Orality and overtness: effects on Spanish subject use

3.1 Context for the corpus

CorDELES was created to investigate the historical path that led to the higher rate of overt SPE in Caribbean Spanish as seen in (1). Considering BP has also undergone this change (Duarte 1993), their shared language contact history is likely responsible. It is crucial to remember that language contact, as well as language change, must always be viewed through the lens of language acquisition and use (Weerman 1993). It is not that two or more concrete monoliths of language come into contact, but that individual speakers learning and using new languages do. With that in mind, acquisition literature has shown that null subject systems are more difficult to acquire than non-null subject systems, particularly for L2 learners (e.g. Bini 1993; Margaza and Bel 2006; Pérez-Leroux and Glass 1999). For example, (2) is a sample from an intermediate L2 speaker who uses the overt pronoun nosotros ‘we’ without emphasis or focus.^[3]

Different theories have arisen to explain why this might be. For instance, the Interface Hypothesis suggests the null subject system is difficult to acquire because it lies on the interface between multiple domains, specifically pragmatics and syntax (Sorace 2011). Alternatively, the Interpretability Hypothesis works with the concept of uninterpretability from the Minimalist framework, suggesting second language learners past the critical threshold, i.e. adults, are only able to access interpretable (that is, semantically active as well as syntactically present) features (Hawkins and Hattori 2006; Tsimpli and Dimitrakopoulou 2007; Tsimpli and Lavidas 2019). This means that when they come across uninterpretable features such as those on null subjects, they misanalyse when to use them and overcompensate with features that are interpretable: overt subjects. If it is indeed the case that null subjects are harder for adult learners to acquire, then, following Dahl’s (2004) definition of complexity as L2-difficulty, the resultant increased use of overt pronouns is an act of simplification. Thus, language contact is simplifying when the feature being learned is difficult for adults to acquire (cf. Walkden and Breitbarth 2019). Trudgill (2011) further qualifies the type of contact that would lead to strain in acquiring new systems: short-term, loose-knit, adult language learning. Those conditions perfectly describe the context for African learners of Spanish in colonial Latin America.^[4]

During this period, enslaved (West-)Africans were taken to the Caribbean and coastal South America (e.g. Bowser 1974). As adult second-language learners who would have had to quickly learn Spanish without consistent access to a dense network of native speakers, they would have struggled to acquire the L2-difficult null subject system, preferring overt subjects. Their children would then nativize this system. This origin is exactly what Sessarego (2013) proposes for the genesis of Afro-Hispanic Languages of the Americas (AHLAs). These varieties have been spoken by the subsequent generations of the original enslaved Africans and they demonstrate the kind of change predicted, namely an increased rate of overt pronouns. To verify this account, we need to look into the diachronic trajectory of SPE. Ideally, we would do this using historical AHLA data, but not enough exists in the written record to compare systematically over centuries. What we can do is track if this potential change might have bled into other varieties of Latin American Spanish, comparing it with AHLA data when we have it. To this end, I constructed CorDELES to attempt to find out whether overtness rates increase over the last five centuries, and if they do, if they are higher in countries with larger Afro-Hispanic populations.

3.2 Corpus composition

Although several western European languages are represented in parsed corpora that allow the simple extraction of null subjects, such a Spanish corpus (especially a Latin American one) does not exist. With this in mind, and due to the difficulty of counting null subjects, it was clear from the start that they would need to be annotated by hand. The simplest method would have been to use an extant corpus to export a list of every finite verb that I would have then annotated by hand. However, despite there being many excellent historical and/or dialectal Spanish corpora (e.g. CDE [Davies 2002], CORDIAM [CORDIAM n.d.], CORDE [Real Academia Española n.d.], and CDH [Real Academia Española 2013]), none provides the necessary metadata or is sufficiently POS-tagged or lemmatized so as to produce such a list. Instead, I created a corpus that I could balance and annotate directly.

CorDELES was designed to consist of 64 texts pulled from eight countries over the period 1500–1899 that are split into two genres. Table 1 shows a grid of the dimensions of the corpus as well as where each individual text (marked by an abbreviation of its title) fits into it. In the event that a prose text could not be found, one written in verse was included. Unfortunately, an extant text for every cell combination still could not be found, resulting in a shortage of seven texts. For a full listing of texts, please see Appendix A.

The eight countries can be grouped into three sections on a geographical basis: Caribbean/Central America (Dominican Republic, Cuba, and Panama), South America (Peru, Colombia, Bolivia, and Venezuela), and Peninsular Spanish as a baseline. Each Latin American country was selected based on whether a known AHLA community lives there^[5] (Klee and Lynch 2009). AHLA populations are densest in coastal areas, reflecting where their enslaved ancestors were originally taken to. The Caribbean islands therefore hold the densest populations and should be the most likely to use overt subjects (Orozco and Guy 2008). To check the relationship between expected population level^[6] and pronoun usage, countries with low (Panama, Bolivia, Peru), mid (Venezuela, Colombia), and high (Cuba, Dominican Republic) Afro-Hispanic populations from the Caribbean, Central America, and South America were all represented. Admittedly, it is less than ideal to categorize Latin American varieties by country, as many of these countries already contain their own well-differentiated dialectal regions, not to mention that many of these dialects cross borders (cf. Orozco and Díaz-Campos 2016; Orozco 2018). However, the limited availability of texts made it difficult enough to fill each cell in Table 1 based only on country. Narrowing the pool to certain sub-regions would have resulted in even more gaps in the corpus. Further, information on provenance beyond country was not even available for many texts. For these reasons, the corpus was regionally delineated by country. Genre was accounted for in a rough literary versus non-literary break to balance the corpus in case it influenced SPE. Because the corpus is so small (3,773 tokens of pronominal subjects) in relation to the number of parameters and there is already a paucity of texts for certain country/century combinations, the corpus would not have been able to be adequately balanced for genre on a more fine-grained level. However, as Section 4 demonstrates, orality is a much better metric for analyzing textual effects on SPE in the corpus. Diachronically, the 16th century is included to create a baseline of the Spanish spoken in Latin America when the first Africans arrived.

The main sources for the texts were Cervantes Virtual (Biblioteca Virtual Miguel de Cervantes 1999), Biblioteca Digital Hispánica (La Biblioteca Nacional de España 2008), and the Digital Library of the Caribbean (Digital Library of the Caribbean 2004). A continuous sample of 2,000–3,000 words was randomly selected from each text. CorDELES totals 151,295 words. Some supplemental texts that were of interest but did not fit into the main corpus were also set aside, most of which will be included with data from the main corpus in the following section. The first of these texts is a transcript from an Afro-Bolivian field interview conducted in 2010 (pulled from several such transcripts provided in Sessarego [2011]). The rest are from Cantos populares de mi tierra and Secundino el zapatero, a series of poems and a play written by Candelario Obeso, an Afro-Colombian poet and playwright from the 19th century. Obeso wrote in Afro-Colombian, providing rare insight into actual historical AHLA language. These four texts also make for a fascinating comparison as they are written by the same author but split across three genres: play, poetry, and prose (pulled from the dedication). Obeso also provided translations of his poetry into standard Spanish, so any marked difference in SPE would be very telling. Each sample was then transcribed where necessary, POS-tagged by an automated tagger, and annotated by hand in XML.

3.3 Annotating null subjects

Part of the necessity of creating a new annotated resource rather than comparing figures from previous studies is that null subjects are very difficult to count consistently. Many judgements need to be made as to what counts as a null subject, and no two studies have taken the exact same approach, making their data tough to compare (as pointed out in de Prada 2009; Martinez-Sanz 2011; Travis 2007). In order to reliably compare frequencies from different countries, they need to have been annotated under the same guidelines. So, before looking at the data for this corpus, we must walk through some methodological decisions that needed to be made. The motivation behind these decisions was to only operate within the envelope of variation (cf. Orozco and Hurtado 2021a; Otheguy and Zentella 2007); that is, to only extract tokens for which both the presence and absence of a subject pronoun were possible. In other words, the goal was to annotate every environment where alternation was possible and only those environments.

Firstly, the focus of CorDELES is the relationship between the subject and finite verb of each clause, meaning subject pronouns of nonfinite clauses were not considered. Imperatives were also excluded due to their unique relationship with subject pronouns that allows their omission even in NNSLs. Similarly, cases of coordination were counted as only one realization of the pronoun since NNSLs are able to omit the second pronoun. E.g. come y bebe mucho ‘he/she eats and drinks a lot’ would count as only one null subject (that was elided for the second VP) rather than two.^[7] Fixed expressions such as hace tiempo ‘time ago’ and es decir ‘that is to say’ were excluded.

Most importantly, null pronouns were tagged for animacy and referentiality because it is not possible (or at least exceptionally rare in the case of inanimacy) in standard Spanish for inanimate, non-referential/impersonal subjects to be overt. If a variety were to realize such an expletive subject overtly, it would be immediate proof that it could no longer be classified as a consistent NSL (cf. Roberts and Holmberg 2010). If this were to occur, it would be one of the final stages in the transition from an NSL to a NNSL. However, such a change is not expected to appear in the corpus (and indeed does not), so I am currently only interested in the environment where overt pronouns might first start overtaking null pronouns: animate, referential subjects. In addition to SPE, each clause was tagged for clause type, inversion, verbal and pronominal morphology, co-referentiality with the previous subject, priming, and interrogative status. The purpose of the present study is to analyze the potential regional and diachronic trends, so I will only consider (referential, animate) pronouns at the moment.

The biggest challenge for replicability came with some of the oldest texts in the corpus whose syntax was often difficult to parse. They contained many run-ons accompanied by lots of relative clauses that left whether a subject was coordinated or not, as well as whether it shared reference with the previous subject, unclear. In those instances, I defaulted toward creating a new null token with a different referent. I was as consistent as possible in my annotation so the texts are as comparable as can be reasonably expected, especially as I was the sole annotator. Additionally, sometimes handwritten texts could not be fully transcribed as some words were lost to bindings or smudged. If these words seemed to be the subject or finite verb of a clause, they were marked ‘unintelligible’ and excluded.

3.4 Measuring orality

Recall that Rosemeyer (2019) measured orality based on the presence of the private verbs achar ‘mean’, pensar ‘think’, acreditar ‘believe’, and crer ‘believe’; the present progressive, the demonstrative neuter pronouns isso and isto ‘this’; the time and place adverbs aqui ‘here’ and agora ‘now’; and the discourse markers né ‘isn’t it?’, bom ‘well’, pois ‘so’, então ‘so’, and olha ‘listen’. These factors were chosen to measure BP orality, but, because of its genetic and typological closeness to Spanish, they can easily be adapted for use in my corpus. To create a slightly easier number of variables to find and count, I narrowed the category of private verbs down to creer ‘to believe/think’ and pensar ‘to think’ and excluded discourse markers. The rest of the variables translate easily, leaving us with the progressive; the demonstrative pronouns esto ‘this’, eso ‘that’, and aquello ‘that’; and the time and place adverbs aquí ‘here’ and ahora ‘now’. The oral nature of these factors is confirmed when they are searched for by genre in the Corpus del Español (Davies 2002). Figures 1–8 each demonstrate that although the frequency of every oral marker fluctuates over time, they all consistently increase in tokens from the genre with the lowest orality to the highest: academic (ACAD) → news (NEWS) → fiction (FICT), → oral (ORAL).^[8]

Figure 1:

Creo ‘I think/believe’ (n = 35,204).

Figure 2:

Pienso ‘I think’ (n = 8,992).

Figure 3:

Estoy + gerund ‘I am X-ing’ (present progressive) (n = 5,126).

Figure 4:

Esto ‘this’ (n = 163,710).

Figure 5:

Eso ‘that’ (n = 87,152).

Figure 6:

Aquello ‘that’ (n = 19,949).

Figure 7:

Aquí ‘here’ (n = 79,282).

Figure 8:

Ahora ‘here’ (n = 64,980).

Once the validity of the orality measurement had been confirmed, I applied it to CorDELES. In order to quantify the gross counts in a comparable manner, I created the ORSCORE, which is derived by firstly aggregating all of the frequencies for each text, dividing them by the word count of that text, and then multiplying this figure by 100. For instance, the 19th century Spanish play CPC (the most oral text in the corpus) had 11 instances of private verbs, 6 of the progressive, 22 of demonstrative neuter pronouns, and 14 of the time and place adverbs, totaling 53 tokens. This gross frequency was then divided by CPC’s word count of 2,260, resulting in the quotient 0.02345. To make this figure more accessible, it was multiplied by 100 to reach the final ORSCORE of 2.35. On the other end of the spectrum, the 16th century Dominican court document SDJ had only 1 instance of a private verb and 0 instances of any of the other factors. Its word count was 2,008, creating the quotient 0.0005 and the final ORSCORE of 0.05.

3.5 Testing with mixed-effects models

A series of statistical models were run to test the effects of orality, time, and region on SPE. Each model was generated using the lme4 package (Bates et al. 2015) in R (R Core Team 2021). A simple linear regression model (lm) is used to first establish the relationship between orality and SPE. In other words, a single continuous response variable (the proportion of overt subject pronouns per text) is modeled as a function of the predictor (ORSCORE). A regression line can then be plotted to represent the conditional mean formed by the model. Coefficients are the numerical values that describe this relationship and are the basis for predicted values. The differences between the model’s predicted values and the actual data points are called residuals. These residuals can be used to determine how well the model explains the data by being compared with the residuals generated by a null model. This measurement of model fit is called R².

The contact hypothesis, however, cannot be evaluated by a simple linear model for several reasons. First, it deals with more than one predictor (time, region, and eventually orality). One of the benefits of looking at more than one predictor in a model is that one can test for interactions. If there is an interaction between two predictors, it means that one predictor influences the response variable differently depending on the second predictor. For instance, an interaction between gender and age might mean that younger men behave differently than older men in a way that cannot be accounted for by just age or just gender. Second, at least one of these predictors is a categorical variable (region). Third, the response variable is also a categorical variable. For the linear model, SPE is turned into a continuous variable (the proportion of over subject pronouns) in order to simply visualize and analyze the relationship between orality and SPE with each text as a single data point. However, when looking at multiple predictors with numerous levels, it is better for the model to be fed each token as a single data point. Because there are only two options for the response variable (null or overt), it can be called binary. The categorical nature of the response variable means a logistic (not linear) regression needs to be run. In other words, the probability of the subject pronoun being overt instead of null is predicted. Lastly, random effects also need to be accounted for. All of the predictors previously listed are fixed effects which means that they are uniformly predictable, constant, and replicable. Random effects, on the other hand, are idiosyncratic and thus unique to each study. For instance, in an experiment, ‘participant’ would be treated as a random effect to control for differences because tokens from Participant A are more likely to pattern closely together than they are with Participant B’s tokens. For a corpus this could be anything like author, text, or newspaper. For CorDELES, the unique code for each text (docID) is specified as a random effect. Mixed-effects models are used to analyze both fixed and random effects. To fit a logistic regression to a mixed-effects model, glmer is used. This time, model fit is assessed by the Akaike information criterion (AIC) whereby the smaller the number is, the better the fit. The regression results for each model, including coefficient size, p-value, and model fit, are reported in Tables 2–4^[9] in the following section.

4.1 The orality correlation

The ORSCORE of each text in the corpus is plotted against the proportion of overt subject pronouns in each text in Figure 9. We can see that, indeed, an orality effect is at play in the corpus data as there is a clear positive correlation between ORSCORE and the rate of overt SPE in a text. This effect is corroborated by the results of a linear model. I ran lm in R (R Core Team 2021), with ORSCORE as a predictor of the proportion of overt pronouns, finding a clearly significant relationship (β = 0.100, p > 0.0001, R² = 0.351 ± 0.02). These results are summarized in the middle column of Table 2. The R² score of 0.351 means that 35 % of the variance in the data can be explained by this model which is a relatively high level of explanation given all the additional factors that have been known to affect SPE (e.g. focus, switch-reference, person, etc. cf. Section 1).

Figure 9:

Orality and overtness regression line (n = 37). Each plot in this paper was created using ggplot2 (Wickham 2016) in R (R Core Team 2021).

There may appear to be two outliers with respective ORSCOREs of 1.77 and 2.35 in the upper right of the graph that are skewing the data. However, as the rightmost column in Table 2 shows, the p-value is still clearly significant (p < 0.002) when these two texts are excluded. In fact, the coefficient is actually larger, as illustrated by the steeper slope of the regression line in Figure 10, showing the strength of the relationship. Further, the two apparent outliers are the Bolivian oral interview (IIA) and a rapid-paced Spanish play (CPC), the two most impressionistically oral texts of the corpus, as demonstrated by examples (3) and (4). This again confirms the power of the measurement in quantifying orality.

Figure 10:

Orality and overtness regression line (‘outliers’ excluded) (n = 35).

Since orality has been demonstrated to have a significant effect on SPE, its influence on the rest of the corpus can now be investigated.

4.2 SPE over time

First, let us look at the effects of time and region on SPE without orality being accounted for. The following bar charts (Figure 11a–e) show the proportion of null and overt subject pronouns per text. Each bar is a text identified by an abbreviation of its title, e.g. ‘ENT’ refers to the 16th century Dominican text titled Entrémes (cf. Table 1 and Appendix A for a full listing). These texts are ordered chronologically. The non-literary genre is illustrated with a solid border, the literary genre with a dotted border. The bars with a dashed border are the supplemental texts from Afro-Hispanic speakers.

Figure 11:

The bar charts in (a–e) give the proportion of SPE in Bolivia (n = 512), Colombia (n = 1,223), Dominican Republic (n = 573), Panama (n = 575), and Spain (n = 1,043).

The proportion of overt subjects appears to slightly increase in the last two centuries for most of the countries (setting aside the supplemental texts from Colombia), the only exception being the Dominican Republic which remains pretty consistent apart from ASD. That being said, these trends need to be taken with a grain of salt. Due to text scarcity, Bolivia and Panama are both missing multiple cells from this period, including the entire 17th century for Bolivia and the 18th for Panama. Further, even though there is a large gap between the last two Panamanian texts (MPE and HS), they are separated by only three years, not nearly long enough for such a rise to be the result of diachronic change alone. With that taken into consideration, there is no longer any obvious sizeable increase for Panama. This leaves Colombia as the only clear Latin American country with a noticeable increase in overt subjects. However, Spain also shows a steady increase during this period, contradicting a colonial contact account. On the other hand, since there are only a maximum of two texts per century-country combination, it is necessary to first rule out any other potential effects before drawing any conclusions.

Although there does not appear to be any genre distinction in any of the countries,^[12] I ran a mixed-effects logistic regression model using glmer from the lme4 package (Bates et al. 2015) (cf. Section 3.5) to see if it could pick up any underlying patterns that are unclear from visual inspection of the bar charts. SPE was run as a binary dependent variable while country, year (z-scored), and genre (literary vs. non-literary) were fixed effects, and docID was a random effect.^[13] Two other regional groupings with fewer levels were also run: region (Caribbean, South America, Spain) and macro-region (Spain vs. the Americas). Over 20 versions of the model were run, testing for every possible combination of fixed effects with and without interactions. Some iterations of the model did not have enough data to converge. These were versions with multiple interactions or interactions with country due to the numerous levels involved. Of the models that converged, year was the only predictor to ever emerge significant (β = 0.458, SE = 0.152, p < 0.003). Similarly, the model with the best fit according to AIC scores was the model that only specified year as a fixed effect. These results are reported in Table 3. In sum, a positive relationship between year and overt SPE was found, i.e. pronouns are more likely to be overt over time.

Even though genre does not have a significant effect in the model, something similar to genre seems to be influencing the data. The texts with some of the highest overt SPE rates also show high levels of dialogue: ASD is an 18th century Dominican court document that transcribes the testimony of witnesses describing events; HS is a 19th century Panamanian series of poems that frequently address a specific audience in 2nd person; the poetry from the supplemental 19th century Afro-Colombian text CPMT similarly uses frequent dialogue;^[14] IIA is the 21st century transcript of an Afro-Bolivian oral interview; DQ is a sample from the 17th century Spanish novel Don Quixote with a lot of dialogue; and CPC is the 19th century Spanish play with a lot of rapid back-and-forth. Given IIA and CPC also had the two highest ORSCOREs in the corpus (cf. Section 4.1), orality seems to be the key to this pattern. Before verifying orality’s role by introducing it into the model, it is important to rule out other potential sources for this discourse-heavy trend.

All of these texts seem to have rich environments for discourse-switching in which the referent for each subject frequently changes clause to clause. Given that switch-reference has a well-known effect on SPE (e.g. Abreu 2012; Alfaraz 2015; Otheguy and Zentella 2007; Silva-Corvalán 1994; Travis 2005), I had already tagged each pronoun for whether it shares reference with the previous subject token or not (cf. Section 2). However, co-reference levels did not correlate with overt SPE in any of the countries. If the high discourse levels are not a function of switch-reference, then maybe genre was on the right track and this pattern is the result of textual differences.

The next logical step following genre would be to look at subgenre/text-type, specifically plays since they are intuitively expected to have the highest levels of dialogue. There are four plays in the corpus (ENT from 16th century Dominican Republic, ARJD from 18th century Spain, CPMTpl from 19th century (Afro-)Colombia, and CPC from 19th century Spain), and they make up a wide range of overt pronominal rates: 11 %, 3.8 %, 10 %, and 22 %, respectively. This variation is particularly striking considering the lowest and highest of the plays’ frequencies are from the same country and only a century apart. Recall, however, that Rosemeyer (2019) points out that it is misleading to assume that all plays should behave uniformly because they may differ in their degree of orality. For BP, stylistic change in the level of orality represented in plays was creating ‘apparent’ change in the diachronic patterning of certain wh-interrogatives (cf. Section 2). In order to evaluate whether orality is doing the same thing to SPE rates in this corpus, I use the measurement of orality outlined in Section 3.4 to analyze its effects in CorDELES in the next section.

4.3 An interaction between orality and time

In Section 4.1 I discovered a relationship between orality and overt SPE in CorDELES. In the same corpus, I also found a significant correlation between year and overt SPE in Section 4.2. The question follows as to whether the former pattern is a real trend responsible for an artificial inflation of the latter. This would mirror Rosemeyer’s (2019) findings but be detrimental for the ability of this corpus to accurately investigate the contact hypothesis. There is another possible eventuality in which the inclusion of orality would give the mixed-effects model more clarity, revealing nuances to the relationships SPE has with year and region that could not be captured with the orality noise unaccounted for. Recall that the R² value for the linear model run in Section 4.1 was only 0.351, meaning only 35 % of the variance in the data can be explained by the relationship between orality and proportion of overt pronouns. Despite the strong relationship in Figure 9, there is still clearly a high degree of variation in overt SPE rates for texts of a similar ORSCORE. For instance, the three texts DLYD (from 17th century Panama), ADLA (from 19th century Bolivia), and DQ (from 17th century Spain) all had similar ORSCOREs of 0.48, 0.42, and 0.44, respectively, whereas their proportion of overt subjects varied appreciably with rates of 2.9 %, 10 %, and 17 %, respectively. Much of this variation is likely due to linguistic influences such as pragmatic conditioning (e.g. focus, emphasis, contrast), but it also leaves room for the possibility that there are underlying regional (and perhaps further diachronic) trends in the corpus still to be uncovered. The next step is to try and tease those apart from this robust orality effect. Either way,^[15] the inclusion of orality in the model is paramount to get an accurate analysis of the diachronic and regional trends in CorDELES.

The best way to do this is to include orality as a predictor in the mixed model. Glmer from lme4 (Bates et al. 2015) was again run in R (R Core Team 2021) with SPE as a binary dependent variable and docID as a random effect. Again, every possible combination of predictors and interactions that included orality was run. Once more, the number of tokens did not span far enough to cover iterations with multiple interactions or interactions with country. Predictably, every model that did converge found ORSCORE to be significant. What is new is that certain combinations found each regional grouping and/or its interaction with ORSCORE to be significant. Of all these models, the one with the lowest AIC (i.e. the model that best described the spread of data) had ORSCORE, year (both z-scored), and macro-region (Spain vs. the Americas) as fixed effects with an interaction between year and ORSCORE.^[16] The results of this model are summarized in Table 4 which reveals that this final model returns three significant findings. First, as expected, ORSCORE is the most significant predictor with an effect size of 1.030 (this is the coefficient that determines the slope of the regression line, cf. Section 3.5) and a p-value less than 0.001.

Second, now that orality is included in the model, year is no longer significant on its own; however, its interaction with orality is (β = −0.43702, SE = 0.20634, p < 0.034177). The negative coefficient means that for every year that increases, the effect of orality on overtness decreases. In other words, the texts are becoming more oral over time, but their levels of overtness are also increasing beyond what orality can explain on its own. This trend is clear with the use of plogis in R (R Core Team 2021) to calculate the predicted probabilities for this interaction given a specific ORSCORE. In other words, one can provide a start year, end year, and precise ORSCORE and the function will return the proportion of overt SPE that the model predicts for the given orality level for each year. For instance, for the median ORSCORE of 0.45, the rate of overt SPE is predicted to increase from 5.7 % to 11 % from 1,500 to 1,900. These predictions are plotted in Figure 12. Note that the gaps between lines become smaller over time, narrowing toward each other. This is the effect of the interaction between orality and year at play: the difference in SPE rates between orality levels is not as large in 1,900 as it is in 1,500. Surprisingly, while the minimum (0.05), median (0.45), and mean (0.66) ORSCOREs all show a steady increase over time, the maximum ORSCORE (2.35)^[17] predicts a decrease from 79 % to 30 %. This prediction does not match the actual data as none of the SPE proportions pass 30 % to begin with. This discrepancy is likely because there are only three texts with ORSCOREs above 1.00, two of which are more than double and triple the mean ORSCORE. Since the model has so little data for such high ORSCOREs (and none before the year 1877), it likely cannot make accurate diachronic predictions for ORSCOREs that large. This is not to say that the interaction is not present for high ORSCOREs, just that they are more likely to pattern like the rest of the ORSCOREs affected by the interaction. In other words, the overt SPE rates predicted for an ORSCORE of 1.08 would increase over time,^[18] but the gap between them and the rates for a 0.66 ORSCORE would still shrink.

Figure 12:

Predicted SPE across ORSCOREs over time.

In addition to the interaction between year and orality, an effect of macro-region has also come out significant in the new model (β = 0.62880, SE = 0.31997, p < 0.049389). This means that the American data favors overt SPE more than Spain. Although this is hardly surprising as the dialectal differences in SPE have been well-attested (thus motivating the present study, cf. Section 1), it was troubling that the original model did not find a difference. That being said, while the effect is consistent with the regional half of the contact hypothesis, it does not fully answer it. The contact hypothesis specifically predicts an interaction between macro-region and year or between year, orality, and macro-region. This would have shown that the two regions do not just behave differently, but came to behave differently diachronically, i.e. after the catalyst of language contact in the Americas. However, neither of those models found these interactions significant. Indeed, the three-way interaction model did not find anything significant at all, even orality. Their AIC scores were accordingly higher than the final model’s, meaning they did not explain the data’s distribution as well.

In sum, the new model proves that there is a diachronic effect in the corpus which is tied to orality. It has also demonstrated that a regional trend seems to have been obscured by orality, something not apparent in either Figure 11a–e or the initial models. The following section will flesh out the implications of these findings and the role of orality in general.