if not, the reason, with possible responses including, “I had a miscarriage or a chemical pregnancy,” “I had an induced or therapeutic abortion,” “I had an ectopic pregnancy,” “I had a blighted ovum,” “I had a stillbirth,” and “I already had my baby.” We included reports of miscarriage, biochemical/chemical pregnancies, and blighted ova in our definition of spontaneous abortion. For pregnancy losses reported on all questionnaires, we asked participants for the date of loss and weeks of gestation at time of the loss (i.e., completed weeks since the last menstrual period [LMP] date).

For participants lost to follow-up, we attempted to identify outcome information by contacting participants via email or searching on social media for evidence of pregnancy loss or birth. We additionally identified participants who did not experience a spontaneous abortion by linking to birth registry data in selected states (California, Florida, Massachusetts, Michigan, New York City, Ohio, Pennsylvania, and Texas).

We have asked about use of home pregnancy tests since the study began in 2013 (i.e., “How was this pregnancy confirmed?,” with an option to select “home pregnancy tests” and allowing for multiple methods). Pregnant participants reported the following methods of confirmation of their pregnancy: urine test at home (95%), urine test in a doctor’s office (32%), gynecological examination (12%), blood test (38%), and ultrasound (39%). We added questions on timing and results of home pregnancy tests to our questionnaires in October 2017. The median weeks gestation at the first positive pregnancy test among women who had a study-related pregnancy was 4 (range = 3–13). We began follow-up time at the completed weeks gestation at the first positive pregnancy test. For women with missing data on pregnancy testing (68% of the sample, as questions on pregnancy testing were added to the questionnaires in July 2018), we began follow-up at four weeks of gestation. We ended follow-up at the occurrence of spontaneous abortion, induced abortion, ectopic/tubal pregnancy, completion of early pregnancy questionnaire (median 8 completed weeks’ gestation [interquartile range = 6–10] for women lost to follow-up after completing the early pregnancy questionnaire), or 20 weeks (for women who did not experience a loss).

Statistical Analysis

To measure seasonal variation in spontaneous abortion, we fit a periodic regression model with an annual cycle.¹⁴ This model assumes an annual unimodal cyclic risk with peak and trough exactly six months apart. We converted the date of the first day of each gestational week to an angle ($x$ = 360° × [day of the year]/365) and then used Cox proportional hazards regression models, with gestational weeks as the time scale, t (i.e., completed weeks since LMP), to estimate the hazard of spontaneous abortion:

$$\mathrm{h}(t)={{\mathrm{h}}_0(t)}^{\ast }\text{exp}({{\beta }_1}^{\ast }\cos (x)+{{\beta }_2}^{\ast }\sin (x)+\dots )$$

where $\mathrm{h}(t)$ is the probability that a spontaneous abortion occurs in week $t+1$, given that the pregnancy has survived past $t$ completed weeks. The Cox model allowed us to account for the gestational time when the pregnancy was recognized and assessment of outcomes could begin, as well as right censoring (i.e., loss to follow-up before 20 weeks’ gestation). From this model, we estimated two aspects of seasonal occurrence: peak/low ratio and timing of the peak. The peak/low ratio is a measure of the intensity of seasonal occurrence and is interpreted as the hazard of spontaneous abortion at the peak relative to the trough.¹⁵ We calculated the peak/low ratio as $\text{exp}(2\times \text{sqrt}[{{\beta }_1}^2+{{\beta }_2}^2])$. We estimated the timing of the peak as $\arctan ({\beta }_2∕{\beta }_1)$ and converted the estimated angle to a date, identified using the signs of the coefficients ${\beta }_2$ and ${\beta }_1$. In primary models, we defined season as the date of each gestational week at risk. In secondary analyses, we defined season as the date of conception.

We first ran models unadjusted for covariates. We evaluated the goodness of fit of these models with a 2-degree-of-freedom chi-square test, gauging the improvement in fit provided by the model including the sine and cosine terms (i.e., the model assuming a shifted sinusoidal function) compared with the null model where both sine and cosine terms equal zero.¹⁶

Seasonal variation in when couples begin trying to conceive¹⁷ can lead to seasonal variation in early reproductive events¹⁸ because the most fertile couples conceive close to their desired month of conception and may have lower risk of spontaneous abortion. Therefore, to determine the intensity of seasonal occurrence that was not due to seasonal variation in initiating pregnancy attempts, we fit a model adjusting for the month the pregnancy attempt began (as indicator variables), reported at baseline.

We fit separate models for gestational weeks <8 and ≥8 to assess proportional hazards and to determine whether season is more strongly associated with early or late spontaneous abortion. The influences on spontaneous abortion risk may vary across gestational ages through different biologic mechanisms including chromosomal abnormalities, problems with implantation or placentation, and fetal development. Our cut point of 8 weeks was informed by karyotype data demonstrating that a higher proportion of spontaneous abortion before 8 weeks is due to chromosomal abnormalities.¹⁹ We fit separate models across strata of latitude (<35, 35–39, 40–44, and ≥45° N) and geographic region (Northeastern United States, Southern United States, Midwestern United States, Western United States, and Canada) because seasonal exposure patterns may vary across North America. We also stratified by body mass index (BMI) to assess whether individuals with high BMI are more susceptible to seasonally varying exposures (e.g., heat).²⁰ Finally, we fit separate models by annual household income (<$100,000/year vs. ≥$100,000/year), as the downstream sequelae of season may vary by socioeconomic status. Results for each stratum are represented by a point on a polar graph, in which distance from the origin is proportional to the amplitude (indicating the strength of the peak/low ratio) and angle on the circle represents the timing of the peak.^21,22

We used fully conditional specification methods to multiply imputed missing data. We imputed gestational weeks at spontaneous abortion for the 6% of women missing these data. We generated five imputation data sets and statistically combined estimates across data sets.

The peak/low ratio is an unusual measure in that the null value and the minimum value are equivalent (1.00) because by definition, the peak can never be lower than the trough for any given sine curve. Therefore, even in the absence of seasonal variation, random error will generate a peak/low ratio above 1, thus introducing positive bias. To estimate the magnitude of the peak/low ratio for these data that could readily be explained by random error alone, we conducted a plasmode simulation,²³ similar to that conducted in previous studies.^7,24 In each of 1000 simulation trials, we randomly assigned a day of the year to the first day of the first gestational week of observation for each woman and then calculated dates for the subsequent weeks of observation. We kept all observed covariate and outcome data. From the simulated data, we fit the periodic regression model and estimated the peak/low ratio. We report the mean and the 2.5th and 97.5th percentiles from the 1000 simulations as an estimate of the expected peak/low ratio under the null hypothesis of no seasonal variation, along with the 95% simulation interval. This calculation estimates the bias in the peak/low ratio induced by randomness, and the interval describes the precision of the bias correction.

RESULTS

During 77,401 gestational weeks, the 6104 participants in our cohort reported 1188 spontaneous abortions (19.5%). The median weeks’ gestation at loss was 6 (interquartile range: 5–9). Most spontaneous abortions were reported as intervening losses between questionnaires (68%), whereas 29% were reported on the late pregnancy questionnaire and 3% were identified through social media or by contacting participants directly. Eighteen percent of the participants were censored at their gestational week at last contact (because they were either lost to follow-up or still <20 weeks pregnant), and 62% were censored at 20 weeks’ gestation. Loss to follow-up did not differ by season of enrollment (data not shown). Participant characteristics are shown in the Table. Most participants were non-Hispanic White (86%) and had a college degree or higher (79%). Almost half of the participants had annual household incomes of over $100,000 US dollars per year (47%). Participants resided in all 50 US and 10 Canadian provinces, with the largest proportion in the Northeastern United States (24.2%).

TABLE.

Preconception Characteristics of 6104 Female Participants in PRESTO (2013–2020)

Characteristica	%	Mean
Age (years)		29.9
Gestational weeks at first positive pregnancy test		4.2
Calendar year
2013–2014	22
2015–2016	19
2017–2018	41
2019–2020	18
Race/ethnicity
Non-Hispanic White	86
Non-Hispanic Black	2
Non-Hispanic Asian	2
Hispanic	6
Multiracial/other race	4
Educational attainment
≤High school	3
Graduate degree	43
Annual household income (USD)
<$50,000	16
≥$200,000	19
Geographic region
Northeast United States	24
South United States	22
Midwest United States	22
West United States	16
Canada	16
Latitude
≥45°	16
40–44°	46
35–39°	23
<35°	16
Sleep <7 hours/night	22
BMI (kg/m2)		26.8
Current regular smoker	5
Alcohol (drinks/week)		3.1
Sugar-sweetened soda (drinks/week)		1.2
Vigorous physical activity (hours/week)		2.7
Moderate physical activity (hours/week)		7.0
Sitting watching TV (hours/day)		2.4
Sitting doing something else (hours/day)		4.5
Gravid	51
Time to conception (cycles)		6.2
Flu vaccination in past 2 months	37
Medication use in past 4 weeks
Antibiotics	8
Migraine medication	8
Asthma medication	6
Hay fever medication	5
Antidepressants/antianxiolytics	12

^aCharacteristics measured at baseline or follow-up questionnaire most proximate to, but preceding, the pregnancy.

USD indicates United States dollars.

In the plasmode simulation, the mean peak/low ratio was 1.1 (95% simulation interval = 1.0, 1.3), indicating that observed peak/low ratios in this range could reasonably be expected to stem from chance variation alone.

We plotted the ratio of observed-to-expected spontaneous abortions across months and compared with the fitted observed-to-expected ratio from the periodic regression model (Figure 2A) to examine visually how well our data follow an annual sinusoidal pattern. In this analysis, we found a peak in June, July, and August. In analyses adjusted for the month pregnancy attempt began, spontaneous abortion risk peaked in late August, and the adjusted peak/low ratio was 1.3 (95% confidence interval [CI] = 1.1, 1.6; eTable 1; http://links.lww.com/EDE/B906). Next, we fit separate models for early (<8 weeks) and late (≥8 weeks) gestation (Figure 2B). These unadjusted analyses show a peak in July for early losses and little consistent seasonal variation in late losses. After adjustment for the month pregnancy attempt began, risk in spontaneous abortions before 8 weeks’ gestation peaked in mid-August (peak/low ratio = 1.4; 95% CI = 1.2, 1.8 [goodness-of-fit P value = 0.04]), whereas for risk at 8 weeks’ gestation or later, the peak/low ratio was notably closer to the null value (1.2; 95% CI = 1.0, 1.6; goodness-of-fit P value = 0.93).

FIGURE 2.

Unadjusted observed and expected numbers of spontaneous abortion across months of the year, PRESTO. A, The solid black line shows the observed over expected number of spontaneous abortion for each month, and the dotted black line shows the fitted sine curve for these data. B, The solid black line shows the observed over expected number of early spontaneous abortions (<8 weeks’ gestation) for each month, and the dotted black line shows the fitted sine curve for the hazard of early spontaneous abortion. The solid gray line shows the observed over expected number of late spontaneous abortions (≥8 weeks’ gestation) for each month, and the dotted gray line shows the fitted sine curve for the hazard of late spontaneous abortion.

We observed differences in seasonal variation of spontaneous abortion across geographic regions (Figure 3; eTable 1; http://links.lww.com/EDE/B906). Associations were strongest in the Southern and Midwestern United States (peak/low ratios = 1.6 [95% CI = 1.1, 2.3] and 2.2 [95% CI = 1.4, 3.6], respectively), with peaks in late August and early September, respectively. There was no appreciable variation in magnitude of the peak/low ratio across latitude or household income. Seasonal variation in spontaneous abortion was strongest among women with BMI <25 (peak/low ratio = 1.5; 95% CI = 1.2, 1.9) and ≥35 (peak/low ratio = 1.5; 95% CI = 1.0, 2.5) kg/m², although results were imprecise. The observed peak was in early September for BMI <25 kg/m² and late August for BMI ≥35 kg/m². Peak/low ratios for BMI categories of 25 to 29 and 30 to 34 were weaker and could more easily be attributable to chance.

FIGURE 3.

Polar graph showing the size and timing of seasonal peaks in spontaneous abortion risk, across categories of weeks’ gestation, latitude, geographic region, annual household income, and BMI. The size of the peak is determined by the distance from the origin (amplitude). The timing of the peak is indicated by the angle on the graph. The size of the data points is proportional to the inverse variance of the estimate.

When we examined season of conception rather than season of each gestational week at risk, the magnitude of the peak/low ratio was similar, but the timing of the peak shifted several weeks earlier (eTable 2; http://links.lww.com/EDE/B906). For example, the overall peak was in early August (compared with late August in the main analysis), and the peak for early spontaneous abortion (<8 weeks’ gestation) was in late July (compared with early August in the main analysis). In sensitivity analyses restricted to women who conceived in <12 cycles (i.e., women who would not be medically considered infertile), results for early spontaneous abortion were slightly stronger: the peak/low ratio was 1.5 (95% CI = 1.2, 1.9) with a peak in late August. Likewise, when we restricted to women who had been trying to conceive for <3 cycles at study entry, the peak/low ratio for early spontaneous abortion was 1.5 (95% CI = 1.2, 2.0), with a peak in early September.

DISCUSSION

In this geographically heterogeneous North American prospective cohort study of couples trying to conceive, we found some support for the presence of seasonal variation in risk of early spontaneous abortion, with a peak in late August, when risk was estimated to be 44% higher than in late February, 6 months earlier. Risk of all spontaneous abortion, regardless of gestational week at loss, was estimated to be 31% higher in late August compared with late February. The association was stronger among women residing in the Southern or Midwestern United States. This finding, along with our earlier work on seasonal variation in fecundability,¹⁷ indicates that early reproductive events may be susceptible to seasonally varying environmental or lifestyle exposures.

Season should be considered a proxy for environmental exposures and changes in lifestyle that vary across the year. We previously reported that fecundability exhibits a seasonal pattern in this cohort, with the lowest fecundability in late May.¹⁷ We hypothesized that unrecognized early pregnancy loss caused by higher temperatures in the summer could manifest as longer time to pregnancy, particularly in the Southern United States. Results from the present study generally support the hypothesis that exposures that tend to be higher in the summer (for example, heat) may be driving this association. Our observation of stronger seasonal variation among women residing in the Southern and Midwestern United States (where summers tend to be hottest) further suggests that heat may play a role in the etiology of spontaneous abortion. Maternal heat exposure during early pregnancy is associated with birth defects, even in the absence of maternal hyperthermia, and, therefore, exposure to heat during early development could be teratogenic and lead to spontaneous abortion.²⁵ In addition, extreme heat has also been linked to sperm DNA fragmentation,^26,27 which may be related to increased risk of spontaneous abortion,^28,29 although the critical window of exposure for DNA fragmentation is 2 to 3 months before conception. There also may be behavioral responses to the heat (e.g., changes in diet, physical activity) that mediate the relationship with spontaneous abortion. However, other explanations are possible, including that exposures with a seasonal pattern could improve survival of afflicted embryos very early in gestation, allowing them to survive long enough to be detected, which would result in an apparent association between the exposure and spontaneous abortion.

Our results are not entirely consistent with the one earlier preconception cohort study that has examined seasonal variation in spontaneous abortion risk. Among 221 pregnancy planners from North Carolina, enrolled during 1982 to 1985, very early risk of spontaneous abortion (<6 weeks’ gestation) was highest among pregnancies conceived in September,¹² whereas the overall peak in risk in PRESTO was among pregnancies conceived in midsummer, perhaps related to the different geographic distributions between the two studies. However, the outcome assessment in the North Carolina study differed from ours, in that they focused exclusively on pregnancy loss before 6 weeks’ gestation, which constituted 41% of the losses in our study. Our results are consistent with a hospital-based case-control study of Australian women enrolled from 1993 to 1999, where spontaneous abortion risk peaked in March (corresponding to a September peak in the northern hemisphere).⁸ Our results are not consistent with a study of 987 women undergoing fertility treatment in Kentucky, where spontaneous abortion risk followed a bimodal pattern, with peaks in March and December.⁹ However, the etiology of spontaneous abortion may differ between women undergoing fertility treatment and women conceiving spontaneously, which could explain the observed differences.

Our study has several limitations. First, we did not collect daily urine samples for measurement of human chorionic gonadotropin, which has high sensitivity and specificity for detecting pregnancy. Therefore, we certainly missed some early spontaneous abortions. Women test for pregnancy at different times relative to their expected menstrual period and, therefore, recognize their pregnancies (and thus have the ability to detect spontaneous abortion) at different weeks of gestation. However, we observed similar timing of first pregnancy test and pregnancy recognition across months of the year (overall and within subgroups) and, therefore, do not expect that this factor introduced meaningful bias, nor one that would be differential with respect to season. In addition, we did not have information on timing of first positive pregnancy test for all women. Thus, we started follow-up time at 4 weeks’ gestation (median weeks’ gestation at pregnancy recognition among women for whom we had complete data). Because missingness was random (it was based on the date the questions on pregnancy testing were added to our questionnaires), we do not anticipate that this caused any differential misclassification. Second, to identify the date of spontaneous abortion occurrence, women reported when they recognized the loss. Because variation exists in when the spontaneous abortion actually occurs relative to when it is recognized, the dates used in our analysis are likely slightly later than the true dates of loss. Next, we did not have information on whether losses were karyotypically normal, which limits our ability to make inferences about mechanisms of action. Finally, we used periodic regression, which employs assumptions about the shape and frequency of cyclic patterns. To the extent that these assumptions are incorrect, our estimates may be biased.

In this large preconception cohort study of pregnancy planners, we found approximately sinusoidal variation in spontaneous abortion risk across the year, with a peak in the late summer and a trough in the late winter. Results were driven by spontaneous abortion occurring before 8 weeks’ gestation; there was little seasonal variation in later outcomes. Associations were strongest among women residing in the Southern and Midwestern United States. These results indicate that exposures that are more prevalent in the late summer may be related to risk of early spontaneous abortion, prompting greater focus on possible physiologic and behavioral pathways that might mediate this association.

Supplementary Material

Supplemental digital content is available through direct URL citations in the HTML and PDF versions of this article (www.epidem.com).