Bioavailable Testosterone Predicts Autistic Traits in Women with and without Autism

Thursday, May 12, 2016: 11:30 AM-1:30 PM
Hall A (Baltimore Convention Center)
A. L. Pohl1, J. R. Lloyd2, L. Ruta3, B. Auyeung4 and S. Baron-Cohen1, (1)Autism Research Centre, University of Cambridge, Cambridge, United Kingdom, (2)Computational and Biological Learning Lab, University of Cambridge, Cambridge, United Kingdom, (3)Institute of Clinical Physiology, National Research Council (CNR), Taormina, Italy, (4)University of Edinburgh, Edinburgh, United Kingdom
Background:   Individual differences in fetal testosterone levels predict individual differences in autistic traits, specific behaviors (such as eye-contact and language), and performance on specific cognitive tasks (such as those measuring theory of mind or attention to detail) related to autism. It is not yet known if testosterone levels in adults are related to autistic traits. 

Objectives:   To test whether bioavailable testosterone, estimated by the free androgen index (FAI), predicts autistic traits in adults, measured using the Autism Spectrum Quotient (AQ). 


We recruited two cohorts of adults with and without autism, and analyzed each cohort separately by sex. Cohort 1 comprised 164 individuals (28 females with an autism spectrum condition (autism), 55 control females, 32 males with autism, and 49 control males). Cohort 2 comprised 103 individuals (19 females with ASC, 26 control females, 44 males with autism, and 14 control males). We measured total testosterone and sex hormone binding globulin (SHBG) levels in serum, and calculated the ratio of total testosterone to SHBG to get the FAI. We tested whether FAI predicted total AQ scores using linear regression, finite mixture modeling, and Bayesian regression. We did not include diagnosis as a grouping variable, as the multifactorial inheritance of autism suggests no difference between diagnostic groups in response to a risk factor in the absence of gene-environment interaction. Instead, we used data-driven approaches to account for the bimodal distribution of AQ scores in our sample caused by including cases and controls.

Results:   FAI predicted AQ only in females in a linear regression model (Cohort 1 females, BFAI = 7.445, SE=2.780, t = 2.68, p = 0.00903) (Figure 1). However, we observed a multimodal distribution of error, as expected. To address this problem, we used finite mixture modeling, and found a significant relationship between FAI and AQ in Cohort 1 females, as well as two latent clusters that roughly, but not perfectly, correspond to the distribution of cases and controls in our sample (BFAI = 6.35, SE = 1.43, z = 5.90, p = 3.68 * 10-9) (Figure 1). To confirm this finding, we used probabilistic programming to encode a model that could account for possible latent clusters in our data without determining the number of clusters a priori. We performed Bayesian inference in our generative model using data from Cohorts 1 and 2. In females, data from Cohort 1 were consistent with frequentist results, and data from Cohort 2 confirmed a positive relationship between FAI and AQ, and the existence of latent structure not accurately accounted for by diagnostic labels in our data.


Autistic traits in adults are predicted by bioavailable testosterone in females only. This result is consistent with the finding that testosterone administration lowers mind-reading ability in women (van Honk et al, 2011, PNAS). Furthermore, this study demonstrates the importance of choosing appropriate statistical models for regression when the data do not meet the assumption of being normally jointly distributed. This study further supports a role for testosterone in sex-differential autism risk.