Our initial sample was comprised of N = 145 participants. After the initial data inspection, two participants, one from each group, were removed because they were outliers. The final sample had N = 143 participants. Of these, 31 (21.7%) were participants diagnosed with ASD, while 112 (78.3%) were participants from the general population. We estimated our sample size using the pROC package in R [15], which allows power testing for receiver operating characteristic (ROC) curves. For a minimal expected area under the curve (AUC) = 0.7, at a standard $\alpha$ = 0.05 with 80%, the required clinical sample size was N = 30.28 for both groups, following recommendations for ROC-based power analysis [16].

Participants from the clinical group were recruited from the authors’ practices, and references from other mental health specialists in Romania through recruiting announcements made by the authors in the mental health network. For the clinical sample, purposive sampling was chosen to select participants. This non-probability sampling method enabled us to recruit sufficient clinical participants, ensuring adequate statistical power for the primary analyses (criterion validity). All participants from the clinical sample underwent a thorough clinical interview to confirm the ASD diagnosis based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) diagnostic criteria [7] and their personal and family history. Extensive interviews with the patients’ main caregivers were conducted to confirm the existence of ASD symptoms since childhood. This method was chosen due to the absence of a registry for ASD diagnoses in Romania. All patients included in the clinical sample received their ASD diagnosis at an adult age (18 and over), prior to their inclusion in this study.

The non-clinical sample was selected through social media platforms (WhatsApp: https://www.whatsapp.com, Facebook: https://www.facebook.com, and Instagram: https://www.instagram.com) to allow access to a more diverse population in this study. For the non-clinical sample, a non-probability sampling method was also chosen, namely convenience sampling, mainly due to accessibility reasons.

All study participants were native Romanian speakers. The inclusion criteria for the clinical sample were: individuals aged 18 years or older, a diagnosis of ASD without intellectual and language deficits received at age 18 or later, and confirmation of the diagnosis by the research team. All individuals diagnosed before the age of 18, as well as those younger than 18, were excluded. The inclusion criteria for the general population group consisted of individuals aged 18 and older without a diagnosis of ASD.

Both groups completed a Google Forms questionnaire containing 245 items: informed consent, confidentiality agreement, demographics (age, gender, area, education level, level of contact with mental health specialists, family history of ASD), AQ (50 items), EQ (60 items), and PDSQ (125 items), with a 20 to 25-minute completion time. Data was collected during December 2024–February 2025. The anonymized data set can be found at the following link: https://osf.io/qxrmj/?view_only=5a3c3ccfef2c48c19d54bb7aa796c514.

The AQ scale is concordant with the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (DSM-IV-TR) criteria [17], recording the autism spectrum elements in a 50 item self-report questionnaire analyzing 5 functional characteristics of ASD throughout 10 items each: social abilities impairment; attention disturbance (flexibility, mobility, focus); attention to details; communication impairment; and poor imagination [18]. Scoring is made on a 1–4 Likert scale (1—total agreement, 4—total disagreement), the subject receiving one point on each item if an abnormal score is recorded. To avoid bias, about half of the items were negatively formulated, the other half being positively formulated, corresponding to typical responses for people with a diagnosis of ASD without cognitive and language delay. The AQ score varies between 0 and 50 [18, 19].

The EQ scale consists of 60 items, 40 analyzing empathy and 20 fillers with a score varying between 0 and 80 scored on the same Likert scale as the AQ, with 1 point for average empathy and 2 points for high empathy level [20]. Adults diagnosed with ASD without language and cognitive delay score high in AQ and low in EQ, individuals with ASD having a mean AQ of 35.8 compared to the control group (AQ = 16.4) and a mean EQ of 20.4 compared to 42.1 for the control group [20].

AQ and EQ are screening instruments, designed to assess autistic traits and empathic capacity, respectively. The inverse correlation between them suggests that higher levels of autistic traits (as measured by AQ) are typically associated with lower levels of affective and cognitive empathy (as measured by EQ) [21].

PDSQ is a self-report instrument based on Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) [17] designed to screen thirteen of the most frequent psychiatric disorders, consisting of 125 items classified in 13 subscales (one for each disorder), and requiring 15 to 20 minutes to complete [14, 22]. The thirteen disorders screened in PDSQ were classified in 6 categories according to the DSM-IV criteria [17]: eating disorders: bulimia/binge-eating disorder; mood disorders: major depressive disorder (MDD), dysthymia; anxiety disorders: panic disorder (PAN), agoraphobia (AGO), post-traumatic stress disorder (PTSD), OCD, generalized anxiety disorder (GAD), and social phobia (SOC); substance use disorders: alcohol and drug; somatoform disorders: somatization disorder and hypochondriasis, and psychosis (PSY) [23]. Given the fact that DSM-5 takes a dimensional approach rather than a categorial one [7], the original six category classification is not necessary anymore, as PTSD and trauma-related disorders received a different chapter, dysthymia became persistent depressive disorder, and somatization disorder and hypochondriasis became illness anxiety disorder [7], but the questions of the PDSQ are still relevant as they give insight on all the pathologies mentioned above, because the diagnosis criteria have not suffered major changes between DSM-IV and DSM-5 [24]. All the questions of PDSQ are dichotomic scored, every “yes” receiving one point if the item is applicable and every “no” receiving 0 if the item is not applicable. Each of the 13 disorders mentioned above receives a score, and the total score functions as a global indicator of psychopathology. Higher subscale scores for PDSQ suggest a higher likelihood of the corresponding psychiatric condition and may indicate the necessity of further clinical evaluation [23].

To explore potential correlates of the AQ and EQ, we relied on the PDSQ subscales and computed Pearson-correlations in the combined sample. The rationale for using both the clinical and the non-clinical participants is that AQ and EQ are considered screening instruments, thus primarily focused on the general population.

We examined the factorial structure of the AQ to evaluate its structural validity in the combined sample. We conducted a confirmatory factor analysis (CFA) using the diagonally weighted least squares (DWLS) estimation method to do this. The DWLS method is particularly suited for ordinal or dichotomous data, especially when the assumption of normality is not met [25]. To assess model fit, we considered several key indices: the comparative fit index (CFI), the root mean square error of approximation (RMSEA), and the standardized root mean squared residual (SRMR). Additionally, we reported the Tucker–Lewis index (TLI). Interpretation of these indices followed the guidelines of Hu and Bentler [26] and Browne and Cudeck [27], which suggest that CFI and TLI values close to 0.95 or higher indicate a good fit, while RMSEA and SRMR values below 0.08 also suggest an acceptable fit, and 0.09 for smaller sample size in SRMR [28].

To estimate reliability, we reported both Cronbach’s alpha and McDonald’s omega. While Cronbach’s alpha is a widely used measure of internal consistency, McDonald’s omega is considered a more robust alternative, as it is less sensitive to violations of the normality assumption [29].

Criterion validity was deduced by conducting two independent t-tests between the clinical and the non-clinical groups regarding the AQ and the EQ scores. Finally, sensitivity and specificity analyses were conducted using the ROC and the AUC. The ROC curve is a widely used plot of an instrument’s sensitivity relative to one minus specificity at each cut-off score (multiple confusion matrices). For interpreting the AUC (which can take values between 0.5 for random performance and 1.0 for perfect performance), we utilized the recommendations of Hanley and McNeil [30], who suggest that a fair diagnostic accuracy can be situated between 0.70 and 0.79, a good accuracy between 0.80 and 0.89 and an excellent accuracy between 0.90 and 1.00. Since AQ and EQ were designed to screen for ASD manifestations in the general population, the cut-off score was determined by favoring sensitivity over specificity.

All the statistical analyses were conducted in R, version 4.4.1 (R Foundation for Statistical Computing, Vienna, Austria) [15].

Regarding the demographic characteristics of our samples, the age for the clinical group was M = 27.96 (SD = 7.39, MIN = 19, MAX = 49), and for the general group, it was M = 38.98 (SD = 11.61, MIN = 20, MAX = 68). Other relevant data are presented in Table 1 (Yates’ continuity correction for chi-square difference test was applied when the expected frequencies were below 5).

Regarding the possible associated symptoms of ASD, the strongest correlates of AQ scores in the combined sample (clinical and general) were PTSD, OCD, SOC, and GAD manifestations. In our sample, both clinically diagnosed and non-clinical participants experienced more anxiety-related manifestations as the AQ score increased. Correlation data is presented in Table 2.

We also calculated Pearson’s correlations between AQ, EQ, and age in the combined sample. A significant negative correlation was found between AQ and age (r = –0.21, p = 0.010), indicating that older participants tended to have lower scores on AQ, and conversely. However, no such relationship was identified between EQ and age.

Both AQ and EQ were highly reliable instruments in our sample. Regardless of the selected group (combined, clinical, and general), both instruments delivered satisfactory internal consistency indicators (well above the threshold of 0.70 for Cronbach’s alpha and McDonald’s omega). Data regarding their reliability is presented in Table 3.

Regarding the factorial structure of the AQ, we also tested the 5-factor structure in the combined sample that comprised both clinical and non-clinical participants. The CFA on the five-factor model returned adequate fit indices: CFI = 0.933, TLI = 0.930, RMSEA = 0.066, SRMR = 0.151. Although the SRMR exceeds the conventional threshold of 0.08, this is to be expected given the low sample size and the categorical nature of the AQ items (see Dolfi et al., 2025 [31] for an extended discussion regarding the AQ factor structure in the Romanian population). This indicates that when clinical participants are present in the sample, AQ discriminates modestly between social skills, attention switching, attention to detail, communication, and imagination deficits as components of ASD.

Regarding the criterion validity, both AQ and EQ returned significant differences between the clinical and the non-clinical group, indicating their capacity to discriminate autism-related manifestations when the clinical diagnosis is established as a comparison criterion. Regarding the AQ score, a t (49.83) = –13.76, p 0.001 was obtained (M_nonclinical = 16.28, M_clinical = 33.61). Similar results were obtained when EQ was considered: t (53.77) = 9.20, p 0.001 (M_nonclinical = 46.00, M_clinical = 26.38).

It is also worth mentioning that, regarding the EQ filler items, no difference was recorded between the clinical group and the non-clinical group, t (50.32) = 1.08, p = 0.280.

The ROC graph for the AQ scale can be visualized in Fig. 1. The AUC was 97.2%, indicating that, for our sample, AQ had excellent accuracy in detecting ASD cases. Regarding the optimal cut-off score, multiple scores were considered, and subsequent confusion matrices were calculated (see Table 4). The optimal cut-off score that achieved the highest sensitivity and optimal specificity was 21. A cut-off score of 21 for AQ correctly classified 100% of the clinically diagnosed ASD participants as ASD cases (true positives) while correctly identifying 80% of the non-ASD participants as not having ASD (true negatives), with an 84% accuracy, at a 0.06 optimal threshold (Youden’s index = 0.80).

The ROC curve for EQ is presented in Fig. 2, with an AUC of 90%, indicating excellent accuracy of AQ in identifying ASD cases within our sample. To determine the optimal cut-off score, various values were evaluated, and confusion matrices were generated (see Table 5). The score of 26 emerged as the most effective, achieving the highest sensitivity while maintaining optimal specificity. At a 0.43 optimal threshold (Youden’s index = 0.67), EQ correctly identified 74% of clinically diagnosed ASD participants as ASD (true positives) and 92% of non-ASD individuals as not having ASD (true negatives), resulting in an overall accuracy of 88%.