3.1 Code-switching as evidence for linguistic theory

It has long been recognized that intrasentential code-switching is systematic and principle-governed, like language itself (Gingràs 1974; Timm 1975; Pfaff 1979). While even early syntactic work recognized that this systematicity can be accounted for without CS-specific rules or constraints (e.g., Pfaff 1979), MacSwan (1999), taking advantage of the possibilities afforded by the Minimalist Program (Chomsky 1995), was the first to implement such an account. In this view, code-switched sentences are generated using the same mechanisms as monolingual sentences, and any apparent constraints emerge from the interaction of the two languages’ lexical items, with their respective syntactic features, in the syntactic derivation.⁵ In other words, code-switching is simply one expression of a speaker’s linguistic competence, resulting from normal syntactic assumptions.

One important consequence of this approach to code-switching is that we can use it as a tool of linguistic analysis. If we assume that cross-linguistic syntactic variation stems from differences in the formal features on functional heads, we can take advantage of these differences to isolate those features (González-Vilbazo & López 2012; González-Vilbazo et al. 2013). Each language’s functional heads will reflect a particular combination of formal features, so we can use the language of the heads as an index to that feature combination. This, in turn, allows us to correlate changes in acceptability with the presence of particular functional heads (and their features).

3.2 Hypotheses on word order from code-switching research

Taking the above approach, code-switching researchers have also made proposals about how word order is determined in the extended verbal projection, which this section presents. First, though, it is worth noting that some (e.g., González-Vilbazo 2005) have contended that switches between C and T are not possible in general. However, such switches are accepted in judgments (Ebert & Hoot 2018; Sande 2018; Vanden Wyngaerd 2020) and attested (though rare) in corpora (Sankoff & Poplack 1981; Mahootian & Santorini 1996; Callahan 2004; Hebblethwaite 2007; Jansen, Müller & Müller 2012), as in (8–10), so we do not consider that hypothesis further.

Assuming such switches are possible, code-switching researchers have arrived at the same divide that studies based on monolingual data (see §2) revealed: some point to T as the locus of features that determine word order and some point to C.

First, MacSwan (2004) argues the features of T determine subject placement in code-switching. He presents corpus evidence from code-switching between VSO languages and SVO, showing that when the T head is from a VSO language, the subject appears post-verbally, as in (11) and (12), concluding T determines subject position, just as in monolingual approaches that assume the EPP is a feature of T (see also MacSwan 2013).

On the other hand, some researchers point to C as the likely determinant of word order in TP. González-Vilbazo and López (2012) propose the Phase Head Hypothesis, which specifies that “the phase head determines grammatical properties of its complement,” including word order. They note that (13), with Spanish word order and prosody in the lower clause, is possible, while (14), with German word order and prosody, is not.

In a pilot study on Spanish/German CS, González-Vilbazo and López (2013) found support for the notion that C determines word order within TP, and two studies that explicitly considered subject placement and omission in the code-switching of bilingual children arrive at the same conclusion: both Cantone (2007) and Jansen et al. (2012) contend that, in their corpus data from Germanic/Romance bilingual children, it is the language of C that determines the word order within TP, including subject placement.

In summary, two opposed hypotheses find support in both monolingual and code-switching evidence. Experimental approaches can help clarify the empirical picture. Although there has been some experimental work on subject position in Spanish (e.g., Goodall 2004; 2010), for the most part the facts about Spanish have not been in dispute, only the analysis. Thus, we next identify three code-switching experiments useful for situating our project.

3.3 Experimental evidence from code-switching

First, we conducted a previous version of the study we are reporting here, which we consider a pilot for the present work (Ebert & Hoot 2018). That experiment had a similar design, testing pre- and post-verbal subjects in Spanish/English code-switching with an acceptability judgment task when C was English and T Spanish, as in (15), and when C was Spanish and T English, as in (16).

We found that post-verbal subjects were rated lower than pre-verbal ones across the board and that sentences with English C and Spanish T, like (15), were always better than the reverse. However, we failed to find an interaction between the two factors, suggesting that the acceptability of post-verbal subjects did not differ according to the presence of either C or T. This finding led us to suggest that perhaps both C and T need to be in Spanish to license post-verbal subjects. Yet our previous conclusion suffered from several limitations, which motivate the design of the present study. First, our sample was small: 16 participants. Here we have 36. Second, we did not have the full factorial design, including items with C and T in the same language, and so our conclusion was necessarily somewhat tentative: we had to argue from a lack of effect and could not directly show the effect of C and T together. We include the full factorial design here. The same was true for our monolingual stimuli, for which we did not previously have the full paradigm, which we have also rectified here. Overall, our previous work points toward a promising conclusion, but it needed additional empirical verification with improved methods, which is a chief motivation for the present study.

Second, Sande (2018) investigated null subjects—which, as we have discussed, are clearly tied to pre-/post-verbal subjects—in Spanish/English code-switching using an acceptability judgment task. Like our work, Sande’s experiment focused on the role of C and T in licensing pro-drop, so test sentences always presented a switch between C and T, although Sande identified the language of C as the language of a wh-word in CP.

She found that null subjects were rated significantly lower than overt subjects when C was English and T was Spanish, as in (17). When C was Spanish and T was English (18), null subjects were also rated slightly lower, although the difference failed to reach statistical significance (perhaps because, much like what we found, switches in this direction were rated much worse overall). Sande proposes that properties of C and T are both required to license null subjects; neither can do it alone. She contends, following González-Vilbazo (2012), that the T head possesses “strong φ-features” that provide syntactic licensing for null subjects, while the C head possesses features that “allow for the linking of [null pro] to its true referent in the discourse” to provide the necessary discursive licensing (Sande 2018: 139–140). Both are therefore needed to allow a null subject in code-switching. Like our own, Sande’s study has some limitations. First, she did not explicitly test cases where C and T were both in the same language, which is needed to fully support the hypothesis that both heads together can license null subjects. Second, as mentioned above, the crucial comparison of overt/null subjects with Spanish C and English T failed to reach statistical significance and was very similar in the raw ratings. Finally, she determined the language of C by assuming it was always the same as the language of a wh-word, which, although reasonable, does not identify the language of C as unambiguously as a design including overt complementizers would. Nonetheless, that Sande’s results from a related but distinct phenomenon point in the same direction as our own pilot study provides an additional reason to expect that an explanation in terms of C and T together is on the right track.

Finally, Vanden Wyngaerd (2020) conducted an acceptability judgment study investigating V2 word order in Dutch/English code-switching. Traditional accounts of V2 claim it is triggered by a feature on C that causes the verb to raise from T, with a few proposals instead pinning it on features of T. Vanden Wyngaerd found that in code-switching the language of the verb determined the word order: when the verb was in Dutch, participants preferred V2 orders (19), and when the verb was in English, participants preferred non-V2 orders (20).

Although her code-switching data points toward an increased role of T in determining V2 and away from the traditional explanations putting the onus on C, she nonetheless notes some independent evidence against accounts focused entirely on T, leading her to conclude that perhaps some type of Feature Inheritance between C and T is the ultimate explanation, which is one of the hypotheses we consider as well.

6.1 Interpreting judgment data

Interpreting sentence acceptability judgments is not as straightforward as it might appear. Linguistic theory is interested in grammaticality—a binary notion of whether a given sentence can be generated by individuals’ grammars—while judgment experiments provide evidence of acceptability—a gradient notion of subjective “naturalness.” Acceptability is a complex psychological construct, which is affected by many extra-grammatical factors and does not directly map to grammaticality (Cowart 1997; Schütze & Sprouse 2013). For instance, even if they are perfectly grammatical, we might expect post-verbal subjects will always be rated lower than pre-verbal subjects, because they depart from the canonical word order, are less frequent, and are more marked than simple SVO in active sentences with transitive predicates. The same is true of all code-switching between C and T, which we know to be possible but rare (see §3), and which we thus expect to reduce acceptability.

Schütze and Sprouse (2013), among others, advocate for factorial designs as a strategy for addressing the confounds inherent to acceptability judgments. In such a design, the factor of interest is crossed with another factor, producing a 2×2 grid, which allows researchers to isolate the effect of interest by looking at how the relevant effect changes between the two levels of the other factor. Sprouse et al. (2016) call these super-additive effects, and they are revealed by an interaction in the statistical analysis. For example, we expect post-verbal subjects to produce a decrease in acceptability for extra-syntactic reasons. If, however, Spanish T has the syntactic features which make post-verbal subjects acceptable, then post-verbal subjects should be more acceptable than expected when the sentence has Spanish T, i.e. there should be an interaction between the language of T and the position of the subject, which is the super-additive effect. Likewise, if C houses the relevant syntactic features, Spanish C should produce a super-additive effect, and if some of the necessary features are present on C and some on T, we expect a super-additive effect only when both C and T are in Spanish. Another way of understanding the factorial logic is: we expect post-verbal subjects to produce a consistent decrement in acceptability, but the acceptability decrease for post-verbal subjects will be ameliorated under certain conditions, and it is that amelioration that points to the presence of syntactic features permitting post-verbal subjects.

Turning to the details of our analysis, Schütze and Sprouse (2013) argue that all judgment data should be z-score transformed to reduce scale bias and skew, so we use z-scores for the statistical analysis. At the same time, we also recognize that the raw ratings can be informative. Therefore, we report both the z-scores and the raw ratings.

For the statistical analysis, we chose to conduct a linear mixed-effects model (LMM) for each 2×2 factorial design within each of the tasks. For the code-switching experiment, which had a 2×2×2 design, we chose to carry out a series of 2×2 tests rather than a single three-factor model because the interpretation is clearer. In all cases, the dependent variable was the z-score of the ratings, and the fixed effects included in the model were those determined by the experiment design: all the independent variables and their interactions. The random effects structure in each case was the maximal structure that converged for that set of data, following Barr et al. (2013).

We carried out the LMMs using SPSS’s MIXED procedure, and in what follows we report the results of the Type III test of fixed effects, which produces an F statistic. Because what we are interested in is whether there is a super-additive effect, our outcome of interest is always the interaction between Subject Position and the other factors. For the sake of completeness, we report the regression coefficients (beta), but focus mainly on the F statistic and its accompanying p value, setting alpha at .05.⁹ Whenever possible, we also report Bonferroni-corrected post hoc pairwise comparisons to examine differences between conditions, and, when reporting the z-score results for each condition, we report the estimated marginal means generated to make those pairwise comparisons (rather than raw mean z-scores).

Prior to data collection, we carried out a power analysis via simulation, adapted from Lane and Hennes (2018). The challenge with any power analysis is defining a priori what effect sizes are meaningful; one of the best ways to make that determination is to base it on previous research or pilot studies. We chose to look to our previous work on this topic (Ebert & Hoot 2018) to define minimal effect sizes for main effects and interactions of the statistical tests we intended to carry out, as well as to make reasonable assumptions about the amount of variance we expected to find in our data. Under these reasonable assumptions we found that with a sample size of n = 20, we would have 80% power or greater for all our planned statistical tests. We are thus confident that our sample size (n = 36) represents an experiment with acceptable power.

6.2 Monolingual English results

Recall from §5.3.1 that the English monolingual task had a 2×2 factorial design, with Type (Declarative/Question), Subject Position (Pre-/Post-verbal), and their interaction as fixed factors. The maximal random effects structure that converged included random intercepts by subject and item and a by-item random slope over the Type×Subject Position interaction. A Type III test of fixed effects yielded a significant effect for Subject Position (β = 1.62, F(1,40) = 687.6, p < .001) and the interaction of Type×Subject Position (β = 0.59, F(1,46) = 22.1, p < .001), but not for Type (β = 0.07, F(1,46) = 1.3, p = .258). Bonferroni-corrected post hoc pairwise comparisons revealed that for both sentence types post-verbal subjects were rated lower than pre-verbal (both p < .001); when subjects were pre-verbal, declarative sentences were significantly better than questions (p < .001), while when subjects were post-verbal, questions were better than declaratives (p = .016). Table 7 presents the results by condition; Figure 1 visually represents the raw ratings and Figure 2 the z-scores.

Figure 1

Mean raw ratings by sentence type and subject position, monolingual English.¹⁰

Figure 2

Estimated marginal mean z-scores of ratings by sentence type and subject position, monolingual English.

These results show a clear acceptability penalty for post-verbal subjects in English for questions and declarative sentences, as expected. The interaction between the two factors indicates, surprisingly, a small amelioration of that penalty in questions, when compared to declaratives. Given that questions are, as expected, nearly 1 point lower in acceptability even with pre-verbal subjects and that post-verbal subjects are rated near the bottom of the scale for both sentence types, at least part of the unexpected amelioration may therefore be a floor effect (i.e., with post-verbal subjects, the questions cannot be a point lower than the declaratives, because the scale ends). Crucially, the difference in raw ratings between pre- and post-verbal subjects (near the ceiling and floor of the scale, respectively) makes clear that these participants have English grammars of the expected sort, with a strong penalty for post-verbal subjects.

6.3 Monolingual Spanish results

The Spanish monolingual task had the same 2×2 factorial design as the English task; the maximal random effects structure that converged was also the same. The Type III test of fixed effects yielded a significant effect for Subject Position (β = 0.27, F(1,26) = 17.2, p < .001) but not for Type (β = 0.12, F(1,14) = 2.3, p = .154), as well as a marginal interaction for Subject Position×Type (β = 0.26, F(1,26) = 4.1, p = .054). A single post hoc pairwise comparison by Subject Position revealed that post-verbal subjects were significantly worse than pre-verbal overall (p < .001). Table 8 presents the results by condition; Figure 3 visually represents the raw ratings and Figure 4 the z-scores.

Figure 3

Mean raw ratings by sentence type and subject position, monolingual Spanish.

Figure 4

Estimated marginal mean z-scores of ratings by sentence type and subject position, monolingual Spanish.

The Spanish results show that post-verbal subjects were penalized slightly, which is to be expected even if post-verbal subjects are entirely grammatical.¹¹ Like with the English results, the raw ratings help clarify the picture here: despite the expected decrement in acceptability of post-verbal subjects, the ratings are near the top of the scale and decrease appreciably less than the corresponding English stimuli when compared to pre-verbal subjects. We take these results to indicate that these speakers have grammars in Spanish that permit post-verbal subjects, while their grammars in English do not.

6.4 Code-switching results

As explained in §6.1, we chose to carry out a series of 2×2 statistical tests for the sake of clarity of interpretation. For completeness and transparency, though, we also provide a table with the raw ratings for the eight conditions together and a graph of the distribution of each answer on the scale for each condition in Supplementary File 1.

6.4.1 Code-switching results: English C

Let us first examine cases with English C that held constant and compare two fixed factors—Language of T (Spanish/English) and Subject Position (Pre-/Post-verbal)—plus their interaction, presented in Table 9.

Table 9

Code-switching results: English C.

Condition	C	T	Subject Position	Example	Raw mean rating	Estimated marginal mean z-score of rating
Th-En-Pre	That	Eng	Pre	Qué piensa tu padre that your brothers have bought at the store?	5.13	0.29
Th-En-Pos	That	Eng	Post	Qué piensa tu padre that have bought your brothers at the store?	2.97	–0.64
Th-Sp-Pre	That	Span	Pre	What does your dad think that tus hermanos han comprado en la tienda?	5.13	0.29
Th-Sp-Pos	That	Span	Post	What does your dad think that han comprado tus hermanos en la tienda?	3.51	–0.37

The maximal random effects structure included random intercepts by subject and item, a random by-subject slope over Subject Position, and a random by-subject slope over the interaction. The Type III test of fixed effects yielded a significant effect for Subject Position (β = 0.79, F(1,37) = 88.7, p < .001) but not for Language of T (β = –0.13, F(1,64) = 2.7, p = .108) or their interaction (β = 0.27, F(1,74) = 2.8, p = .096).

Visual inspection of both the raw ratings (Figure 5) and the estimated marginal mean z-scores (Figure 6) reveals a pattern like that of a super-additive effect, but the interaction between Language of T and Subject Position fails to reach statistical significance. This fact could indicate there is a small effect here and that our experiment did not have enough power to detect it. However, as noted in §6.1, we conducted a power analysis prior to data collection and found that, under reasonable assumptions about variance and effect sizes, our sample had sufficient power to detect relevant effects. Because the interaction was not significant, we conducted only a post hoc pairwise comparison by Subject Position, which revealed, as is clear from the graphs, that pre-verbal subjects were rated significantly higher than post-verbal subjects (p < .001).

Figure 5

Mean raw ratings by language of T and subject position when C is English that.

Figure 6

Estimated marginal mean z-scores of ratings by language of T and subject position when C is English that.

We take this result to indicate that code-switching sentences behave like English when C is in English, with post-verbal subjects consistently penalized, regardless of the language of T.

6.4.2 Code-switching results: Spanish C

Turning now to Spanish C que (Table 10), the fixed effects had the same 2×2 distribution, and the maximal random effects structure was also the same. However, in this case, the Type III test of fixed effects yielded a significant effect for Subject Position (β = 0.51, F(1,31) = 40.9, p < .001), for Language of T (β = –0.83, F(1,68) = 134.4, p < .001), and for their interaction (β = 0.66, F(1,83) = 20.8, p < .001). Bonferroni-corrected post hoc pairwise comparisons revealed that post-verbal subjects were rated lower overall than pre-verbal (p < .001) and sentences with English T (and thus a code-switch between C and T) were rated lower overall than sentences with Spanish T (p < .001), as expected. Pairwise comparisons also indicated that, when T was in English, post-verbal subjects were rated significantly lower than pre-verbal subjects (p < .001); when T was in Spanish, though, we failed to find a significant difference between pre- and post-verbal subjects (p = .107). Visual inspection of the results in Figures 7 and 8 revealed the characteristic pattern of a super-additive effect.

Table 10

Code-switching results: Spanish C.

Condition	C	T	Subject Position	Example	Raw mean rating	Estimated marginal mean z-score of rating
Qu-En-Pre	Que	Eng	Pre	Qué piensa tu padre que your brothers have bought at the store?	4.37	–0.02
Qu-En-Pos	Que	Eng	Post	Qué piensa tu padre que have bought your brothers at the store?	2.34	–0.86
Qu-Sp-Pre	Que	Span	Pre	What does your dad think que tus hermanos han comprado en la tienda?	5.55	0.48
Qu-Sp-Pos	Que	Span	Post	What does your dad think que han comprado tus hermanos en la tienda?	4.98	0.30

Figure 7

Mean raw ratings by language of T and subject position when C is Spanish que.

Figure 8

Estimated marginal mean z-scores of ratings by language of T and subject position when C is Spanish que.

Both the statistical results and the figures point to an effect of Spanish T in this case: the reduction in acceptability occasioned by post-verbal subjects is ameliorated by the presence of Spanish T alongside Spanish C que.

Although we have reported the outcome of each of the eight conditions in our experiment at this point, it is illuminating to consider the results in a different way: for the next two analyses, instead of holding C constant, we hold T constant and compare the effects of manipulating the language of C.

6.4.3 Code-switching results: English T

Turning first to sentences with English T (Table 11), we had a 2×2 design with Language of C (English that/Spanish que), Subject Position (Pre-/Post-verbal), and their interaction as fixed factors. The maximal random effects structure included random intercepts by subject and item, as well as by-item random slopes over Language of C and over its interaction with Subject Position, and a by-subject random slope over the interaction. The Type III test of fixed effects yielded a significant effect for Subject Position (β = 0.88, F(1,60) = 99.9, p < .001) and for Language of C (β = 0.27, F(1,42) = 8.9, p = .005), but no interaction (β = 0.07, F(1,59) = 0.2, p = .691). Bonferroni-corrected post hoc pairwise comparisons revealed that post-verbal subjects were rated lower overall than pre-verbal (p < .001) and sentences with Spanish C (and thus a code-switch between C and T) were rated lower overall than sentences with English C (p = .005). Visual inspection of the results in Figures 9 and 10 revealed no super-additive effect.

Table 11

Code-switching results: English T.

Condition	C	T	Subject Position	Example	Raw mean rating	Estimated marginal mean z-score of rating12
Th-En-Pre	That	Eng	Pre	Qué piensa tu padre that your brothers have bought at the store?	5.13	0.29
Th-En-Pos	That	Eng	Post	Qué piensa tu padre that have bought your brothers at the store?	2.97	–0.62
Qu-En-Pre	Que	Eng	Pre	Qué piensa tu padre que your brothers have bought at the store?	4.37	–0.02
Qu-En-Pos	Que	Eng	Post	Qué piensa tu padre que have bought your brothers at the store?	2.34	–0.86

Figure 9

Mean raw ratings by language of C and subject position when T is English.

Figure 10

Estimated marginal mean z-scores of ratings by language of C and subject position when T is English.

The result of this analysis shows that acceptability is penalized for switching between C and T and for post-verbal subjects overall, but we find no interaction between post-verbal subjects and the distribution of C/T. Instead, post-verbal subjects are consistently lower in acceptability than pre-verbal subjects with English T, even with a Spanish C.

6.4.4 Code-switching results: Spanish T

Finally, we held T constant in Spanish and made the same 2×2 comparison (Table 12). In this case, the maximal random effects structure included a random intercept by subject and item, and a random by-subject slope over the interaction of the fixed effects. The Type III test of fixed effects yielded a significant effect for Subject Position (β = 0.39, F(1,112) = 23.6, p < .001), for Language of C (β = –0.44, F(1,103) = 33.0, p < .001), and for their interaction (β = 0.50, F(1,103) = 10.3, p = .002). Bonferroni-corrected post hoc pairwise comparisons revealed that post-verbal subjects were rated lower overall than pre-verbal (p < .001) and sentences with English C (and thus a code-switch between C and T) were rated lower overall than sentences with Spanish C (p < .001). Pairwise comparisons also indicate that, when C was in English, post-verbal subjects were rated significantly lower than pre-verbal subjects (p < .001); when C was in Spanish, though, we failed to find a significant difference between pre- and post-verbal subjects (p = .198). Visual inspection of the results in Figure 11 and Figure 12 revealed the characteristic pattern of a super-additive effect.

Table 12

Code-switching results: Spanish T.

Condition	C	T	Subject Position	Example	Raw mean rating	Estimated marginal mean z-score of rating
Th-Sp-Pre	That	Span	Pre	What does your dad think that tus hermanos han comprado en la tienda?	5.13	0.27
Th-Sp-Pos	That	Span	Post	What does your dad think that han comprado tus hermanos en la tienda?	3.51	–0.37
Qu-Sp-Pre	Que	Span	Pre	What does your dad think que tus hermanos han comprado en la tienda?	5.55	0.46
Qu-Sp-Pos	Que	Span	Post	What does your dad think que han comprado tus hermanos en la tienda?	4.98	0.32

Figure 11

Mean raw ratings by language of C and subject position when T is Spanish.

Figure 12

Estimated marginal mean z-scores of ratings by language of C and subject position when T is Spanish.

We observe a penalty for switches between C and T and for post-verbal subjects overall, but the most relevant finding is the super-additive effect: the acceptability penalty for post-verbal subjects is significantly ameliorated when C is in Spanish.

6.4.5 Summary of code-switching results

When we hold C constant, we find that post-verbal subjects are consistently penalized when C is in English (§6.4.1), irrespective of the language of T, whereas the acceptability of post-verbal subjects varies when C is in Spanish (§6.4.2): post-verbal subjects are much more acceptable when T is also in Spanish. A similar pattern emerges when we hold T constant. When T is English (§6.4.3), post-verbal subjects are much worse regardless of the language of C; when T is Spanish (§6.4.4), it varies, such that post-verbal subjects are more acceptable when C is also Spanish.

Figure 13 presents the rating for pre-verbal (blue line with circles) and post-verbal (orange line with squares) subjects for each of the four possible combinations of C and T. On the secondary y-axis, it shows the difference between the two (gray bars). This graph shows that post-verbal subjects are consistently penalized, but that penalty largely disappears when both C and T are in Spanish.

Figure 13

Differences in acceptability, pre- and post-verbal subjects, by language of C and T.

7.1 Evaluating predictions

Returning to our research question, we asked: In Spanish/English code-switching, which functional head (C, T, or both) determines the subject’s position in the clause? We considered three hypotheses and their predictions for each experimental condition. Table 13 reviews those predictions and evaluates them according to the results presented above. Note that for the two key predictions, highlighted in gray, the T Hypothesis and the C Hypothesis were not borne out. Instead, our results are most compatible with the C+T Hypothesis, replicating our previous results, reported in §3.3 (Ebert & Hoot 2018). The remainder of this section considers the implications of our findings.

7.2 Implications: T Hypothesis

The first hypothesis we considered was that T would determine subject position, a view supported by the traditional view of the EPP as a feature of T, including parameterized versions of the EPP (Alexiadou & Anagnostopoulou 1998). Previous work on word order in code-switching (MacSwan 2004; 2013; Vanden Wyngaerd 2020) has found evidence that the language of the verb (taken by those authors to be the same as the language of T) determines word order in the clause, but our results do not support this hypothesis. The crucial piece of evidence is that this hypothesis predicts a sentence from condition Th-Sp-Pos like (23) should be acceptable, which we did not find.

To interpret this result, recall from §2 that theories of null and post-verbal subjects in Spanish that appeal to the EPP have two basic forms: (i) the EPP is parameterized such that the Spanish EPP is satisfied by V-to-T movement or (ii) the EPP is the same in all languages but Spanish allows a silent or elided pronoun in Spec,TP.

Regarding the first account, it seems reasonable to expect that post-verbal subjects should be allowed any time there is V-to-T movement of a verb with rich agreement morphology (i.e., a Spanish verb). It is difficult to reconcile such an account with our findings, and so we take our experiment as evidence against this understanding of the EPP.

Regarding the second account, we cannot rule out the existence of an expletive pro filling Spec,TP in (23) nor a copy of the subject in Spec,TP that is subsequently elided. However, if expletive pro is part of the Spanish lexicon, it could be present in any of our sentences, as could an ellipsis operation that deletes the higher copy of the subject in Spec,TP. We see no reason to expect that these solutions would apply only when both C and T are in Spanish. That is, if pro is available in the lexicon, why can’t it be inserted to salvage sentences with post-verbal subjects when only T or only C is in Spanish? Likewise, if it is possible to copy the subject to Spec,TP to satisfy the EPP and then elide it at PF, why should that only apply to some code-switched sentences and not all? For this reason, we interpret our results as pointing away from these explanations as well.

7.3 Implications: C Hypothesis

One of the most straightforward predictions regarding the role of C, González-Vilbazo and López’s (2012) Phase Head Hypothesis, does not find support in our data. Previous evidence from code-switching has also pointed away from this hypothesis (Ebert & Hoot 2018; Sande 2018; Vanden Wyngaerd 2020). The crucial piece of evidence against the C hypothesis is Qu-En-Pos sentences like (24), which were predicted to be acceptable but were not.

Similarly, we contend that our data is evidence against the initial version of Feature Inheritance introduced by Chomsky (2008), although this concept is much more nebulous and harder to evaluate. Recall that Chomsky explains FI thus: “…for T, phi-features and Tense appear to be derivative, not inherent: basic tense and also tense-like properties (e.g., irrealis) are determined by C…. In the lexicon, T lacks these features” (Chomsky 2008: 143) Our understanding of Chomsky’s proposal is that T is mostly empty of formal features, and when C is merged it copies to T features that include Tense itself, along with phi-features and the EPP (whatever form that takes).

Accordingly, T must inherit its features from C even when there is a code-switch between C and T, despite the two heads surfacing in different languages, given the ample evidence that such switches are possible. In those cases, on the assumption that the apparent language of C reflects its features, English C will cause T to inherit a different set of features than Spanish C. When C is English and T Spanish, English C copies its EPP feature requiring a pre-verbal subject (among other features) to T, predicting only pre-verbal subjects. So far, so good. The problem emerges when C is Spanish and T English, as in (24), which similarly disallow post-verbal subjects. If the EPP feature that requires pre-verbal subjects is inherited from C, where does it come from in this case? Our data seems to point to an English T that has an EPP requirement despite being selected by a C which presumably does not. If T has an EPP without inheriting from C, we take our results to run counter to such a version of FI.

Finally, another proposal that focuses on the role of C is that of Pesetsky and Torrego (2001), who argue that the difference between Spanish and English turns on the difference between complementizers in the two languages. Unfortunately, their proposal does not provide a complete treatment of the Spanish facts (Pesetsky and Torrego include Spanish in the paper to contrast it with English and support the broader argument, but their focus is not on Spanish) and so we are unable to fully evaluate it. Nonetheless, although it hinges on differences in the C head, because it still focuses on the ways C and T interact, further exploration of this option could prove fruitful.

7.4 Implications: C+T Hypothesis

In §4, we made predictions following Sande’s (2018) account because it is the most explicit relevant hypothesis we are aware of, and our results are consistent with her proposal. Recall that for null subjects, Sande contends that they must be licensed by two different types of features: syntactic licensing from T and discursive licensing from C. This view is compatible with a theory that places the EPP entirely on T, as most traditional approaches do, while reconciling it with the code-switching facts, including ours, supporting a role for C. Given that (certain types of) post-verbal subjects are only felicitous under certain discourse conditions, it seems sensible to extend her account and postulate some discourse feature that is required to license post-verbal subjects. It seems equally sensible to locate such a feature in the C-domain, which has long been linked to discourse functions.

Accepting this approach suggests possible avenues for future work in code-switching. If the availability of post-verbal and null subjects in Spanish/English code-switching depends on two types of features, it would be valuable to examine language pairs with different feature arrangements. For example, consistent verb-initial languages like Celtic, Mayan, and Austronesian languages do not require discourse licensing of post-verbal subjects (Clemens & Polinsky 2017; Rouveret 2017). It would be valuable to investigate the distribution of subjects when such languages are combined with languages like English (to examine the locus of the syntactic licensing alone) and languages like Spanish (to examine the role of discourse licensing). Within Spanish/English (or, more broadly, Romance/Germanic) code-switching, it would be valuable to look into different types of post-verbal subjects, some of which are likely tied more strongly to discourse restrictions than others. For example, subject inversion in matrix questions is taken by many to be obligatory in non-Caribbean Spanish, and subjects of unaccusative verbs canonically appear in post-verbal position, appearing pre-verbally only when marked discursively as topics. In both cases, any proposed role for discourse features from C is likely reduced (or, in the case of unaccusatives, perhaps even reversed, where pre-verbal subjects require discursive licensing instead). An experiment testing these cases in code-switching could be a good next step to evaluating Sande’s hypothesis.¹³ A third possibility, suggested by an anonymous reviewer, would be to use code-switching data to investigate the discourse features of C, whether they are universal, and how they vary across languages. It would be interesting in this regard to combine languages that are highly discourse-configurational, like Hungarian, Basque, Catalan, or Turkish, with languages that instantiate discourse features differently, like English, Spanish, or German.

An alternative explanation to the C+T Hypothesis would be to appeal to certain variations on Feature Inheritance. For example, C and T could both be required to be in Spanish because features are shared simultaneously by the two heads (e.g., Ouali 2008) or because the two heads are exact copies of one another which must thus be in the same language (e.g., Gallego 2014).¹⁴ At the very least, the apparent close relationship we observe between C and T suggests some kind of interaction between the two heads, and it is for this reason that both we (Ebert & Hoot 2018) and Vanden Wyngaerd (2020) have pointed generally in this direction previously. Unfortunately, it remains unclear how FI can be extended to CS, and the field has not yet reached a consensus on many aspects of FI more generally. To craft a viable understanding of the functioning of FI in code-switching and test it against our data would require specific answers to numerous questions which are still debated.

7.5 Other considerations

Several other points merit consideration in interpreting our results.

First, we have largely assumed implicitly that something special is required to license post-verbal subjects, whereas pre-verbal subjects require little explanation. Such a formulation may seem sensible to speakers of canonical SVO languages, and we have argued for the justification of our hypotheses, but from a theory-internal perspective, it is curious. After all, the EPP, motivating pre-verbal subjects, is the special stipulation that Chomsky (2008: 156) called “an annoying problem” requiring justification. Yet our data suggest the opposite: post-verbal seems to be the special case, only available under certain circumstances, while subjects must be pre-verbal when either C or T alone is in English. Perhaps an alternative analysis focused on the English EPP could be fruitful; we leave that for future work.

Alternatively, perhaps the default or canonical nature of pre-verbal subjects is the reason we observe “English-like” behavior when we have either English C or English T. Given that code-switching between C and T is rare, perhaps our participants default to SVO, which is canonical in both languages, simply because they are less sure. Another similar idea would be to posit that perhaps subjects are always pre-verbal when the switch is between C and T because pre-verbal subjects are common to both languages, while post-verbal subjects are available only in Spanish. Indeed, this idea is a longstanding one in the code-switching literature, dating at least to Poplack’s (1980) Equivalence Constraint.

Finally, we should recognize potential design limitations of our study. For instance, we felt that increasing the number of code-switches would make our stimuli even more contrived, so there was always a single switch just before or just after the complementizer. Consequently, a given sentence type always started in Spanish and ended in English or vice versa. One potential confound is that the sentences with Spanish C+T that were rated high regardless of subject position always began with an English matrix clause, including an English wh-question, so perhaps there is an effect of the matrix clause. However, the sentences with English C and Spanish T, which were not rated as high, had the same matrix clauses.