Associations of Maternal Androgen-Related Conditions With Risk of Autism Spectrum Disorder in Progeny and Mediation by Cardiovascular, Metabolic, and Fertility Factors

Abstract

Fetal exposure to elevated androgens is thought to contribute to autism spectrum disorder (ASD) risk. However, data rely heavily on in utero androgens measurements, which also reflect fetal secretions. Thus, in utero hyperandrogenemia might indicate adverse autism-related neurogenesis that has already occurred affecting fetal androgen homeostasis, rather than being a cause of the disorder. Associations between maternal androgen-related conditions and ASD could more directly implicate androgens’ etiological role. We examined the association between maternal hyperandrogenemia-related conditions, focusing primarily on polycystic ovarian syndrome (PCOS), and progeny ASD, in an Israeli cohort of 437,222 children born in 1999–2013. Odds ratios and 95% confidence intervals were estimated using generalized estimating equations. Multiple mediation analyses using natural effect models were conducted to evaluate combined mediation of the PCOS effect by androgen-related cardiovascular, metabolic, and fertility factors. Results indicated that children of mothers with PCOS had higher ASD odds compared with children of mothers without PCOS (odds ratio = 1.42, 95% confidence interval: 1.24,1.64), and this effect was only partly mediated by the factors considered. Elevated odds were also observed for other hyperandrogenemia-related conditions. Findings provide support for direct involvement of maternal hyperandrogenemia in ASD etiology. Alternatively, findings might reflect shared genetic and/or environmental factors independently affecting maternal androgen homeostasis and fetal neurodevelopment.

Abbreviations

ASD

autism spectrum disorder

CI

confidence interval

ICD-9

International Classification of Diseases, Ninth Revision

GEE

generalized estimating equations

MHS

Maccabi Health Services

OR

odds ratio

PCOS

polycystic ovarian syndrome

Autism spectrum disorder (ASD) is a neurodevelopmental communication disorder with increasing prevalence reported worldwide. While the etiology of the disorder remains unclear, ASD exhibits a strong male preponderance, suggesting that sex-specific factors might play important etiological roles. The prenatal sex steroid theory of autism specifically posits that fetal exposures to elevated levels of steroid sex hormones contribute to the development of ASD (1, 2). Although not all data support this theory (3–6), it is based on findings from animal and human studies that early androgen exposure produces sex differences in behavior, cognition, and brain structure and function that are relevant for ASD (7, 8) and findings that higher levels of amniotic fluid androgens are associated with poorer scores on scales of social functioning (9–12) and autistic traits (9, 11, 13, 14), as well as a higher risk of clinical ASD diagnosis (15).

One challenge in the study of in utero androgen effects on adverse neurodevelopmental outcomes is establishing temporality. Available data on fetal androgen exposure rely heavily on assessments of levels in amniotic fluid samples (13, 15, 16), which are influenced by fetal androgen secretions (17). Cross-sectional analyses have shown that children and adults with ASD have higher androgen levels and higher prevalence of androgen-related conditions compared with individuals without ASD (18–21). Thus, elevated in utero androgen levels in children later diagnosed with ASD could be subsequent to an atypical ASD-related neurogenesis that has already occurred rather than a cause of the disorder (reverse causation).

An association between elevated maternal prenatal androgen levels and ASD risk could potentially eliminate this possibility of reverse causation and strengthen the etiological evidence for androgens. While androgen concentrations in women are usually low and are not routinely clinically measured (22, 23), certain clinical conditions in women are strongly indicative of having hyperandrogenemia. Recent evidence has indeed linked 2 such maternal conditions, polycystic ovarian syndrome (PCOS) and hirsutism, with elevated progeny ASD risk, although some did not observe this overall association, or observed effects that were specific to certain subgroups (24–30).

Additionally, none of the previous publications specifically examined whether the observed effects are mediated by other androgen-related comorbidities. This question is of high importance, because a relationship between maternal PCOS and child ASD risk might not necessarily be explained by direct effects of atypical fetal androgen exposure. For instance, placental aromatase has been postulated to protect the fetus from direct exposure to abnormally high maternally derived androgens, (31) possibly explaining inconsistent observations concerning the correlation between concentrations of androgens in amniotic fluid and maternal plasma samples (32–35). On the other hand, PCOS (and hyperandrogenism) is associated with several cardiovascular and metabolic comorbidities that have been independently linked with progeny ASD risk, including glucose intolerance, dyslipidemia, hypertension, and excess weight (36–39). Additionally, women with PCOS are more likely to experience psychiatric comorbidity, use fertility treatments, or receive treatments with oral contraceptives, metformin, and antiandrogenic medications, with possible implications for the fetus (36, 40–43). Thus, an association between PCOS and progeny ASD risk could possibly be mediated by one or more of the above co-factors rather than being directly caused by high androgen concentrations. Better understanding of these possible mediating factors could improve our etiological understanding and enable designing interventions to reduce ASD risk.

We used data from a large birth cohort in Israel with detailed clinical and demographic information to further examine the association between maternal PCOS and progeny ASD risk; examine the extent to which any observed association is mediated by maternal cardiovascular, metabolic, psychiatric and fertility-related factors; and additionally evaluate the associations between other maternal androgen-related conditions and ASD risk.

METHODS

Study population

Maccabi Health Services (MHS) is Israel’s second-largest integrated health-care organization, serving as both insurer and health-care provider to 2.1 million members (25% of the Israeli population). MHS is one of 4 insurers providing equivalent, universal coverage mandated by Israel’s National Health Insurance Law. This study included all 482,504 singleton pregnancies ending in a live birth from 1999–2013 to mothers who were members of MHS throughout the year preceding their child’s birthdate. Our main analyses included 437,222 children (229,992 mothers) after excluding children who left MHS before the end of follow-up (December 31, 2016) at an age younger than 8 years, to minimize missing ASD diagnoses possibly given after leaving MHS. However, all singleton births were considered in sensitivity analyses. The study was approved by MHS’s institutional review board and by the Office of Human Research Administration at the Harvard T. H. Chan School of Public Health.

ASD case ascertainment and validation

We initially identified all children with an International Classification of Diseases, Ninth Revision (ICD-9), code of 299.x (autistic disorder) through 2016 (n = 4,758) in their record. Of these, 3,373 (70.9%) were deemed confirmed cases based on receiving benefits from the Israeli National Insurance Institute (NII), which has strict eligibility requirements based on in-person clinical and physical evaluations (44). ICD-9-identified cases not receiving NII benefits might not truly have met NII criteria but might also never have applied to NII, which could be independent of case status and instead related to some families’ choice not to pursue benefits. Therefore, one author (M.D.) individually reviewed the medical records for each of the remaining ICD-9-identified cases (n = 1,385), blinded to maternal PCOS status, and confirmed the diagnosis of 649 (46.9%) of these children based on meeting Diagnostic and Statistical Manual of Mental Disorders (DSM-IV before 2014 and DSM-V after) criteria, for a total of 4,022 confirmed ASD cases considered in main analyses.

Maternal androgen-related conditions

We used ICD-9 codes to retrieve information on maternal clinical conditions that are indicative of hyperandrogenemia. We focused primarily on PCOS (ICD-9 code 256.4) because it is the most common condition in premenopausal women that often relates to excess secretion of androgens from ovarian theca cells (45). In secondary analyses, we explored rarer conditions that could also cause hyperandrogenemia, including other and unspecified anterior pituitary hyperfunction (253.1), hyperaldosteronism (255.1), Cushing syndrome (255.0), acromegaly and gigantism (253.0), other ovary hyperfunction (256.1), and congenital adrenal hyperplasia (255.2). We also obtained information on conditions that could be caused by excess androgens, including acne (706.1), hirsutism (701.1), alopecia (704.0x), nigricans acanthosis (701.2), hidradenitis suppurativa (705.83), amenorrhea (626.0), hypomenorrhea (626.1), irregular menstrual cycle (626.4), and infertility (628.x, excluding cases of tubal (628.2), uterine (628.3), or vaginal and cervical (628.4) origin), with the assumption that if maternal hyperandrogenemia is a risk factor for progeny ASD, these maternal conditions should also show an association with the disorder. Because diagnoses of menstrual irregularities and fertility problems might be erroneously recorded for women whose pregnancy has not been properly recognized or recorded, we excluded diagnoses for these conditions that were recorded near delivery, pregnancy loss, or termination (see Web Appendix 1).

To create a more refined PCOS exposure definition, we also obtained information on available lab results for total and bioavailable testosterone, androstenedione, dehydroepiandrosterone sulfate (DHEA-S), luteinizing hormone (LH), and follicle stimulating hormone (FSH). Abnormal laboratory test results for all hormones were defined based on established clinical threshold levels routinely used in MHS (Web Table 1, available at https://doi.org/10.1093/aje/kwaa219). Because concentrations of androgen hormones change during pregnancy (17), we similarly excluded from main analyses tests performed during gestation, or near delivery, pregnancy loss, or termination (see Web Appendix 1).

Our main ascertainment definition for all conditions was having the relevant ICD-9 code recorded in the mother’s medical file prior to conception. We considered stricter ascertainment definitions for PCOS in sensitivity analyses. These included considering only diagnoses made by specialists (obstetricians, gynecologists, or endocrinologists), only mothers diagnosed with PCOS who also had at least 1 androgen hormone test result exceeding normal reference values, and only those diagnosed who also had a ratio of luteinizing hormone to follicle-stimulating hormone of >2. Because PCOS might have developmental origins (46, 47), it is possible that women first diagnosed with PCOS after delivery already had the condition preconceptionally. Therefore, in sensitivity analyses, we considered all mothers with PCOS, regardless of diagnosis time. For all analyses considering individual conditions, the unexposed (comparison) group included children whose mothers did not meet the ascertainment definitions for the condition considered.

Maternal cardiovascular, metabolic, and fertility conditions

We obtained detailed information on mothers who were clinically diagnosed, treated, or had laboratory test findings indicative of the following conditions: prediabetes, diabetes, and gestational diabetes; hypertension and preeclampsia; hypertriglyceridemia, hypercholesterolemia and hypoalphalipoproteinemia; and excess weight. We additionally used medication dispensing data and records of medical procedures to obtain information on mothers who conceived following assisted reproductive therapy, and we also obtained information on any dispensing of progesterone, metformin, oral contraceptives, and antiandrogenic medications in the 3 months preceding conception through delivery. Information on infant birthweight (grams) and gestational age (weeks) was obtained from hospital birth records. See Web Table 2 for detailed ascertainment definitions.

Covariate information

All models adjusted for the following covariates as recorded at delivery: calendar year, maternal age, residential district, and residential enumeration area (homogeneous geographical unit of approximately 3,000 people, the smallest unit of analysis of the Israeli census) socioeconomic status (on a scale of 1 to 10, based on a poverty index as previously described (48)). Information was also obtained on enumeration areas with high proportions of minority (Israeli Arabs and/or Jewish Orthodox) and immigrant residents. We included only maternal age because paternal age was available for only 60% of the children in our cohort and was highly correlated with maternal age (r = 0.8; P < 0.01). Additionally, thyroid disorders have been associated with PCOS, androgen homeostasis, and ASD risk (49–55). We thus obtained information on all women diagnosed with, treated for, or suspected to have hypothyroidism, hyperthyroidism, or other thyroid-related anomalies preconceptionally (Web Table 2), and we adjusted for these conditions in all analyses. Finally, PCOS and hyperandrogenism have been associated with several psychiatric conditions, and maternal psychiatric morbidity has also been implicated in progeny ASD risk (43, 56, 57). Importantly, whether maternal psychiatric morbidity mediates or confounds the PCOS-ASD association is unclear. Thus, in sensitivity analyses, we additionally included information on several maternal psychiatric conditions (see Web Appendix 2 for additional information).

Statistical analysis

Generalized estimating equations models.

We used generalized estimating equations (GEE) with independence working correlation structure and a logit link to estimate odds ratios and 95% confidence intervals, employing robust variance estimators to account for multiple births per mother (58, 59). An assumption of GEE is that the cluster size (in this case, the number of progeny) is uninformative (i.e., that the risk of the outcome given the model covariates is independent of the cluster size). To evaluate the possibility of informative clustering, we repeated the GEE analyses using the inverse of maternal parity in 1999–2013 as weights in sensitivity analyses (58, 59). In all analyses, we included a quadratic term for SES since multivariable analyses using generalized additive models (GAMs) with penalized splines suggested a curvilinear association with ASD risk. A multiplicative interaction term was used to evaluate effect modification by child’s sex for the main PCOS ascertainment definition. We initially considered PCOS as well as each of the other androgen-related maternal conditions in separate models but additionally explored models that included conditions caused by androgen excess with mutual adjustments. We performed several sensitivity analyses, including adding adjustment for maternal psychiatric morbidity, restricting to mothers with at least 5 years of follow-up in MHS before delivery, and considering all singleton births, regardless of whether they left MHS, using a Cox model, defining time from birth as the timescale and censoring date as the first date of ASD indication, last MHS contact, or end of follow-up, whichever came first.

Multiple mediation analysis.

We conducted a multiple mediation analysis to discern the nature of the association between maternal PCOS and ASD risk. To this aim, we used the medflex package in R (R Foundation for Statistical Computing, Vienna, Austria), which relies on a novel class of conditional mean models (termed natural effect models) that enable users to decompose the estimated total effect of maternal PCOS on ASD risk into a pure direct effect (estimated effect of PCOS not through any of the mediators considered), and a total indirect effect (estimated effect of PCOS through any of the mediators considered, also accounting for possible exposure-mediator interactions). The approach also allows the user to account for exposure-confounder, mediator-confounder, and mediator-mediator interactions. Confidence intervals for these effects can be obtained by bootstrapping (60, 61).

For the main mediation analysis, we considered the following 4 main binary mediator groups: dyslipidemia, glucose intolerance (having either diabetes, prediabetes, or gestational diabetes), excess weight, and hypertension (hypertension or preeclampsia). We included terms for interaction between each of these groups and PCOS and also included all 2-way interactions between the mediators themselves. In a separate model, we tested for effect modification of the direct and indirect effects by child sex. In secondary analyses, we additionally included other possible mediating factors, including assisted reproductive therapy, other hormonal treatments (antiandrogens, progesterone, and/or oral contraceptives), treatment with metformin, gestational age at birth, and birthweight, again allowing for all 2-way interactions. As detailed in Web Appendix 2, in additional analyses we also explored the possibility that maternal psychiatric morbidity might mediate or confound the PCOS-ASD association. All models adjusted for the same set of covariates as the main GEE analyses.

RESULTS

Overall, 17,447 (4.0%) children were born to mothers diagnosed with PCOS preconceptionally. Compared with women without the condition, women with PCOS had a higher prevalence of other comorbidities associated with androgen excess, including menses-related problems, infertility, dermatologic problems, and cardiovascular-metabolic comorbidities. Women with PCOS were also more likely to use assisted reproductive and other hormonal therapies and experienced more pregnancy complications (Table 1). Median age at first ASD diagnosis for the 4,022 children diagnosed with the disorder was 3 (interquartile range, 2–5) years.

Table 1.

Characteristics and Prevalence of Exposure to Maternal Androgen-Related Conditions and Metabolic Disorders According to Maternal Preconception Polycystic Ovarian Syndrome Status for Children Born in 1999–2013 (n = 437,222), Maccabi Health Services, Israel

Characteristic	Maternal PCOS (n = 17,447)		No PCOS (n = 419,775)
	No.	%	No.	%
Female sex (child)	8,335	47.8	203,744	48.5
District
North	3,131	18.0	66,296	15.8
Sharon	3,506	20.1	78,499	18.7
South	2,318	13.3	65,128	15.5
Central	4,614	26.5	106,933	25.5
Jerusalem and Shfela	3,878	22.2	102,919	24.5
High minority and immigrant subpopulationa
Israeli Arab	664	3.8	22,578	5.4
Jewish Orthodox	1,287	7.4	56,270	13.4
Immigrants	1944	11.1	45,154	10.8
Androgen-related conditionsb
Acne	6,915	39.6	95,906	22.9
Alopecia	3,488	20.0	50,137	11.9
Hirsutism	2,557	14.7	10,642	2.5
Acanthosis nigricans	68	0.4	292	0.1
Hidradenitis suppurativa	292	1.7	4,037	1.0
Infertility	9,155	52.5	80,793	19.3
Amenorrhea	7,887	45.2	50,376	12.0
Hypomenorrhea	3,130	17.9	11,408	2.7
Irregular period	7,271	41.7	47,388	11.3
Cardiovascular and metabolic conditionsc
Diabetes mellitus	225	1.3	1,567	0.4
Prediabetes	5,623	32.7	111,524	26.7
Excess Weight	4,617	26.5	48,202	11.5
Hypertension	419	2.4	4,193	1.0
Dyslipidemia	5,400	31.0	67,790	16.2
Assisted reproductive therapiesc
Gonadotropins and clomiphene	5,119	29.3	38,791	9.2
In vitro fertilization	1,292	7.4	14,492	3.5
Other treatmentsc
Progesterone therapy	4,434	25.4	44,270	10.6
Antiandrogens	168	1.0	1928	0.5
Oral contraceptives	799	4.6	14,071	3.4
Metformin	595	3.4	983	0.2
Pregnancy complicationsc
Pre-eclampsia and hypertension in pregnancy	627	3.6	9,514	2.3
Gestational diabetes	2,292	13.1	28,165	6.7
Maternal psychiatric morbidityd
Schizophrenia	66	0.4	1,089	0.3
Episodic mood disorders	253	1.5	3,323	0.8
Delusional and nonorganic psychoses disorders	92	0.5	1,459	0.4
Autism spectrum disorder	13	0.1	161	<0.1
Neurotic disorders and anxiety	2,382	13.7	39,615	9.4
Personality disorders	314	1.8	3,815	0.9
Adjustment disorders	1,447	8.3	19,813	4.7
Depression	1,098	6.3	16,273	3.9
Conduct disorders	40	0.2	592	0.1
Hyperkinetic syndrome	297	1.7	3,489	0.8
Eating disorder	354	2.0	3,906	0.9
Tics	36	0.2	941	0.2
Thyroid conditionsc
Hypothyroidism	1,633	9.4	21,813	5.2
Hyperthyroidism	591	3.4	8,440	2.0
Other thyroid conditions	1,170	6.7	20,152	4.8
Maternal age, yearse	31.2 (4.5)		31.2 (5.2)
Socioeconomic statusaef	6.3 (1.9)		6.0 (1.9)
Gestational age, weekseg	39.0 (1.8)		39.2 (1.7)
Birthweight, kgeh	3.3 (0.5)		3.3 (0.5)

Abbreviations: ICD-9, International Classification of Diseases, Ninth Revision; PCOS, polycystic ovarian syndrome.

^a Based on residential enumeration area, as defined in the main text.

^b Diagnoses based on ICD-9 codes, as defined in the main text.

^c See Web Table 2 for specific ascertainment definitions.

^d See Web Appendix 2 for specific ascertainment definitions.

^e Values are expressed as mean (standard deviation).

^f On a scale of 1 to 10, as defined in the main text.

^g Excluding 5,885 children (1.3%) with missing values.

^h Excluding 3,020 children (0.7%) with missing values.

GEE analyses indicated elevated odds ratios for ASD in children of mothers diagnosed preconceptionally with PCOS (adjusted odds ratio (OR) = 1.42, 95% confidence interval (CI): 1.24,1.64, Table 2). Effect estimates were similar in analysis using stricter ascertainment definitions for PCOS and were slightly stronger in the subgroup of women who also had laboratory results indicating hyperandrogenemia. Results were similar with further adjustment for maternal psychiatric morbidity and in analyses that considered all mothers with PCOS, regardless of diagnosis time. Consistent results were also observed when restricting to mothers with at least 5 years of continuous MHS membership before delivery (n = 15,972, adjusted OR = 1.39, 95% CI:1.20, 1.62) or when using a Cox model to consider all singleton births regardless of whether they left MHS (n = 18,889, adjusted hazard ratio = 1.44, 95% CI: 1.26, 1.64). We did not observe effect modification for PCOS by child sex (P for interaction = 0.59). Results from cluster-weighted models were very similar to those from models without this weighting. The prevalence of other conditions that could cause hyperandrogenemia was too low to draw meaningful conclusions for each condition separately. However, children born to mothers diagnosed with any of these conditions also had an elevated ASD risk (n = 5,172, adjusted OR = 1.41, 95% CI: 1.10, 1.79).

Table 2.

Odds Ratios for Autism Spectrum Disorder in Children Born to Mothers Diagnosed With Polycystic Ovarian Syndrome, Maccabi Health Services, Israel, 1999–2013

PCOS Ascertainment Definition	No. of Children Born to Mothers With the Condition	ASD		Crude a		Adjusted b		Additional Adjustment c
		No.	%	OR	95% CI	OR	95% CI	OR	95% CI
ICD-9 code recorded
No PCOS	419,775	3,788	0.90	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	17,447	234	1.34	1.49	1.30, 1.72	1.42	1.24, 1.64	1.40	1.22, 1.61
Alternative Ascertainment Definitions
Diagnosis by specialists
No PCOS	421,448	3,807	0.90	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	15,774	215	1.36	1.52	1.31, 1.75	1.45	1.25, 1.67	1.42	1.23, 1.65
LH:FSH ratio > 2
No PCOS	429,423	3,914	0.91	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	7,799	108	1.38	1.53	1.25, 1.87	1.45	1.18, 1.78	1.42	1.16, 1.74
Androgen levels testedd
No PCOS	423,198	3,826	0.90	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	14,024	196	1.40	1.55	1.34, 1.81	1.48	1.27, 1.72	1.45	1.25, 1.69
Elevated androgen levelse
No PCOS	433,042	3,957	0.91	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	4,180	65	1.56	1.71	1.33, 2.21	1.64	1.27, 2.12	1.61	1.25, 2.08
Ever diagnosis
No PCOS	414,340	3,726	0.90	1.00	Referent	1.00	Referent	1.00	Referent
PCOS	22,882	296	1.29	1.44	1.27, 1.64	1.43	1.26, 1.62	1.40	1.24, 1.59

Abbreviations: ASD, autism spectrum disorder; CI, confidence interval; FSH, follicle stimulating hormone; ICD-9, International Classification of Diseases, Ninth Revision; LH, luteinizing hormone; OR, odds ratio; PCOS, polycystic ovarian syndrome;

^a Crude (unadjusted) analysis.

^b Adjusted for the following covariates as recorded at the time of birth: calendar year, maternal age, residential district, socioeconomic status, and high minority (Israeli Arabs and/or Jewish Orthodox) and immigrant subpopulations at residential enumeration area.

^c Similar analysis to that described in footnote b, with additional adjustments for maternal psychiatric morbidity (see Web Appendix 2 for details).

^d Mothers with a PCOS diagnosis based on ICD-9 codes who also received laboratory tests for serum androgen levels, regardless of tests results.

^e Mothers with a PCOS diagnosis based on ICD-9 codes who also had laboratory test results indicating hyperandrogenemia (see Web Table 1 for clinical reference levels for androgen hormones).

Multiple mediation analysis indicated that the estimated total effect of PCOS on ASD risk was driven largely by the pure direct effect through pathways that were independent of the 4 main metabolic and cardiovascular mediator groups considered (Table 3). There was no evidence for effect modification of either the direct or indirect estimated effects by child sex. Additional consideration of assisted reproductive therapy, other hormonal treatments, and metformin use as potential mediating factors, as well as of birthweight and gestational age, resulted in a higher total indirect effect, yet a direct, unmediated pathway was still implicated. Findings were consistent with further adjustment for maternal psychiatric conditions as possible confounders, or when psychiatric morbidity was included as an additional potential mediator in the analysis.

Table 3.

Multiple Mediation Analysis of the Estimated Total Effect of Maternal Polycystic Ovarian Syndrome on the Risk of Autism Spectrum Disorder in Progeny, Maccabi Health Services, Israel, 1999–2013

Mediators Considered	Pure Direct Effect		Total Indirect Effect a
	OR	95% CI b	OR	95% CI b
Cardiovascular and metabolic comorbiditiesc	1.41	1.22, 1.64	1.02	0.97, 1.07
Additional consideration of ART, other hormonal treatments, and treatment with metformind	1.28	1.09, 1.52	1.12	1.02, 1.21
Additional consideration of birthweight and gestational agee	1.28	1.10, 1.52	1.11	1.02, 1.21
Additional consideration of maternal psychiatric morbidityf	1.26	1.08, 1.47	1.13	1.04, 1.23

Abbreviations: ART, assisted reproductive therapy; CI, confidence interval; OR, odds ratio.

^a Sum of the pure indirect effect (effect due to mediation only) and the mediated interaction effect (effect due to mediation and exposure-mediator interactions).

^b Normal approximation bootstrap confidence intervals based on 1,000 bootstrap samples.

^c Glucose intolerance, dyslipidemia, hypertension, and excess weight, as defined in the main text and in Web Table 2.

^d Same as in footnote c, with additional consideration of ART (in vitro fertilization, gonadotropins, and/or clomiphene treatment), other hormonal treatments (antiandrogens, progesterone, and/or oral contraceptives), and metformin treatment (see Web Table 2 for specific ascertainment definitions).

^e Same as in footnote d, with additional consideration of birthweight and gestational age. For the children with missing gestational age (1.3%) or birth weight (0.7%) data we assigned the mean values in our study population but excluded this group in sensitivity analyses to confirm consistency of the results.

^f Same as in footnote e, with additional consideration of maternal psychiatric morbidity. See Web Appendix 2 for details on maternal psychiatric conditions.

In secondary analyses we evaluated other conditions that might be caused by hyperandrogenemia to investigate whether the association is specific to PCOS (Table 4). Elevated odds ratios were observed for children born to mothers diagnosed preconceptionally with hirsutism, acne, and fertility problems, while null results were observed for all menses-related morbidities and alopecia. The results were overall consistent when mutually adjusting for all these conditions simultaneously.

Table 4.

Odds Ratios for Autism Spectrum Disorder in Children According to Mothers’ Preconception Diagnosis With Other Androgen-Related Conditions, Maccabi Health Services, Israel, 1999–2013

Condition	No. of Children Born to Mothers With the Condition	ASD		Crude Analysis a		Adjusted Analysis b		Mutually Adjusted c
		No.	%	OR	95% CI	OR	95% CI	OR	95% CI
Menses-related condition
Amenorrhea	58,263	602	1.03	1.15	1.05, 1.25	1.09	1.00, 1.19	1.05	0.95, 1.15
Hypomenorrhea	14,538	146	1.00	1.10	0.92, 1.31	1.01	0.85, 1.21	0.94	0.79, 1.13
Irregular menses	54,659	532	0.97	1.07	0.97, 1.17	1.05	0.95, 1.15	1.00	0.91, 1.10
Infertility	89,948	1,021	1.14	1.32	1.22, 1.42	1.19	1.10, 1.28	1.18	1.09, 1.27
Cutaneous-related conditions
Acne	102,782	1,058	1.03	1.16	1.08, 1.25	1.15	1.07, 1.24	1.13	1.04, 1.22
Alopecia	53,625	563	1.05	1.17	1.06, 1.28	1.02	0.93, 1.12	0.99	0.90, 1.09
Hidradenitis suppurativa	4,329	54	1.25	1.37	1.04, 1.80	1.22	0.93, 1.61	1.20	0.91, 1.58
Nigricans acanthosis	360	5	1.39	1.52	0.64, 3.62	1.42	0.59, 3.40	1.34	0.56, 3.21
Hirsutism	13,199	165	1.25	1.38	1.17, 1.62	1.30	1.10, 1.53	1.23	1.05, 1.46

Abbreviations: ASD, autism spectrum disorder; CI, confidence interval; OR, odds ratio.

^a Crude (unadjusted) univariable analysis, considering each condition separately. Control group for each model includes all women without a diagnosis for the specific condition considered.

^b Similar analysis to that described in footnote a, but with adjustment for the following covariates as recorded at the time of birth: calendar year, maternal age, residential district, socioeconomic status, and high minority (Israeli Arabs and/or Jewish Orthodox) and immigrant subpopulations at residential enumeration area.

^c Considering all conditions simultaneously, with adjustments for the same set of covariates as in footnote b.

DISCUSSION

We found that ASD risk is higher in children of mothers with PCOS. Effect estimates were consistent when using different PCOS ascertainment definitions and were more robust in the subgroup of women who also had laboratory results indicative of biochemical hyperandrogenemia. Our mediation analyses additionally indicated that the estimated effect of PCOS on ASD risk was only partly mediated by androgen-related metabolic, cardiovascular, and psychiatric comorbidities or by hormonal and reproductive therapies that are associated with hyperandrogenemia. Our results for other conditions that could cause, or be caused by, excess androgens were less conclusive but generally consistent with the hypothesis that maternal hyperandrogenemia is a possible risk factor for ASD.

Our results agree with previous publications that reported worse scores in scales of ASD traits and a higher risk of clinical ASD diagnosis in offspring of mothers with PCOS (24, 25, 27–30). Our findings that this estimated effect appears to be only partly mediated by other comorbidities, combined with previous observations that prenatal excess androgen exposure is associated with cognitive, behavioral, and neuroanatomical changes that are typical of ASD (13, 16, 62, 63), add to a growing body of evidence supporting a possible direct involvement of androgens in ASD etiology. If true, the null results we observed for other androgen-related conditions, including some menses-related disorders and alopecia, and the relatively modest effect estimate observed for acne, might reflect the multifactorial nature of these relatively common conditions, for which hyperandrogenemia is just one of many possible causes (64).

There has been some debate concerning the effectiveness of the placental aromatase barrier. While evidence concerning the strength of the correlation between androgen concentrations in maternal circulation and amniotic fluid samples remains inconsistent, elevated in utero androgen levels have been observed in pregnant women with PCOS (33, 34), suggesting that at least in some pregnancies, the barrier might be ineffective or incomplete. Alternatively, the placental aromatase aromatizes androgens to estrogens, which have many regulatory neurodevelopmental properties (65), and a recent analysis of stored amniotic fluid samples observed a positive correlation between estrogen levels and subsequent ASD diagnosis (66). Thus, it is possible that PCOS-related hyperandrogenemia translates to elevated prenatal estrogen exposure, increasing ASD risk.

Additional considerations for the observed findings also merit discussion. Women with PCOS have higher risks for several neuropsychiatric disorders, including ASD (20, 25). Interestingly, risks for psychiatric morbidities are also higher among unaffected female and male siblings of these women (43). While this finding could be related to the higher burden of other endocrine problems also seen in these siblings, it might also reflect genetic predispositions that increase risks for both PCOS and neuropsychiatric developmental disorders, possibly through independent mechanisms (43, 67, 68). However, shared familial factors did not seem to explain the observed PCOS-ASD association in a recent analysis that compared PCOS-exposed offspring with PCOS-unexposed cousins (69). Alternatively, shared environmental factors could also drive this association, and recent evidence indeed linked perinatal exposure to several classes of environmental toxicants with both PCOS and ASD (46, 70). Additionally, a link between maternal PCOS and ASD risk does not necessarily imply a unidirectional association between in utero androgen concentrations and ASD risk. In a recent analysis of singleton male newborns in this cohort, we observed that ASD risk was higher among newborns with hypospadias or cryptorchidism, conditions that are highly indicative of a hypoandrogenic in utero environment (48). This could indicate a potentially more complicated relationship between fetal androgenic environment and neurodevelopment, possibly suggesting that both hyper-and hypoandrogenemia, or alternatively general maternal endocrine dysregulation, play etiological roles in ASD.

Strengths of this study include prospectively collected medical information from a large population-based cohort with universal access to health-care services; complete validation of ASD cases; and having access to medication-dispensing information and laboratory measurements. There are also several limitations that merit discussion. While we relied on several data sources for ascertaining maternal comorbidities, we might have missed some women with subclinical symptoms. However, the readily accessible community-based health-care system in Israel likely minimizes the likelihood of completely missing patients with milder symptoms. Additionally, we relied on maternal conditions that are indicative of excess androgens, but we were only partly able to verify the presence of biochemical hyperandrogenemia. Women with PCOS have low levels of sex hormone–binding globulin, and direct assays of total testosterone, as were primarily used in our cohort for assessment of hyperandrogenemia due to the high cost of tests for bioavailable testosterone (fraction not bound to sex hormone–binding globulin and thus biologically active), are highly variable in these women and might not accurately reflect the true biological activity of the hormone (64). Nonetheless, we observed more robust effect estimates in women whose hyperandrogenemia was ascertained using this cruder biomarker. Further, while the comorbidities we considered as mediators could be linked with ASD risk through multiple different pathways, some might cause lower concentrations of sex hormone–binding globulin, thus possibly resulting in elevated free testosterone levels and consequent exacerbation of hyperandrogenemia-related effects (64). Therefore, it is possible that some of the indirect effect of PCOS on ASD risk that was attributed in our analysis to mediation still operates via androgen-specific pathways. Finally, like other analyses, the estimation of direct and indirect effects requires various assumptions of no unmeasured confounding (71). While we controlled for an array of possible confounders and conducted various sensitivity analyses, some residual confounding might still have influenced our results.

Overall, our study suggests that PCOS and other maternal conditions indicative of excess androgen are a risk factor for ASD and that the underlying mechanisms appear to be only partly mediated by comorbidities and treatments that are linked with hyperandrogenemia. Future research should focus on assessing the involvement of environmental factors in the etiologies of both maternal hyperandrogenemia and ASD, as well as examining the familial co-aggregation of ASD and androgen-related conditions to assess the contribution of genetic factors.