The longitudinal relationship between prenatal maternal cortisol and preschool internalising problems

Intro: Prenatal exposure to synthetic glucocorticoids may increase the risk of emotional symptoms in childhood partly by reducing fetal growth. We explored if physiological levels of prenatal maternal cortisol were associated with internalising problems in boys and girls and if this was mediated by birth weight. Methods: Mother-child dyads from the prospective Odense Child Cohort (n = 1162) were included if maternal serum cortisol (3rd trimester), offspring birth weight, and Child Behaviour Checklist (CBCL) assessments in preschool age were available. Crude and adjusted associations between cortisol and internalising problems were determined in linear mixed models stratified by offspring sex. Covariates included parental psychiatric history, parity, maternal age, education, smoking during pregnancy, and gestational age at birth. In the presence of significant associations, we evaluated the potential mediating role of birth weight. Results: The study sample included 601 boys and 561 girls and internalising problems were assessed at mean ages 2.3 ( ± 0.4) and 5 ( ± 0.5) years. In the crude analysis, cortisol was positively associated with internalising problems in boys (p-value 0.017) and in girls (p-value < 0.0001). In the adjusted analyses, there was no statistically significant association between cortisol and offspring internalising problems in boys or girls (all p-values > 0.15). There was no mediation by birth weight. Discussion: Maternal serum cortisol was positively associated with offspring internalising problems in boys and girls, but there was no association following adjustment for potential confounders and no mediation through birth weight. Maternal third-trimester cortisol levels do not predict preschool offspring internalising problems in our study.


Introduction
Anxiety and depression are leading causes of disability worldwide (Ferrari, 2022) and the incidence rates of clinical diagnoses and population reports of anxiety and depressive (internalising) symptoms are rising for children and adolescents (Durbeej et al., 2019;Piao et al., 2022).An adverse intrauterine environment may affect fetal development and increase the risk of internalising disorders (Mathewson et al., 2017;Tien et al., 2020).This is supported by recent studies showing that exposure to synthetic glucocorticoids (sGC) -a pharmacological equivalent to high levels of maternal stress -holds potential neurodevelopmental effects (Räikkönen et al., 2020;Wolford et al., 2020).In pregnancy, sGC are primarily administered during the third trimester to improve fetal outcomes at the risk for preterm births (Liggins and Howie, 1972); however, a population-wide Danish sibling-control study suggests that sGC administration may increase the risk of offspring internalising disorders (Laugesen et al., 2022), in part by reducing fetal growth.These in-vivo quasi-experiments highlight the pressing need for translational studies of the maternal endogenous stress hormone, cortisol.
Maternal cortisol increases physiologically throughout pregnancy reaching a peak in the third trimester (Jung et al., 2011).Unlike sGCs, cortisol is inactivated by the placenta through most of pregnancy, thereby protecting the fetus from untimely exposure (Seckl and Holmes, 2007;Gitau et al., 1998).However, in the third trimester reduced cortisol inactivation facilitates prepartum organ maturation (Fowden et al., 1998) increasing the viability of the offspring.Despite this vital benefit of cortisol, higher maternal levels during pregnancy have been associated with potentially adverse effects, in particular reduced fetal growth (VALLADARES et al., 2009;Goedhart et al., 2010;Cherak et al., 2018) and altered neurodevelopment (Buss et al., 2012a;Graham et al., 2019).Community-and clinic-based longitudinal studies have shown a relation between increased prenatal maternal cortisol and anxiety symptoms in 6-9-year-olds (N=170) (Davis and Sandman, 2012) and internalising problems at age nine (N=70) (Isaksson et al., 2015).A meta-analysis highlighted inconsistencies in the reported results (Zijlmans et al., 2015) with some proposing that the associations between cortisol and internalising problems depend on offspring sex (Sandman et al., 2013;Sutherland and Brunwasser, 2018).
A female fetus is associated with higher maternal prenatal cortisol levels (Andersen et al., 2019;Vrijkotte et al., 2023;Bleker et al., 2017) and girls also have higher diagnostic rates of internalising disorders compared to boys (Dalsgaard et al., 2020).A study of children (N=65) from the US reported that the association between high prenatal maternal cortisol and depressive symptoms at age 7, was limited to females (Buss et al., 2012a).Another US study was able to replicate this female-driven association between maternal cortisol and internalising problems at 2 years of age (N=70) (Graham et al., 2019), although others could not (Sutherland and Brunwasser, 2018;Wright et al., 2019).Sex differences in fetal neurodevelopment of the amygdala has been proposed as a mediator of this sex-dimorphic association (Buss et al., 2012a;Graham et al., 2019;Stoye et al., 2020), but to our knowledge, the role of fetal growth remains unexamined.Females have lower average birth weight than boys and population-based studies suggest sex-specific associations between birth weight and childhood behaviour (Dooley et al., 2023).
In summary, the literature suggests a female-driven association between high prenatal maternal cortisol levels and more offspring internalising symptoms, mediated by fetal development (Sandman et al., 2013;Sutherland and Brunwasser, 2018); however, systematic reviews conclude that existing studies have small sample sizes, heterogeneity in timing and analysis of cortisol, and limited attention to confounders (Zijlmans et al., 2015;Sutherland and Brunwasser, 2018).Furthermore, the potential mediating role of reduced fetal growth in this relationship has not been explored.

Aim
To explore the association between prenatal maternal plasma cortisol and internalising problems in sex-stratified preschool offspring and to investigate whether birth weight mediates a potential association using crude and multivariable mediation analyses.

Participants
This study was part of the Odense Child Cohort (OCC) study (Kyhl et al., 2015).The OCC recruited pregnant women within the Municipality of Odense, Denmark, between January 1, 2010 and December 31, 2012.Of 6707 pregnancies, 4017 women received information about OCC, and 2874 pregnant women were included in the cohort.They were invited to complete several questionnaires and clinical examinations during pregnancy, and at child ages 2-3 and 5 regarding their child's behaviour.OCC participants eligible for this study had a singleton pregnancy, a third-trimester cortisol measurement and at least one completed childhood behavioural assessment.To keep observations independent, we only included the first sibling, in the case where a mother had participated in the OCC with more than one pregnancy.

Cortisol sampling and analysis
Mothers in the OCC were invited to provide standardised thirdtrimester fasting blood samples (Andersen et al., 2019).Maternal venous samples were collected in the morning between 7:40-10:00 am in GA week 27-30 (mean 29.0 (± 0.6)) with the majority of the blood samples (82.3 %) being collected between 8.00 am and 8.59 am.Samples were centrifuged for 10 minutes at 3000 revolutions per minute, and stored at − 20 • C. Samples were analysed for cortisol using liquid chromatography-tandem mass spectrometry (LCMS/MS).Plasma samples were extracted by supported liquid extraction, and the eluate was evaporated and reconstituted before analysis.The LCMS/MS analysis was calibrated by in-house prepared calibrators, and the relative SD was <10 %.Cortisol analysis quality was assured by monthly participation in the external quality control program for steroid hormones from The United Kingdom National External Quality Assessment Service.A subgroup of the women with a serum-cortisol measure also provided a 24 hour urine collection from GW 27-28.The time for voiding before attendance at the department was noted.Maternal urine samples were collected from that time until the morning of the second day (24 hours).On day 2, the morning urine was collected separately and included in the sampling period.Urine samples were kept frozen at − 80 • C until analysis until analysis by LCMS/MS.

Childhood internalising problems
Offspring internalising problems were assessed using the maternal reported Child Behaviour Checklist for ages 1.5-5 years (CBCL/1½− 5) at ages 2-3 and 5 years.CBCL/1½− 5 consists of 99 items rated on a 3point Likert scale (0=absent, 1=occurs sometimes, 2=occurs often).The questionnaire comprises a 36-item subscale of internalising (emotional) problems (range 0-72) covering emotional reactivity, anxiety/depression and somatic complaints, and a 24-item subscale of externalizing (behavioral) problems (range 0-48) covering attention problems and aggressive behaviour.The primary outcome was internalising problems as a continuous score.Externalising problems were analysed as a secondary outcome to examine whether a potential association with maternal cortisol would be specific to internalising problems.

Perinatal variables
Birth weight was recorded by midwives shortly after delivery and extracted from birth records, and kept as a continuous variable in kilograms.Gestational age (GA) was ultrasound-determined, and GA at birth was kept as a continuous variable in days.Preterm birth was defined as birth <37 weeks +0 days (used in sensitivity analysis).

Covariates
Maternal age (years), maternal pre-pregnancy BMI, maternal smoking during pregnancy (no, yes), education status (low: high-school or less, middle: high-school +1-4 years, high: high-school +>4 years), and parity (multipara vs nullipara) were obtained from questionnaires and hospital records.Information on maternal and paternal psychopathology before childbirth was obtained from the Danish Psychiatric Central Research Register (PCR) and dichotomised as psychopathology for any parent (yes/no) (Mors et al., 2011).The PCR contains information on all psychiatric disorder diagnoses assigned at inpatient or outpatient mental health services from 1995 onwards.Finally, offspring age (years) at each CBCL/1½-5 assessment was recorded.

Statistics
We used linear mixed models to allow for repeated CBCL/1½-5 assessment of children at approximately age 2-3 and 5 years and thereby minimise the loss to follow-up.Potential confounders (Zijlmans et al., 2015) were identified a priori based on directed acyclic graphs (DAG) shown in Fig. 1a, and participants with missing data were excluded from multivariable analysis.Selection bias was assessed by comparing the analytic sample to excluded participants.
For our primary analysis, we divided the sample by offspring sex recorded at birth.Age at CBCL/1½-5 assessment was included in the crude and multivariable analysis because CBCL/1½-5 responses vary by age, which has been taken into account in the original validation of the questionnaire (Rescorla, 2005).Given the dimensional nature of child behaviour and emotional problems we used raw scores and not clinical cut-offs in our analysis (Kotov et al., 2017;Caspi and Moffitt, 2018).We treated the ordinal CBCL/1½-5 data as a continuous measure, which due its many levels should not affect inference in most situations (Robitzsch, 2020).In both crude and adjusted analyses, we used a modified alpha of 0.2 to test if associations could be subject to mediation before proceeding with mediation analysis.Prior to mediation analysis, we tested for interaction between cortisol and birth weight on internalising problems as part of a quality control and proceeded only if the interaction was insignificant (VanderWeele, 2014).
Model assessment indicated some deviations from normality in residuals and random effects.We therefore used a case bootstrap with 5000 resamples to estimate the standard errors used in the statistical inference.To estimate confidence intervals for direct-and mediated effects of cortisol, we used quasi-Bayesian confidence intervals (Tingley et al., 2014).The variables of primary interest in our mediation model were the total effect of cortisol, and the indirect effect of cortisol mediated by birth weight (in the original units of cortisol).As this was an exploratory analysis, we determined statistical significance by examining the 95 % confidence intervals (CI).Prespecified sensitivity analyses included analysis of term-born only, nulliparous pregnancies only and both sexes combined.We also wanted to examine the specificity of the association between maternal cortisol and internalising problems; therefore, we reanalysed our crude and adjusted models using externalising problems as the outcome.Post hoc, we re-analysed our models using 24-hour urine cortisol as the exposure variable to test if the effect size of maternal cortisol would increase.This was done because 24-hour urine cortisol reflects unbound serum cortisol (El-Farhan et al., 2017) which may more readily diffuse into the fetal circulation and affect fetal development.In order to achieve statistical power to illustrate the effects of confounding on the mediating role of birth weight, we pooled the data after testing for a sex interaction.We then performed mediation analysis on the adjusted mixed model in the pooled sample and explored a parsimonious model adjusting for parity and maternal age.

Results
A total of 1991 (69.3 % of the original cohort) mother-child dyads had at least one CBCL/1½-5 assessment, and for 1215 (42.3 %) participants, maternal cortisol measurement in pregnancy was also available.Next, we excluded twins and siblings, and the study sample comprised 1162 (40.4 %) children (601 males and 561 females) for crude analyses (Fig. 2 + Table 1).Cortisol concentrations ranged from 112.5 -1737.3nmol/L and information on additional variables for the study population are described in Table 1 with variable correlations illustrated in Fig. 1b.Multivariable modelling was performed for 596 males and 552 females with available information on all relevant confounders.Compared to the excluded participants, the study sample was characterized by higher education and less smoking during pregnancy (Supplementary Table 1).
In females, the crude analysis showed a statistically significant total effect of cortisol, with 1 nmol/l increase in cortisol resulting in a 0.00214 (CI: [0.00106, 0.00321], P-value: < 0.0001) points increase on the CBCL internalising problem subscale.In the adjusted analysis, the total effect of cortisol was not statistically significant (β: 0.00085, CI: [-0.00032, 0.002], P-value: 0.15).In males, the crude analysis showed a statistically significant total effect of cortisol with 1 nmol/l increase resulting in an increase of 0.00157 (CI: [0.00032, 0.0029], P-value: 0.017) points on the CBCL internalising problem subscale.In the adjusted analysis the total effect of cortisol was not statistically significant (β: 0.00042, CI: [-0.001, 0.002], P-value: 0.58) (Table 2).We then pooled both sexes together (n = 1162) and included a sex-by-cortisol interaction.We found no evidence for an interaction in crude or adjusted models.In the combined sample, the crude analysis showed a statistically significant total effect of cortisol, with 1 nmol/l increase in cortisol resulting in a 0.00184 (CI: [0.00106, 0.00321], P-value: < 0.0001) points increase on the CBCL internalising problem subscale.In the adjusted analysis, the total effect of cortisol diminished and was not statistically significant (β: 0.00085, CI: [-0.00032, 0.002], P-value: 0.15) (Supplementary Figure 1 and Supplementary Table 2).
In crude mediation analysis stratified by fetal sex, we found no indirect effects mediated by offspring birth weight in males or females (Table 2).In crude mediation analysis of the combined cohort, there was a statistically significant indirect effect of cortisol mediated by birth weight (p-value: 0.006) (Supplementary Table 2).Despite no overall effect of cortisol, we performed mediation analysis to illustrate the impact of confounding on the mediating role of birth weight.As expected, there were no significant direct effects of cortisol or indirect effects mediated by birth weight in adjusted analyses (Supplementary Figure 2).Similar results were seen when adjusting only for maternal age and parity (supplementary table 2) In the planned sensitivity analyses, we reanalysed our adjusted model with prespecified subgroups.We first analysed term-born children only (n males = 578, n females = 544); second, we analysed offspring from nulliparous mothers only (male n = 335, female n = 314).There were only modest changes in effect sizes (Supplementary Table 3-4).As a prespecified supplementary analysis, we changed the dependent variable in our models to parent reported externalising problems measured by the CBCL/1½-5 externalising problems subscale (male n = 600, female n = 561).Unlike the crude models for internalising problems, there were no statistically significant associations between cortisol and offspring externalising problems for males or females.This did not change in the adjusted model (Supplementary Table 5).Post hoc, we analysed levels of 24-hour urine cortisol collected in a subgroup of individuals (male n = 166, female n = 133).Using this exposure variable did not change the null-finding in the adjusted analysis (Supplementary Table 6), with similar results in a combined analysis.

Discussion
Using repeated behaviour and emotional assessments in a large cohort, our study demonstrated crude relationships between maternal serum cortisol levels in the third trimester and internalising problems in pre-school children, that disappeared after confounder adjustment.
This study adds to the literature in several ways.Our data and analyses confirmed crude associations between prenatal cortisol and offspring internalising problems reported in previous studies (Graham et al., 2019;Buss et al., 2012b).However, it questions the causal pathway from cortisol exposure through altered fetal development to childhood internalising problems (Seckl and Holmes, 2007;Buss et al., 2012a;Graham et al., 2019) (Supplementary Figure 2).Our large and data rich birth cohort sample allowed us to conduct systematic and thorough confounder adjustments and after these, there was no longer significant total effects of cortisol.This is in line with concerns from a recent review that highlighted insufficient confounder-adjustment as a likely source of error in interpretating associations between maternal prenatal cortisol and child outcomes (Zijlmans et al., 2015).Second, our results question the proposed female vulnerability to prenatal cortisol exposure (Sandman et al., 2013;Sutherland and Brunwasser, 2018).We observed a numerically larger association among females compared to males but this was not statistically significant in the adjusted analysis, and effect sizes were small (1 SD increase in cortisol corresponding to a 0.19 point increase in CBCL/1½-5 internalising score (range 0-72 points).This suggests that, if an effect in females exists, it has negligible clinical significance.Third, this study is unique in using serum cortisol as opposed to other studies that were based on maternal saliva cortisol (Buss et al., 2012a;Graham et al., 2019;Davis andSandman, 2012, 2010;Isaksson et al., 2015;Wright et al., 2019;Braithwaite et al., 2017).Auto sampling of saliva may be subject to variation in timing; additionally, the placenta and salivary glands have similar enzymatic inactivation of cortisol.Therefore, high saliva cortisol could reflect impaired hormone metabolism and not just high serum cortisol (Blair et al., 2017).
Our findings contrast the paper by Isaksson et al (Isaksson et al., 2015)., who found increased CBCL externalising and internalising problems at age nine in Nicaraguan children (N=111).In addition to examining a more socially deprived sample at an older age and using saliva cortisol, the authors did not adjust their analyses for the two important confounders, maternal age and parity.Our findings also contrast studies on the long-term effects of sGCs (Räikkönen et al., 2020;Laugesen et al., 2022).The difference may be explained by the higher potency of sGC compared to physiological cortisol and its superior ability to traverse the placenta (Gitau et al., 1998).It is also possible that altered signalling of sGC could affect the fetus in a way qualitatively different from that of cortisol (Timmermans et al., 2019).
The conducted sensitivity analysis of term-born pregnancies only, resulted in negligible changes in effect sizes.Parity could modify the effects of confounders on cortisol concentration (Bleker et al., 2017), but the sensitivity analysis of nulliparous pregnancies did not notably change effect sizes.As expected, the supplementary analysis of externalising problems did not result in significant crude or adjusted effects of cortisol.This suggests that cortisol associates more closely to internalising than externalising problems but that differences are due to underlying confounders.Our results based on serum cortisol could have been biased because protein-binding of cortisol inhibits signalling activity, but analysis of maternal 24-hour urine cortisol (reflecting unbound cortisol) supported our primary results.In the analysis of both sexes combined, we did not find a statistically significant sex interaction in crude or adjusted models.This suggests that the association between maternal cortisol and internalising problems does not differ between males and females and that cohort stratification is unnecessary.The crude analysis for the combined sample agreed with the a priori hypothesis of a mediating role of birth weight for the effect of cortisol on internalising problems; however, consistent with the primary analysis, all associations disappeared in the adjusted analyses (Supplementary Table 2 and supplementary Figure 1 + 2), again highlighting the importance of confounder adjustment.
The present study has several strengths.The sample size was considerable compared to previous studies investigating emotional and behavioural outcomes of prenatal cortisol exposure (Zijlmans et al., 2015;Sutherland and Brunwasser, 2018), and to our knowledge, the largest study to date.We used a validated questionnaire with repeated measures in more than half of the analytic cohort to assess internalising problems.Healthcare professionals recorded birth weight immediately after birth and potential confounders were collected at baseline with a limited risk of recall bias.The cortisol measurements were subject to stringent methodology in collection and analysis.The fasting morning cortisol samples collected in a narrow gestational period and analysis by mass spectrometry should reduce bias caused by diurnal variation in cortisol and the progressive rise during pregnancy (Bleker et al., 2017;Wilhelm et al., 2007) as well as the systematic cortisol underestimation seen with immunoassays (Jung et al., 2011).Finally, our third trimester samples are particularly well-suited to address our research question given reports of changing associations between maternal cortisol and internalising problems across trimesters (Buss et al., 2012a;Davis and Sandman, 2010) and for comparability to timing of sGC administration.
The study also has limitations.Participants in birth cohort studies tend to be of better health and have higher socioeconomic status than the general population, which is also the case in OCC (Kyhl et al., 2015).Our analytic sample had more favourable baseline characteristics than the excluded OCC participants.This may have influenced our results because passage of cortisol to the fetal circulation may relate to maternal psychosocial stressors (Seth et al., 2015).Our inclusion of covariates influencing fecundity and internalising problems should minimise collider-bias from conditioning on live births but the possibility of residual bias remains from an unknown confounder (Liew et al., 2015).The time from awakening to venous sampling was unknown, potentially increasing the variability in cortisol levels; however, we deem it unlikely that the time from awakening would be associated with behavioural assessment years later to meaningfully bias the effect estimates.Finally, the OCC study participants are ethnically homogenous and our results may not generalise to other populations.
In summary, our study indicates that maternal cortisol in the third trimester is not an important determinant of preschool internalising problems in privileged cohorts, and we encourage careful confounder adjustment in future studies to prevent spurious findings.

Fig. 1 .
Fig. 1.Conceptual framework of the analysis and correlation matrix (a) Directed acyclic graph for the primary adjusted mediation analysis in males and females separately.Green nodes represent exposure and mediator, pink nodes represent confounders, blue nodes represent outcome, or ancestors of outcome.(b) Correlation of variables in primary-and supplementary analysis.Cortisol = Fasting morning cortisol in maternal serum in week 28 of pregnancy, Birth weight = midwifemeasured birth weight, Internalising problems = Child Behaviour Checklist (CBCL) internalising problems, Parity = nulliparous vs multiparous mother, Maternal age = maternal age at study inclusion, Gestational age = gestational age at birth in days, Maternal BMI = pre-pregnancy BMI, Maternal education = short, medium or long education, Parental psychiatric history = zero vs at least one parent with a diagnosis of mental illness, Maternal smoking = yes/no, Age at assessment = offspring age at CBCL assessment, Urine cortisol = 24-hour urine cortisol.

Fig. 2 .
Fig. 2. Participant selection.Flowchart of the selection of the analytic cohort.CBCL = Child Behaviour Checklist, OCC = Odense Child Cohort.Made with diagrams.net.
) † Any psychiatric diagnosis prior to pregnancy in father or mother, but recorded after 1995 † † Low = Primary school and short youth education, Medium = Short tertiary education, High = University diploma

Table 1
Baseline demographics for the OCC analytic sample (n = 1162).