Exogenous female sex steroid hormones and new-onset asthma in women: a matched case–control study

The West Sweden Asthma Study (WSAS) is a population-based, longitudinal study established in West Sweden in 2008, which has been described in detail elsewhere [24]. At study baseline in 2008, the first questionnaire survey was sent to 30,000 randomly selected adults aged 16–75 years in western Sweden, of which 15,003 were women (out of which 9897 (66%) responded) (Fig. 1). Of the respondents, we excluded 1103 women who reported ever had asthma or ever diagnosed with asthma by a doctor. In 2016, the second questionnaire survey was sent to the remaining 8794 women, out of which 6295 (72%) responded. Of the respondents, 114 developed asthma during the 8-year follow-up. For use of hormonal contraceptives, the study population was based on all responding women, including 114 who had new-onset asthma and 6181 who had never had asthma by 2016. For use of MHT, the study population was restricted to 3641 women of menopausal age (\(\ge\) 45 years) at baseline, including 54 who had new-onset asthma and 3587 who had never had asthma by 2016.

In 2018–2020, the GA²LEN Women’s Questionnaire survey was sent to the 114 cases that had developed asthma during the 8-year follow-up to obtain information on hormonal exposures, out of which 72 (63%) responded. The 72 responding cases were individually matched to 602 controls, out of which 281 (47%) responded to the survey. The matching variables were exact age in years in 2008, place of residence (in or outside Gothenburg), and smoking status (never smoker, former smoker, or current smoker). The choice of a relatively high number of controls per case was to account for potential non-response among the controls. More details on individual matching are provided in Additional file 2 [3, 6, 9, 13, 17, 19, 24‐91].

The exposures of interest included ever use of hormonal contraceptives or MHT, which were reported by participants in the Women’s Questionnaire survey in 2018–2020. Use of hormonal contraceptives was measured by “Have you ever taken a treatment which contains hormones to stop you from getting pregnant (e.g., tablets, depot injection, hormonal coil)?” Use of MHT was measured by “Have you ever taken a treatment which contains hormones to reduce the symptoms or effects of menopause (e.g., ‘HRT’ tablets, depot injection, patches, gels, but not vaginal creams or pessaries)?” If participants answered “yes” to either question, they were further asked “How old were you when you first took the treatment?” The respective comparator group was never use of hormonal contraceptives or MHT.

Women who reported never having had asthma or doctor-diagnosed asthma during the first questionnaire survey in 2008 but later reported that they had asthma or doctor-diagnosed asthma during the second questionnaire survey in 2016 were considered as having developed new-onset asthma. Women who reported asthma were further asked “How old were you when you got asthma?”

We used causal directed acyclic graphs (DAGs) to represent potential systematic biases in our study [30, 92]. Details on DAGs are provided in Additional file 2. For confounding bias, we identified potential common causes of each hormonal exposure and new-onset asthma based on previous literature (Additional file 2: Table S1); then, we applied the backdoor criterion to determine a sufficient set of adjustment variables required to minimize confounding (Additional file 2: Figs. S1 and S2) [26]. For use of hormonal contraceptives, the adjusted variables included age, place of residence, level of education, age at menarche, gynecological conditions (including endometriosis, polycystic ovarian syndrome, gynecological acne, and hysterectomy with or without oophorectomy), and tobacco smoking. For use of MHT, the adjusted variables included age, place of residence, level of education, body mass index, tobacco smoking, environmental tobacco smoke, age at menopause, physical exercise, and gynecological conditions. The definitions of the adjusted variables are summarized in Additional file 3 [93]. The DAGs for selection bias and measurement bias are respectively presented in Additional file 2: Figs. S3 and S4. Particularly, for many women, hormonal exposures happened before the study had started; if hormonal exposures had a causal effect on new-onset asthma, restricting the study population to those who had never had asthma at baseline would likely result in differential proportion of susceptible individuals after baseline, thereby introducing selection bias [27].

We developed a priori an analysis protocol including justifications for the statistical methods applied in this study (Additional file 2). In brief, for each hormonal exposure we applied Frequentist conditional logistic regression model to adjust for the matching sets and measured confounding variables [38, 49, 51]. We used multiple imputation (MI)—full-conditional specification (FCS) [44, 46]—to impute the missing data and fitted Frequentist conditional logistic regression model to the multiply imputed datasets (\(m=\) 100). Then, we conducted complete-case analysis as a sensitivity analysis, that is, restricting the analysis to individuals with complete data on all variables included in the model. The results were summarized as odds ratio (OR) with 95% confidence interval (CI). We conducted subgroup analyses by baseline age (above or below an age cut-off) to evaluate potential selection bias [94] and calculated E-value [69] to assess the robustness of the estimated causal effects to potential residual confounding.

We applied Bayesian conditional logistic regression model in the multiply imputed datasets (Additional file 2) [64, 66]. The Bayesian model approximated the posterior probability distributions over all possible values of the parameters of interest, conditional on the prior probability distributions, statistical model and observed data [58]. For parameters of hormonal exposure variables, we derived a priori original prior distributions (use of hormonal contraceptives: \(\mathrm{log}OR\sim N(-0.26,{0.20}^{2})\); use of MHT: \(\mathrm{log}OR\sim N(0.17,{0.13}^{2})\)), based on our previous review work [3, 9, 13] as well as newly published studies [11, 14, 15]. In addition, we set three alternative prior distributions to represent our uncertainty about the original prior distribution and used a flat prior distribution (which assigns equal prior plausibility to all possible values of a parameter) to understand the influence of our prior knowledge compared to that of the observed data on the model results [63]. The process of deriving the prior distributions is available in Additional file 4 [3, 10‐15, 69]. Finally, we used the Markov chain Monte Carlo (MCMC) method to approximate the posterior distributions [64, 66] and calculated the median and the 95% central posterior interval (PI) on OR scale [68]. The 95% central PI means that, given the prior distributions, statistical model, and observed data, the true causal effect has a 95% probability of falling within this range [58]. We estimated the probability that each hormonal exposure would increase the risk of new-onset asthma in women [58].

As asthma is relatively rare in our study population (< 15% by the end of follow-up), the estimated ORs can be approximately interpreted as risk ratios [69]. All statistical analyses were performed using the R software (version 4.0.4) [95]. The R scripts and packages [45, 64, 69, 96‐106] are available in Additional files 5 and 6.

In total, 72 cases and 281 controls responded to the Women’s Questionnaire survey, among which 62 cases (86.1%) and 204 controls (72.6%) had ever used hormonal contraceptives. The median age when starting hormonal contraceptives was 18 years (range: 13–49 years). Thirty-five cases and 150 controls were aged 45 years or older at baseline, among which eight cases (22.9%) and 25 controls (16.7%) had ever used MHT. The median age of starting MHT was 52 years (range: 38–64 years). Forty-two of the 114 cases did not respond to the Women’s Questionnaire survey and were matched with additional 115 controls, resulting in a total of 114 cases and 717 controls. Tables 1 and 2 summarize the background characteristics for all cases and matched controls and for hormone users and never users, respectively. The median age of participants at baseline was 44 years (range: 19–74 years). Cases were more likely to have a BMI of \(\ge\) 30 kg/m² and a higher level of education (although not statistically significant). Hormonal contraceptive users were more likely to be younger and have a lower BMI and a higher level of education, compared to never user. MHT users were more likely to have a lower BMI, be a former smoker, and live in Gothenburg city, compared to never user. The PECOS components, the results from complete-case analysis, and the MI process are detailed in Additional file 3. Below, we describe the results from both Frequentist and Bayesian analyses based on the multiply imputed datasets.

In Frequentist analysis, ever use of hormonal contraceptives compared to never use was associated with an increased risk of new-onset asthma in women. Subgroup analyses showed that the association consistently became stronger among women of older age at baseline. Among menopausal women, ever use of MHT compared to never use was statistically non-significantly associated with an increased risk of new-onset asthma. In Bayesian analysis, although the 95% PIs for both use of hormonal contraceptives and MHT included the null value, there was a 72% and 91% probability that the OR was larger than one for hormonal contraceptives and MHT, respectively.

There are several strengths in our study. First, being a case–control study nested within an ongoing prospective cohort, we could ascertain the temporality between the exposure and the outcome and thereby apply models to allow estimation of potential causal effects of exogenous female sex hormones on asthma risk. Second, we built causal DAGs based on published literature and our a priori subject-matter knowledge to identify potential confounding variables for each hormonal exposure [26, 30], which provides an explicit framework to minimize confounding. Third, in order to reduce the potential bias introduced by item non-response, we implemented MI to impute the missing data [39]. Fourth, we applied Bayesian statistical model which incorporated our prior background knowledge into the analysis; in this way, the current analysis provides a solid basis for future analyses. Finally, we adopted open and reproducible research practices [107]—we developed a priori statistical analysis protocol, documented in detail the research process, and made R scripts publicly available.

Certain limitations need to be taken into account in the interpretation of our findings. First, in our study, women with ever asthma at baseline were excluded, and for many women, hormonal exposures occurred before the study had started. This could have introduced selection bias, especially among older women, only if hormonal exposures had a causal effect on new-onset asthma (Additional file 2: Fig. S3) [27]. For example, if hormonal contraceptives increased the risk of new-onset asthma, the more susceptible individuals would have developed asthma before baseline in the exposed group than in the unexposed group; consequently, restricting to individuals who had not developed asthma by baseline would likely result in less susceptible individuals after baseline in the exposed group than in the unexposed group, thereby attenuating the magnitude of the effect estimate for the true harmful effect or even biasing the effect estimate towards the opposite direction [27]. Contrary to the hypothetical example, in our study, we found that when the age at baseline increased, the magnitude of point estimate for use of hormonal contraceptives consistently increased. We suspect that selection bias due to selection of women based on baseline asthma status may likely be the main explanation for the increase of the point estimate with increasing baseline age. This suggests that use of hormonal contraceptives may in fact have a protective effect on new-onset asthma, as opposed to a harmful effect indicated by our results. Notably, this type of selection bias may arise in any study that attempts to estimate the causal effect of an (lifetime) exposure that occurs before the study has started [27, 94, 108, 109]. Second, because some identified confounding variables were not available in our dataset, we had to rely on proxy variables (e.g., for socioeconomic status), or could not adjust for them at all (e.g., diet, alcohol) for use of MHT and new-onset asthma. Third, the information on hormonal exposures and asthma status was obtained retrospectively by questionnaire survey. Thus, an individual’s ability to recall their medical history may affect the measurement of both hormonal exposures and asthma (Additional file 2: Fig. S4) [92]. In addition, because hormonal exposures were ascertained by recall after asthma had occurred, asthma status might affect the recall of hormonal exposures. However, we expect that the influence was likely to be minimal (if present), because the causal relationship between asthma and hormonal exposures had not been well established. Fourth, more than half of the cases did not report age at asthma onset, which could potentially affect estimation of the temporal relationship between hormonal exposures and new-onset asthma. However, for the cases who reported age at asthma onset, asthma occurred after use of hormonal contraceptives or MHT. Fifth, although WSAS is a population-based study of a representative sample in western Sweden, given the baseline participation rate of 66% and the follow-up rate of 72%, potential selection bias (i.e., bias due to unit non-response) may have arisen, which could potentially affect the estimation of causal effects of exogenous female sex hormones on asthma risk as well as the generalization of our results to the source population. Unfortunately, we were unable to account for this type of bias in our study. Sixth, we could not investigate the potential varying causal effects of use of hormonal contraceptives and MHT by subtypes, doses, durations of use, and routes of administration, because the information on these factors could not be reliably determined from the questionnaire survey. Seventh, the study population for MHT use included women aged \(\ge\) 45 years at baseline, which was used as a proxy measure to identify menopausal women. Since it was not based on the actual menopausal status, this may affect the generalization of our results to the menopausal group of women. Finally, due to data unavailability, we were unable to address the potential modifying effect of BMI for the effects of exogenous female sex hormones on asthma risk.

To our knowledge [3], three cohort studies have investigated the causal effects of use of hormonal contraceptives on the risk of new-onset asthma, which excluded women with ever asthma at baseline to form the study population [10‐12]. Interestingly, a similar upward trend existed in the effect estimate with increasing baseline age across or within studies (Additional file 7: Fig. S1 [10‐12]). Likewise, we suspect that selection bias may likely explain the increase and the reversal of the relative risk with increasing baseline age.

Recently, an umbrella review including five cohort studies [13] and a nested case–control study based on the Danish registers [14] reported that use of MHT was associated with an increased risk of new-onset asthma in menopausal women. In contrast, a UK national cohort study [15], the largest longitudinal study on the topic to date, found that use of MHT was associated with a decreased risk of new-onset asthma and that longer duration of use was associated with a lower asthma risk than shorter duration. It is unclear whether different subtypes of MHT, population characteristics, or asthma phenotypes could explain the heterogeneity in these results.

First, despite intensive investigations, the epidemiologic evidence on female sex hormones and asthma risk remains equivocal [3, 9, 16]. Future longitudinal studies that account for systematic biases will help advance the evidence. This will benefit from making explicit the causal goals of research [17], and using causal diagrams (e.g., DAGs) to represent and classify systematic biases [25], which formulates a clearer framework for evaluating the proposed causal structures and analytical approaches, thus facilitating explicit causal reasoning of the results. Second, in studies of hormonal exposures that occur early in life (e.g., use of hormonal contraceptives), special attention needs to be paid to potential selection bias [27]. Indeed, for the upward trend in the effect estimate with increasing baseline age observed in our study and across previous studies, other explanations may include effect modification by age (i.e., use of hormonal contraceptives beneficial at younger ages but harmful at older ages) or residual confounding (i.e., the strength of residual confounding varied systematically across different age groups). Future studies that account for different systematic biases will hopefully provide further insights. Third, longitudinal studies that investigate the potential varying causal effects of use of hormonal contraceptives and MHT by subtypes, doses, durations of use, and routes of administration on various phenotypes of asthma are warranted. This can help generate evidence for more individualized asthma prevention strategies for women. Fourth, future studies are needed to investigate the potential modifying effect of BMI for the effects of exogenous female sex hormones on asthma risk. Notably, because BMI can arguably be a mediator along the causal pathway between hormonal contraceptives and new-onset asthma [80‐82], future studies that carefully collect data on relevant key variables are warranted (e.g., physical activity, alcohol, stress, diet) (Additional file 7: Fig. S2).

The Bayesian framework naturally incorporates the investigators’ prior background knowledge about the parameters of interest before observing the data (i.e., prior beliefs) and updates these beliefs about the parameters after observing the data (i.e., posterior beliefs) [19]. This can help improve precision in effect estimate when the sample data is relatively small. For example, the incorporation of our prior knowledge from previous studies of use of MHT and asthma produced an effect estimate with substantially more certainty than that from only relying on the sample data. On the other hand, when the quality of the sample data is not optimal, the Bayesian method can help to some degree mitigate the influence of potential systematic biases on the results; in contrast, the Frequentist method relies only on the sample data and thus is highly dependent on data quality [19]. For example, the inclusion of our prior knowledge on use of hormonal contraceptives and asthma yielded a conservative result as opposed to that from Frequentist analysis; the wide spread of the posterior distribution means that more evidence is needed, which favorably kept us from making an overconfident claim that use of hormonal contraceptives would increase asthma risk in women. Furthermore, Bayesian analysis allows us to make intuitive probabilistic statements about the parameters [19, 110]. We can say that, for example, there was a 91% probability that ever use of MHT would increase asthma risk in menopausal women, conditional on the priors, statistical model and observed data. However, Frequentist estimates are often misinterpreted as Bayesian estimates [19]. A criticism of Bayesian analysis is that the priors are subjective [19]. However, it is noteworthy that this subjectivity creeps into both Frequentist and Bayesian analyses, where in Frequentist analysis, flat priors are implicitly assumed and may not always realistically capture a priori knowledge [19, 20]. The Bayesian framework makes explicit this aspect of subjectivity and uses different distributions to quantify this subjectivity [20].