Figure 7 illustrates the rate of correct responses provided by children and adults in lexical and sluiced indirect wh-questions according to the Extraction Site of the wh-element. Overall, all participants performed above-chance, with accuracy increasing with age. Compared to their younger peers, older children pattern more closely to adults, who performed at ceiling in all conditions. Across all ages, children were more accurate in lexicalized than in sluiced indirect wh-questions. They were also consistently more accurate with subject- than object-extracted sentences, both in sluices (70.8% vs. 53.5% at age 3, 80.1% vs. 59% at age 4, 85.3% vs. 70.7% at age 5) and in lexicalized interrogatives (84.7% vs. 68.8% at age 3, 90.4% vs. 72.4% at age 4, 91.3% vs. 79.3% at age 5).

To statistically verify the results observed in Figure 7, participants’ responses were fitted to linear mixed-effects logistic regression models (GLMMs) in R (R Core Team 2023). The analyses were conducted using the glmer function from the lme4 package (Bates et al. 2015), with the specification of the logit link transformation, the binomial family and the bobyqa optimizer to enhance convergence stability during model fitting. Score was analyzed as a binary dependent variable indicating response accuracy (correct, incorrect). As for predictors, we included Age (4 levels: Age 3, Age 4, Age 5, Adults), Realization (2 levels: Sluiced, Lexicalized), Extraction Site (2 levels: Subject, Object), Scenario (2 levels: Yes-Scenario, No-Scenario), also specifying their interactions. Factor levels were sum-contrast coded to aid interpretability. We also included random intercepts for Participants and Items to account for baseline variability in response accuracy across individuals and test items. Additional contrasts between factor levels were further checked using post-hoc pairwise comparisons with Tukey’s correction. The estimates of the fixed effects of the model that best fitted our data are reported in Table 3.11

The model revealed a significant main effect of the predictor Age, as well as significant two-way interactions Extraction Site*Age and Realization*Age in the contrast Children vs. Adults.

As expected, the first Age contrast (Children vs. Adults) indicates that, overall, children did not perform in an adult-like fashion yet (β = 3.64, SE = 0.59, z = 6.14, p < .001). The other Age contrasts show that 3-year-olds significantly differ from 4-year-olds (β = 0.74, SE = 0.24, z = 3.11, p = .002), and 4-year-olds significantly differ from 5-year-olds (β = 0.75, SE = 0.25, z = 3.01, p = .003), suggesting that, altogether, children’s performance gets progressively better with age.

The difference between children and adults clearly arises with respect to the lexicalization of the wh-embedded interrogative sentences, as indicated by the two-way interaction Realization*Age (Children vs. Adults) (β = –1.98, SE = 0.82, z = –2.43, p = .015): while children performed better with lexicalized than sluiced sentences, adults were at ceiling in both.

Turning to our second research question, namely the subject>object asymmetry, our model statistically confirms the descriptive overview reported in Figure 7. A significant Extraction Site*Age (Children vs. Adults) interaction (β = 3.14, SE = 1.07, z = 2.95, p = .003), shows that adults performed at ceiling for both subject- and object-extracted stimuli, while children responded more accurately to the former than the latter.12 Post-hoc comparisons showed that this asymmetry is consistent within each age group of children (age 3: β = 0.87, SE = 0.21, z = 4.24, p < .001; age 4: β = 1.18, SE = 0.22, z = 5.46, p < .001; age 5: β = 0.98, SE = 0.24, z = 4.11, p < .001).

Notably, model comparisons showed that the Extraction Site*Realization interaction did not significantly improve model fit (χ²(1) = 0.215, p = .643), as expected under our theoretical assumption. Indeed, since lexical and sluiced interrogative questions involve syntactically equivalent structures, the asymmetry was predicted to arise in both sentences.

To verify the consistency of our statistical results within participants, we conducted two separate individual analyses. The first analysis investigated the Realization*Age interaction, aiming to determine whether the lexicalization of wh-embedded interrogative sentences constrains each child’s acquisition path. For each participant, we counted the number of correct responses in the two Realization conditions (Sluiced, Lexicalized). A structure type was considered acquired if the participant provided at least 8/12 correct responses, as indicated by a binomial test. Response patterns could fall within one of the following four: (i) No acquisition of indirect wh-questions, comprising children who have not yet acquired either lexicalized or sluiced sentences (Pattern 1);13 (ii) Acquisition of lexicalized interrogatives, encompassing children who have acquired lexicalized wh-questions but not their elliptical counterparts (Pattern 2), and (iii) Acquisition of both lexicalized and sluiced interrogatives, including children who handle both sentence types (Pattern 3); (iv) Acquisition of sluiced questions but not their lexicalized counterparts (Pattern 4).

As shown in Table 4, the distribution of these patterns across age groups is not random (Fisher’s Exact Test with Monte Carlo method, p = .049)14, suggesting an acquisition path roughly correlated with age. Numerically, Pattern 1 primarily comprises some residual younger children, and Pattern 2 is mostly found among 3- and 4-year-olds. Pattern 3 mainly encompasses 5-year-old children (FET-MC, p = .037, Bonferroni-corrected), along with all adult participants. Crucially, virtually no children showed a pattern consistent with understanding sluices but not lexicalized wh-questions (Pattern 4). This is consistent with theoretical expectations since sluices presuppose the structure of lexicalized wh-questions and require additional operations.15

Hence, the comprehension of indirect wh-questions proceeds in a three-step fashion: (i) STAGE I, in which children do not comprehend indirect wh-questions, (ii) STAGE II, in which they comprehend lexicalized but not sluiced questions, (iii) STAGE III, in which they successfully comprehend both structures.

The second individual analysis examined our second statistical result, i.e. Extraction Site*Age, within participants. We counted the number of correct responses for subject- and object-extracted interrogatives, collapsing both sluiced and lexicalized interrogatives, thereby mirroring the interaction detected by the model. A threshold of 8/12 correct responses was maintained to define a structure type as acquired. Again, three response patterns emerged: (i) No acquisition of indirect wh-questions, comprising children who have not acquired either subject- or object-extracted sentences (Pattern 1), (ii) Acquisition of subject-extracted sentences, encompassing children who comprehended subject-extracted sentences but not object-extracted ones (Pattern 2), and (iii) Acquisition of both subject-extracted and object-extracted sentences, including children who have acquired both sentence types (Pattern 3). Crucially, in line with the predictions of fRM, no participant instantiated the reverse of Pattern 3, i.e. Pattern 4 —that is, having acquired object-extracted but not subject-extracted sentences.

As shown in Table 5, the distribution of patterns across age groups suggests a possible age-related acquisition trajectory. However, this variation did not reach statistical significance (FET-MC, p = .12), possibly because most children had already acquired subject-extracted sentences, and a finer-grained analysis of the acquisition of object dependencies may be required (see Experiments 2 and 3 in Sections 3.2 and 3.3). Descriptively, Pattern 1 occurs only among some 3-year-olds, whereas Pattern 2 is mostly instantiated by 3- and 4-year-olds. Pattern 3, besides including all adult participants, is primarily observed in 5-year-olds.

Thus, the comprehension of indirect wh-questions proceeds in a three-step fashion: (i) STAGE I, in which children do not comprehend indirect wh-questions; (ii) STAGE II, in which children comprehend subject but not object wh-questions, (iii) STAGE III, in which children comprehend both subject and object wh-questions.

Figure 12 illustrates our results. As in Experiment 1, participants performed above chance overall, with accuracy increasing progressively with age: older children provided more target-like responses and patterned more closely with adults. Consistent with the previous experiment, subject sluices were well comprehended at all ages, whereas object sluices remained challenging, even for older children.

However, Figure 12 also shows that object sluices are better comprehended when involving the bare wh-element chi (i.e., mismatch conditions) than the lexically restricted quale+NP (i.e., match conditions) (69.4% vs. 56.9% at age 3, 80.2% vs. 61.7% at age 4, 85.9% vs. 71.4% at age 5), while no such difference appeared in subject sluices (73.6% vs. 78.5% at age 3, 87.7% vs. 87% at age 4, 94.3% vs. 92.7% at age 5).

To statistically verify the results observed in Figure 12, participants’ responses were analyzed using linear mixed-effects logistic regression models (GLMMs) in R (R Core Team 2023). The analyses were conducted using the same statistical tools outlined for Experiment 1. As for predictors, we included Age (4 levels: Age 3, Age 4, Age 5, Adults), Extraction Site (2 levels: Subject, Object), Lexical Restriction (2 levels: Match, Mismatch), Scenario (2 levels: Yes-Scenario, No-Scenario), also specifying their interactions. Random intercepts were also specified for Participants and Items, to model baseline differences in accuracy of responses. The estimates of the fixed effects of the model that best fitted our data are reported in Table 7.

The model revealed main significant effects of the three predictors Age, Extraction Site and Lexical Restriction, as well as the two-way interaction Extraction Site*Lexical Restriction.

As in Experiment 1, the first Age contrast (Children vs. Adults) indicates that, overall, children gave significantly fewer correct responses than adults, who performed at ceiling in all conditions (β = 4.59, SE = 0.72, z = 6.41, p < .001). The progressively increasing accuracy across age groups shown in Figure 12 is further confirmed by the remaining Age contrasts: 3-year-olds differed from 4-year-olds (β = 1.14, SE = 0.23, z = 4.88, p < .001), and 4-year-olds differed from 5-year-olds (β = 1.14, SE = 0.23, z = 4.84, p < .001).

The significant effect of Extraction Site indicates higher accuracy in subject than in object sluices (β = –0.96, SE = 0.19, z = –5.05, p < .001), thus replicating the asymmetry observed previously.

Notably, the model also detected a significant main effect of Lexical Restriction (β = 0.43, SE = 0.19, z = 2.25, p = .024), indicating that preschoolers performed better on sluices involving chi (i.e., mismatch condition) than on those with quale+NP (i.e., match condition). Moreover, the significant interaction between Extraction Site and Lexical Restriction shows that accuracy was higher in object sluices with bare wh-elements than in those with lexically restricted ones (β = 0.91, SE = 0.38, z = 2.41, p = .016). Importantly, no such difference emerged for subject sluices, as further checked by post-hoc Tukey’s tests (β = 0.03, SE = 0.28, z = 0.11, p = 1.0).

To verify the consistency of our results within participants, we conducted an individual analysis consisting of two subparts.

First, as we did for Experiment 1, we examined the subject>object asymmetry detected by the statistical model within each individual’s acquisition path. Specifically, we counted the number of correct responses across the two conditions related to the variable Extraction Site (Subject, Object). To define a structure type as acquired, we fixed a criterion of 8/12 correct responses for the given structure, as indicated by a binomial test. We identified three different response patterns: (i) No acquisition of sluicing, comprising children who have not yet acquired either subject or object sluices (Pattern 1); (ii) Acquisition of subject sluices, including children who correctly comprehended subject sluices but still struggled with object sluices (Pattern 2); (iii) Acquisition of both subject and object sluices, including children who comprehended both sentence types (Pattern 3). Notably, the fourth logical possibility, i.e. having acquired object sluices but not subject sluices (Pattern 4), was not instantiated in our dataset.

Since lexical mismatch improved children’s comprehension of object sluices (see Section 3.2.3.1), we further included this variable in Pattern 3 (the only one in which object sluices were acquired). In this sub-part of the individual analysis, with 6 stimuli for each condition, acquisition was defined as at least 5/6 correct responses, as indicated by a binomial test. Accordingly, Pattern 3 was split into two sub-groups: (i) Acquisition of object sluices involving a lexical mismatch configuration, comprising children who correctly comprehended object sluices involving the bare wh-element chi, but not those involving the lexically restricted wh-element quale+NP (Pattern 3A); (ii) Acquisition of both object sluices involving a mismatch configuration and those involving a match configuration, encompassing children who performed above-chance in both object sluices with chi and those with quale+NP.

In addition to Patterns 3A and 3B, two other possibilities arise: (i) the reverse of Pattern 3A, i.e., having acquired object sluices with a match configuration but not those with a mismatch configuration (henceforth, Pattern 3C); and (ii) the NAC (Nearly-Above Chance) Pattern, i.e. scoring exactly 4/6 on both object sluices with lexical match and those with lexical mismatch, yielding a total of 8/12 correct responses in object sluices overall. While no participant displayed Pattern 3C, as predicted by fRM, four 3-year-olds and two 4-year-olds exhibited the NAC Pattern.

Leaving the NAC Pattern aside, as being rather uninformative, and comparing the other patterns, we can trace an acquisition path. Table 8 reports participants’ distribution across patterns, suggesting a four-stage acquisition path correlated with age (Fisher’s Exact Test with Monte Carlo method, p < .001).16

Specifically, only a few 3-year-olds exhibit Pattern 1 (FET-MC, p = .014, Bonferroni-corrected), whereas Pattern 3B, at the opposite pole of the acquisition path, primarily appears among 5-year-olds (FET-MC, p = .001, Bonferroni-corrected) alongside all the adults. Intermediate Patterns, i.e. Pattern 2 and Pattern 3A, are instantiated by children of all age groups, though the former is mostly numerically represented by those aged 3 and 4, and the latter by those aged 4 and 5. Thus, the comprehension of sluicing proceeds in a four-step fashion: (i) STAGE I, where children fail to comprehend both subject and object sluices; (ii) STAGE II, where children comprehend subject sluices but not object sluices; (iii) STAGE III, where children comprehend subject sluices and object sluices involving a lexical mismatch; (iv) STAGE IV, in which children comprehend subject sluices as well as object sluices, both with and without a lexical mismatch.

Our results are reported in Figure 17. As in Experiments 1 and 2, adults performed at ceiling in all conditions, and children showed an overall above-chance performance, with accuracy increasing with age.

Figure 17 further suggests the presence of a subject>object asymmetry in the sluices, mirroring the trend of the previous experiments: altogether, children gave more correct responses for subject sluices than for object sluices. Nevertheless, such an asymmetry seems to be mitigated in the presence of Number mismatch in object sluices. Specifically, Italian preschoolers performed better in object sluices with Number Mismatch than in those with Number Match (61.3% vs. 54% at age 3, 76.9% vs. 60.3% at age 4, 90.2% vs. 76.4% at age 5). In stark contrast, the proportion of correct responses in subject sluices is similar across the two different Number conditions (72.7% vs. 74% at age 3, 81.4% vs. 84.6% at age 4, 92.5% vs. 89.7% at age 5).

To statistically verify the results observed in Figure 17, participants’ responses were analyzed using linear mixed-effects logistic regression models (GLMMs) in R (R Core Team 2023). The analyses were conducted using the same statistical tools outlined for Experiments 1 and 2. As for predictors, we included Age (4 levels: Age 3, Age 4, Age 5, Adults), Extraction Site (2 levels: Subject, Object), Number (2 levels: Match, Mismatch), Scenario (2 levels: Yes-Scenario, No-Scenario), also specifying their interactions. We also specified random intercepts for Participants and Items. The estimates of the fixed effects of the model that best fitted our data are reported in Table 10.

The model revealed significant main effects of Age, Extraction Site and Number, as well as the two-way interaction Extraction Site*Number.

The first Age contrast (Children vs. Adults) shows that children’s correct responses are significantly lower than adults’, who performed at ceiling in all conditions (β = 4.18, SE = 0.60, z = 7.00, p < .001). As in previous experiments, the proportion of target responses increase with age: 3-year-olds differ from 4-year-olds (β = 1.32, SE = 0.25, z = 5.30, p < .001), and 4-year-olds differ from 5-year-olds (β = 1.50, SE = 0.26, z = 5.79, p < .001).

Additionally, the main effect of Extraction Site confirms the asymmetry depicted in Figure 17: subject sluices are better comprehended than object sluices, as shown by a significantly lower proportion of correct answers for the latter (β = –0.78, SE = 0.16, z = –4.84, p < .001).

Eventually, the model also detected a significant main effect of Number (β = 0.33, SE = 0.16, z = 2.07, p = .038), indicating that Italian preschool children provided a higher number of correct responses in the sluices with Number mismatch than in those with Number match. In this respect, the significant Extraction Site*Number interaction further indicates that children responded more accurately to object sluices with Number mismatch (β = 0.73, SE = 0.32, z = 2.28, p = .023), while the accuracy rate in subject sluices is not affected by the Number condition, as further confirmed by post-hoc Tukey’s tests (β = 0.03, SE = 0.24, z = 0.14, p = .99).

To verify the consistency of our statistical results within participants, an individual analysis was conducted, consisting of two subparts as in Experiment 2. First, we examined the effect of Extraction Site, focusing on whether the subject>object asymmetry holds within each individual’s acquisition path. To this end, we counted the number of correct responses across the two conditions related to the variable Extraction Site (Subject, Object). With a threshold of 8/12 correct responses to define a structure as acquired, participants were classified into three response patterns: (i) No acquisition of sluicing, comprising children who have not yet acquired either subject or object sluices (Pattern 1), (ii) Acquisition of subject sluices, including children who correctly comprehend sluices but still exhibit difficulties with object sluices (Pattern 2); (iii) Acquisition of both subject and object sluices, including children who manage both sentence types (Pattern 3). Notably, the fourth logical possibility, i.e., having acquired object sluices but not subject sluices (Pattern 4), was again not instantiated by any participant.

The second subpart of the analysis focused on the two-way interaction Extraction Site*Number, splitting Pattern 3 into two sub-conditions based on the Number manipulation. In this sub-analysis, since there were 6 stimuli for each condition, a structure type was defined as acquired with 5/6 correct responses, as indicated by a binomial test. Accordingly, two sub-groups within Pattern 3 emerged: (i) Acquisition of object sluices involving a number mismatch configuration, including children who correctly comprehended object sluices involving a plural subject DP, but not those in which the subject DP, being singular, match in the Number specification with the wh-object (Pattern 3A); (ii) Acquisition of both object sluices involving a Number mismatch configuration and those involving a Number match configuration, encompassing children who performed above-chance in both types of object sluices, that is not only when the two DPs involved in the stimuli mismatched in Number, but also when they both were specified as singular with respect to the Number feature (Pattern 3B).

Once again, the two other possibilities besides Pattern 3A and Pattern 3B would have been (i) the reverse of Pattern 3A, i.e. having acquired object sluices with a Number match but not those with a Number mismatch, henceforth Pattern 3C; (ii) the NAC (Nearly-Above Chance) Pattern, where participants scored exactly 4/6 in both object sluices with Number match and in those with Number mismatch, yielding 8/12 correct responses overall. While no participants exhibited Pattern 3C, as predicted by fRM, two 3-year-olds, one 4-year-old, and one 5-year-old exhibited the NAC Pattern.

The stepwise acquisition path that emerges putting together the different patterns identified, roughly correlated with age (Fisher’s Exact Test with Monte Carlo method, p < .001)18, is depicted in Table 11.19

Numerically, Pattern 1 is instantiated only by some residual younger children, while the opposite pole of the path, i.e. Pattern 3B, is mostly found among the older ones, as well as all the adults. As for the intermediate patterns, Pattern 2 is mostly exhibited by children aged 3 and 4 (FET-MC, p = .003, Bonferroni-corrected), whereas Pattern 3A is mostly instantiated by those aged 4 and 5 (FET-MC, p = .032, Bonferroni-corrected). This suggests a four-stage acquisition path: (i) STAGE I, where children do not comprehend either subject or object sluices; (ii) STAGE II, where children comprehend subject sluices but not object sluices; (iii) STAGE III, where children comprehend both subject sluices and object sluices involving a Number mismatch; (iv) STAGE IV, where children comprehend both subject sluices and object sluices, both with and without a Number mismatch.