Table 1. Participant Characteristics by Diagnostic Group

Below, CB recreated KC’s demographic table to ensure that we wrangled the raw data in the same way.
Table 1. Participant Characteristics by Diagnostic Group
TD (n = 19) ASD (n = 16) Range t p
Age (years) 15.26 (3.16) 16.75 (3.32) 9.00–22.00 -1.35 0.187
ADOS-2 Total 2.68 (3.23) 12.75 (7.28) 0.00–26.00 -5.12 < .001
CELF-5 Expressive Language Index 106.32 (13.93) 81.94 (19.45) 47.00–126.00 4.19 < .001
CELF-5 Language Memory Index 106.95 (14.80) 80.62 (20.77) 45.00–128.00 4.24 < .001
CELF-5 Formulated Sentences 42.63 (4.99) 32.81 (13.48) 3.00–48.00 2.76 0.013
CELF-5 Recalling Sentences 67.68 (6.63) 46.94 (21.43) 11.00–78.00 3.73 0.002
CELF-5 Semantic Relationships 17.11 (2.66) 8.88 (7.17) 1.00–20.00 4.34 < .001
CELF Syntax Composite 0.47 (0.37) -0.55 (1.18) -2.91–0.94 3.32 0.004
CELF Semantic Composite 0.49 (0.40) -0.58 (1.07) -2.85–0.96 3.82 0.001
DAS-II Nonverbal Reasoning 107.32 (15.03) 85.44 (22.61) 40.00–133.00 3.30 0.003
1st-Order ToM 6.61 (0.50) 5.43 (1.70) 2.00–7.00 2.52 0.024
2nd-Order ToM 5.74 (0.56) 4.35 (1.49) 1.00–6.00 3.21 0.006
2nd-Order ToM Justification 5.13 (0.94) 2.77 (1.69) 0.00–6.00 4.58 < .001
Aliens Total Points 169.26 (96.26) 124.62 (64.04) 30.00–330.00 1.64 0.112
Aliens Accuracy (%) 51.18 (20.04) 41.56 (13.41) 22.50–85.00 1.69 0.101
Aliens MLU 5.50 (1.13) 5.06 (2.44) 1.82–9.86 0.66 0.516

Start of Bayesian Modeling

For all of the following models, weakly informative priors were used in combination with a gaussian distribution to regularize parameter estimates while remaining permissive of theoretically plausible effects. Given the modest sample size and correlated predictors, these priors helped stabilize estimation and reduce the influence of implausibly large parameter values without imposing strong prior assumptions on the models. All models converged, as indicated by the R-hat statistic value (i.e., R-hat = 1.00).

1st-order ToM and Age, Diagnostic Group, DAS, & CELF

Table 2. Bayesian Structural Language Models Predicting 1st-Order Theory of Mind

Table 2. Bayesian Structural Language Models Predicting First-Order Theory of Mind
Predictors Controls Only Model Controls + CELF Syntax Model Controls + CELF Semantics Model Full Model
Age 0.26 [-0.06, 0.58] 0.10 [-0.16, 0.36] -0.03 [-0.32, 0.26] -0.03 [-0.30, 0.24]
Nonverbal Cognition (DAS-II) 0.75 [0.41, 1.07] 0.35 [0.03, 0.68] 0.22 [-0.15, 0.59] 0.17 [-0.17, 0.52]
Group -0.47 [-1.05, 0.11] -0.15 [-0.67, 0.36] 0.04 [-0.53, 0.60] 0.07 [-0.48, 0.60]
Syntax Composite — 0.77 [0.42, 1.11] — 0.50 [0.08, 0.92]
Semantic Composite — — 0.91 [0.49, 1.33] 0.56 [0.05, 1.08]

Initial descriptive analyses indicated lower ToM performance among autistic participants relative to TD participants (Table 1). However, in the Bayesian models, group effects weakened substantially.

Model comparison using leave-one-out (LOO) cross-validation suggested that models incorporating CELF composite scores demonstrated substantially improved predictive performance for first-order ToM (ToM1) relative to the controls-only (i.e., age, group, DAS) model. The full model (containing both CELF Syntax & Semantics) demonstrated the strongest overall predictive fit, outperforming the controls-only model (elpd_diff = -10.4, SE = 5.4).

Both the CELF Syntax and Semantic Composites individually demonstrated strong positive associations with ToM1 performance. In the Syntax-only model, the CELF Syntax was positively associated with ToM1 performance. Similarly, in the semantic-only model, the CELF Semantic Composite was positively associated with ToM1 performance. Importantly, when syntax and semantic composites were entered simultaneously, both predictors retained positive associations with ToM1 performance, over and above age, DAS, and group.

Figure 1. CELF Syntactic & Semantic Composites positively associated w/ 1st-Order ToM performance

Figure 1. Posterior distributions of standardized regression coefficients from the combined Bayesian model predicting first-order Theory of Mind (ToM1) performance. Distributions shifted further from zero indicate stronger directional associations, with darker shading reflecting positive associations and lighter shading reflecting negative associations. Wider distributions (i.e., flattened peaks) reflect greater uncertainty in parameter estimates. Vertical dashed lines indicate zero effect.

Do the LIWC measures explain anything beyond CELF Syntax & Semantics?

Starting with CELF Syntax:

Table 3. Exploratory LIWC Models Predicting 1st-Order ToM Beyond CELF Syntax

Table 3. Exploratory LIWC Models Predicting First-Order ToM Beyond CELF Syntax
Predictors Cognitive Processes Model Insight Model Tentative Model
Syntax Composite 0.74 [0.36, 1.12] 0.76 [0.40, 1.12] 0.78 [0.43, 1.11]
Cognitive Processes 0.02 [-0.07, 0.12] — —
Insight Language — 0.02 [-0.08, 0.13] —
Tentative Language — — -0.08 [-0.29, 0.13]
All models controlled for age, DAS, and group.

Exploratory follow-up Bayesian models examined whether LIWC language features contributed predictive utility beyond structural syntax for first-order ToM performance. Across all follow-up models, the CELF Syntax Composite remained the strongest positive predictor of ToM1 performance. In contrast, Cognitive Processes, Insight, and Tentative language demonstrated little evidence of unique associations with ToM1 performance beyond syntax.

Figure 2. LIWC variables contribute little beyond CELF Syntax for 1st-Order ToM Performance

Figure 2. Posterior distributions of LIWC language predictors from exploratory Bayesian models examining contributions beyond CELF Syntax in first-order ToM performance. All models controlled for age, nonverbal cognition, diagnostic status, and the CELF Syntax Composite.

And now CELF Semantics:

Table 4. Exploratory LIWC Models Predicting 1st-Order ToM Beyond CELF Semantics

Table 4. Exploratory LIWC Models Predicting First-Order ToM Beyond CELF Semantics
Predictors Cognitive Processes Model Insight Model Tentative Model
Semantics Composite 0.88 [0.42, 1.31] 0.91 [0.47, 1.33] 0.91 [0.49, 1.33]
Cognitive Processes 0.17 [-0.11, 0.45] — —
Insight Language — 0.04 [-0.23, 0.30] —
Tentative Language — — -0.07 [-0.33, 0.19]
All models controlled for age, DAS, and group.

Exploratory follow-up Bayesian models examined whether LIWC measures explained additional variance in first-order ToM performance beyond semantic language ability. Across all models, the CELF Semantic Composite remained a strong positive predictor of ToM1 performance, whereas each individual LIWC measure showed weak and uncertain associations, with credible intervals overlapping zero.

Figure 3. LIWC variables contribute little beyond CELF Semantics for 1st-Order ToM Performance

Figure 3. Posterior distributions of LIWC language predictors from exploratory Bayesian models examining contributions beyond CELF Semantics in first-order ToM performance. All models controlled for age, nonverbal cognition, diagnostic status, and the CELF Semantics Composite.

2nd-order ToM and Age, Diagnostic Group, DAS, & CELF

Table 5. Bayesian Structural Language Models Predicting 2nd-Order ToM

Table 5. Bayesian Structural Language Models Predicting Second-Order ToM
Predictors Controls Only Model Controls + CELF Syntax Model Controls + CELF Semantics Model Full Model
Age 0.01 [-0.33, 0.34] -0.10 [-0.41, 0.22] -0.17 [-0.51, 0.17] -0.16 [-0.49, 0.16]
Nonverbal Cognition (DAS-II) 0.53 [0.19, 0.86] 0.26 [-0.08, 0.61] 0.20 [-0.19, 0.60] 0.17 [-0.21, 0.54]
Group -0.71 [-1.30, -0.08] -0.46 [-1.03, 0.11] -0.37 [-0.99, 0.24] -0.35 [-0.95, 0.25]
Syntax Composite — 0.60 [0.24, 0.96] — 0.45 [-0.00, 0.90]
Semantic Composite — — 0.61 [0.17, 1.02] 0.29 [-0.24, 0.82]
All models controlled for age, DAS, and group.

Model comparisons suggested that the CELF Syntax model showed the best estimated predictive performance for ToM2 performance (elpd_diff = -4.7, SE = 4.4), with the CELF Syntax Composite positively associated with ToM2 performance. The CELF Semantic model also demonstrated improved predictive performance relative to the controls-only model, although overall fit was somewhat weaker than the syntax model (elpd_diff = -1.6, SE = 2.5). When syntax and semantic composites were entered simultaneously (Model 4), the CELF Syntax composite remained more positively associated with ToM2 performance than the semantic composite. Importantly, however, the model produced minimal improvement in predictive performance (elpd_diff = -0.2, SE = 0.9).

Figure 4. CELF Syntax strongly predicts 2nd-Order ToM Performance

Figure 3. Posterior distributions of standardized regression coefficients from the combined Bayesian model predicting second-order ToM performance.

Do the LIWC measures explain anything beyond CELF Syntax?

Table 6. Exploratory LIWC Models Predicting Second-Order ToM Beyond CELF Syntax

Table 6. Exploratory LIWC Models Predicting Second-Order ToM Beyond CELF Syntax
Predictors Cognitive Processes Model Insight Model Tentative Model
Syntax Composite 0.47 [0.11, 0.83] 0.60 [0.24, 0.96] 0.61 [0.26, 0.95]
Cognitive Processes 0.12 [0.00, 0.23] — —
Insight Language — 0.02 [-0.10, 0.15] —
Tentative Language — — 0.21 [-0.04, 0.45]
All models controlled for age, DAS, and group.

Exploratory follow-up Bayesian models examined whether LIWC measures contributed predictive utility beyond CELF Syntax. Model comparison suggested that the model including Cognitive Processes language demonstrated the strongest predictive performance among the follow-up models. Relative to the syntax-only model (Model 2), adding Cognitive Processes language modestly improved out-of-sample predictive fit (elpd_diff = 1.4, SE = 1.3). Within this model, both the CELF Syntax Composite and Cognitive Processes language demonstrated positive associations with ToM performance. Tentative and Insight language contributed little beyond syntax alone.

Figure 5. Cognitive Processes language shows unique contributions beyond syntax in 2nd-Order ToM performance

Figure 4. Posterior distributions of LIWC language predictors from exploratory Bayesian models examining contributions beyond CELF Syntax in second-order ToM performance. All models controlled for age, nonverbal cognition, diagnostic status, and the CELF Syntax Composite.

Overall Summary

Group-Level Findings

Autistic participants demonstrated weaker ToM performance than non-autistic participants in the descriptive statistics. However, across both 1st- and 2nd-order ToM analyses, once age, nonverbal cognition, and especially language abilities were statistically accounted for, diagnostic group differences weakened substantially and often no longer showed clear independent associations with ToM performance.

1st-Order ToM (ToM1)

CELF Composites. Both CELF Syntax and Semantic composites were strongly associated with stronger ToM1 performance, even after accounting for age, nonverbal reasoning, and diagnostic group. Importantly, when Syntax and Semantics were examined together, both continued to show positive relationships with ToM1 performance.

LIWC Measures. The exploratory LIWC analyses suggested a more limited role for these naturalistic language features. Participants’ use of Cognitive Processes, Tentative, and Insight language explained little variation in ToM performance, particularly once broader language abilities were taken into account. In contrast, the CELF Syntax and Semantic composites consistently remained the strongest predictors of ToM1 performance.

2nd-Order ToM (ToM2)

CELF Composites. When examined separately, both CELF Syntax and Semantic composites were positively associated with ToM2 performance. However, when both language measures were included in the same model, Syntax showed the more stable relationship with ToM2 performance, whereas the association with Semantics became notably weaker.

LIWC Measures. The LIWC findings for ToM2 were somewhat mixed but still relatively modest overall. Most LIWC measures again showed weak or uncertain associations once broader language abilities were included in the models. However, there was limited evidence that greater use of Cognitive Processes language (i.e., words related to thinking or reasoning) may have contributed some additional predictive value beyond Syntax for ToM2 performance. Even so, syntactic language remained the strongest and most consistent predictor across the models.