An improved test of the constant rate hypothesis: late Modern American English possessive have

1 Introduction

Competition is a frequent form of language change that proceeds as an opposition between variants of a linguistic variable (e.g., Kroch 1994;^[1] Zehentner 2019: Ch. 6.3). Most research has focused on the dynamics of two-variant systems. Competition normally arises from a small modification of a linguistic pattern, which thereby creates a new virtually synonymous alternative that is incompatible with the old form and hence vies with it for the same opportunities to be uttered (e.g., Mufwene 2009: 20, similarly Ritt 1995: 54). The innovative variant begins to co-exist side by side with the conservative one in the mind of a speaker leading to variation within an individual (e.g., Roeper 1999). It also spreads from one speaker to another so that the competing grammatical options propagate through the population resulting in variation within the speech community (e.g., Labov 2001). While several outcomes may develop from this state of affairs, such as a failed change, stable differentiation or free variation (e.g., De Smet et al. 2018; Postma 2017; Roberts 2007: Ch. 4), the most salient consequence and focus of this paper is the complete replacement of the old with the new form across several generations during a transitional period, a time when both variants have substantial currency in a population.

The Constant Rate Hypothesis (CRH) is the prediction that “when one grammatical option replaces another with which it is in competition across a set of linguistic contexts, the rate of replacement, properly measured, is the same in all of them” (Kroch 1989: 200). More technically speaking, the CRH hypothesizes that a regression of the probability of a new form as a function of time and grammatical context will return the time parameter as independent of the grammatical context parameter(s) even though contexts may be significantly different from each other. My own formulation of the CRH takes the form of the following argument:

1. Premise (i) – Distinct grammatical structures represented in the mind can access the innovative/conservative variant in competition^[2] thereby creating different grammatical contexts for the change.

2. Premise (ii) – The frequency of use of the innovative/conservative variant in competition (in any context) depends solely on its entrenchment.^[3] In the case of a replacement change, entrenchment, shared between the two variants, gradually strengthens/weakens^[4] for the new/old form to the point of its complete dominance/non-availability over successive lifespans and generations of speakers.

3. Premise (iii) – There are diachronically stable factors that influence whether a particular context would occur with the innovative/conservative variant in the first place.^[5]

4. Conclusion – The rising/falling diachronic pattern of the frequency of use of the innovative/conservative variant is independent of grammatical contexts.^[6]

The following points concerning the CRH are particularly noteworthy. First, the CRH stands as an alternative to the claim that rates of change tend to be inherently variable across contexts, or are perhaps predictable from some feature of contexts. For example, Bailey uses a concept of “weighting”, defined by markedness, rule ordering or frequency, to propose that “environments are […] faster and slower according to their relative heaviness” and that changes “operate ‘faster’ in heavier environments than in lighter ones” (1973: 44–5, 82, 110–3, 143). Second, the CRH’s implication – if a single innovation is expressed in distinct grammatical contexts, it changes at the same rate in all of them – is sometimes turned around so that identity in the rates of change is taken as evidence for the relatedness of contexts. This logical inversion is not legitimate (e.g., Kauhanen and Walkden 2018: 508) and reasoning along such lines should be avoided, or at least be signposted explicitly as an extension of the CRH, not the CRH itself. Third, the CRH is frequently (mis)interpreted to claim that a change is actuated earlier in some contexts than in others (e.g., “changes may occur in a particular syntactic context before spreading to another”, Kastovsky 2012: 158). Rather, it asserts that once actuated, a change has full potential to be used in all contexts and independent factors explain why corpora record a new form more frequently in one over some other context. Fourth, the CRH has been criticized as merely relating “to an emergent property of language at the level of aggregate corpus data” (anonymous, p.c.). This is unlikely to be correct because the theoretical underpinnings of the CRH (as explicated in premises (i) to (iii) above) are rooted in individual minds. If constant rate effects are not caused by cognitive mechanisms in individuals, one would leave their emergence unexplained. Finally, the CRH is quite special in that it posits the truth of a null hypothesis (Paolillo 2011: 264). Null hypothesis significance testing has the potential to provide evidence for significantly different rates of change across contexts but cannot demonstrate that the rate of change is not different across contexts. Analytical advancements to address this issue in the context of logistic regression have been discussed in the literature (e.g., Ecay 2015: 36–40, 63–66; Kauhanen and Walkden 2019 for remarks on power analysis), but have not yet become standard practice. Hence, while this paper is careful to use state-of-the-art techniques in well controlled conditions to detect significant contextual divergences in the rate of change, it cannot resolve the fundamental problem that its methodology makes it impossible to prove conclusively the absence of such an effect.

Taken at face value, the CRH is a straightforward, empirically testable hypothesis. It can be refuted directly by observing that the rate of change depends on the grammatical contexts. In practice, the CRH comes with a number of conceptual and methodological challenges so difficult that the field has seen great confusion about the CRH’s interpretation and controversy about its applicability. Some researchers endorse the CRH, writing that “the notion of a constant rate has accumulated enough support over the last three decades for this to be referred to as the Constant Rate Effect” (Kauhanen and Walkden 2018: 484, referencing Pintzuk 2003). Others express great skepticism towards the CRH, stating that “[c]hanges may occur at different times and at different rates (pace Kroch 1989)” (Traugott and Trousdale 2010: 26) and that “the Constant Rate Hypothesis should not be considered valid unless tested on more reliable data” (Feltgen 2017:133). The purpose of this paper is therefore, firstly, to list common misconceptions surrounding the treatment of the CRH in previous case studies and, secondly, to present a new, improved enquiry that is free of the problems thus identified.

Section 2 will present a summary of the most important shortcomings of previous studies and derive a list of desiderata that an investigation of the CRH should observe. Section 3 reports an improved test of the CRH that meets those desiderata. Its empirical basis concerns possessive have in late Modern to Present-Day American English (c. 1800–today). I will outline the relevant linguistic developments as well as an elementary syntactic model, describe the database used for the collection of the material, the Corpus of Historical American English (Davies 2010), and evaluate the CRH using mixed-effect logistic regression models. Section 4 discusses the findings. Section 5 concludes.

2 Problems of previous studies and desiderata for a new study

This section discusses important requirements for an empirical test of the CRH.

3 A new test of the CRH

I will now present a study testing the CRH that meets the desiderata outlined in the last section.

3.1 Material: change in late modern English possessive have

The empirical material for this study concerns finite possessive have in late Modern American English (c. 1800–today). The term ‘possessive have’ is here defined by complementation pattern, as any instance of a finite form of the lexeme have co-occurring with a nominal object, have + DP. As such, the structure contrasts with other complementation patterns of have, perfect auxiliary, have + past participle (have done something), causative have, have + non-finite clause with an overt subject (have someone do something), its passive variant (have something done), modal have, have + to-infinitival clause (have to do something), etc. Semantic considerations regarding the denotation of the DP do not play a role for my definition.^[9]

The verb has changed its realization in several syntactic contexts. The diachronic variation is illustrated below. Example (1) exemplifies the conservative, example (2) the innovative, use of possessive have.

(1)

conservative variant:

“have you not still some water left in the sprinkling can!” “To be sure I have,” Uncle Wiggily whispered back.

(from: Star Tribune Journal, Minneapolis, July 11, 1916, p. 18)

(2)

innovative variant:

“Perhaps she is lonely. Do you not still have your sewin’ circle?” “We do, but Abigail says Loretta will not come”

(from: Marti Talbott’s Marblestone Mansion, Book 6, published 2013, p. 70)

In (1), not negates the proposition directly (have not), while in (2) negation requires do-support (do not have). Questions are formed by inverting have and the subject in (1) (have you), but by inserting the dummy auxiliary do in (2) (do you have). Example (1) places possessive have in front of a VP-adverb (have still); example (2) places it after such an item (still have). Finally, (1) forms an elliptical answer with have (I have), whereas (2), once again, uses the word do (We do).^[10]

The change in possessive have can be analyzed as follows. Modern English makes a fairly sharp distinction between two kinds of verbs, commonly called auxiliaries versus lexical main verbs. I will model the two types as two different word class categories and call the former I (“inflection”) and the latter V (“verb”) (following Chomsky 1986). There is an ordering restriction such that I occurs earlier in the clause than V. I formalize this in terms of a hierarchical constituent structure, where I dominates V, so that I is “higher” and V is “lower”.

These assumptions account for differential behavior of auxiliaries and lexical main verbs with respect to negation, inversion, VP-ellipsis (the first three of the so-called NICE properties, Huddleston 1976) as well as the placement of VP-adjuncts (e.g., Pollock 1989). The negator not is sandwiched between higher I and lower V. It follows that auxiliaries must be placed before not (3a) whereas lexical verbs are ungrammatical in this position (3b). Instead, they require do-supported negation, (3c), which, conversely, is not a grammatical option for auxiliaries, (3d).

(3)

a. Mary [ I may] not [ V study] for her exam.

b. *Mary studies not for her exam. (V before not)

c. Mary does not study for her exam.

d. *Mary (does) not may study for her exam. (not before I)

Inversion is a word order pattern that results from targeting material exclusively found in higher I and placing it in front of the subject. Therefore, auxiliaries can invert with the subject, (4a), whereas full lexical verbs cannot do so, (4b). Rather, do-support is needed in the absence of another auxiliary, (4c), but cannot be employed if another auxiliary is present, (4d).

(4)

a. [ I Was] his theory [ V accepted]?

b. * Became his theory accepted? (inversion of V)

c. Did his theory become accepted?

d. * (Did) his theory was accepted? (non-inversion of I)

Adverbs are typically adjoined to VP. Such VP-adjuncts are thus placed in front of finite transitive verbs, (5a), and cannot occur post-verbally before a verbal complement, (5b). However, this diagnostic is not as stringent as the negation or inversion tests for the distinction between auxiliaries and lexical verbs. While VP-adjuncts typically occur after auxiliaries, they are also acceptable in front of them, (5c). The former option, however, is considerably more common.

(5)

a. They [ I have] probably [ V visited] Spain.

b. * They have visited probably Spain. (VP-adjunct after V)

c. ok They probably have visited Spain. (VP-adjunct before I, rare)

VP-ellipsis, constructions involving an incomplete target sentence without a verb phrase reconstructed on a preceding, complete antecedent or source clause, requires the presence of higher I whereas lower V must be a part of the omitted material. As a consequence, VP-ellipsis structures can be licensed by an auxiliary, (6a), but not by a lexical main verb, (6b). If a source clause does not include an auxiliary, VP-ellipsis requires the dummy auxiliary do, (6c). In contrast, if a source clause involves an auxiliary, the target clause is degraded if it uses do-support instead of the contextually salient antecedent auxiliary (6d).

(6)

a.

You [ I were] [ V reading] a book.

And they were _, too./Or were you _?/But I was not _.

b.

You read a book.

* And they read _, too./* Or did you read _?/* But I did not read _. (V survives ellipsis)

c.

You read a book.

And they did _, too./Or did you _?/But I did not _.

d.

You were reading a book.

?? And they did _, too./?? Or did you _?/?? But I did not _.

(I does not survive ellipsis)

The four patterns differentiating lexical verbs and auxiliaries are exactly the four contexts in which late Modern American English possessive have has changed its realization. This is unlikely to be an accident. Hence, one should characterize the difference between lexical verbs versus auxiliaries and the change in possessive have in closely-related terms. Specifically, conservative possessive have should be grouped together with Modern English auxiliaries as an element of category I. Innovative possessive have should be grouped together with Modern English lexical main verbs as an item of category V. The change can then be understood as dual representation of possessive have as both I and V with replacement of the former with the latter during the transitional period. This is sketched in (7).

(7)

The present case study thus meets the first desideratum listed in the previous section: A grammar-theoretical model identifies a single linguistic change, the reassignment of the category of possessive have from I to V, with realization in four distinct environments, negation, inversion, VP-adjunction and VP-ellipsis. The CRH thus predicts that the innovation in possessive have will spread at identical rates of change in all four contexts.

Note that the relatedness of the contexts depends on the correctness of the aspect of the grammatical theory from which they are deduced. While it is impossible to prove that any particular formalism articulating a theory must be correct, I would like to point out that the analysis presented here is extremely consensual as far as such matters go in linguistics. This lends further credence to the claim that the change in possessive have falls under the domain of the CRH.

3.3 The data set

I will mention some noteworthy aspects regarding the final data set. Figure 1 presents a graphical representation of the data distribution for the negation context. The graph plots the proportion of innovative do-supported negation on the Y-axis against time on the X-axis. There is enough material for 198 data points, every year from 1810 to 2009 except for 1812 and 1813. The size of a data point is proportional to the overall number of tokens for that year. The graph includes a fitted line from a simple logistic model regressing the innovative form against time only as a guide for the eye illustrating the degree to which the data points conform to or diverge from a monotonic progression of the change.

Figure 1:

Distribution of possessive have data (negation context).

Innovative do not have completely replaces conservative have not and such substitution is likewise apparent in the other contexts. This satisfies the fourth desideratum, i.e., a replacement change.

The next desideratum, monotonicity, is slightly more complicated. There is an obvious S-shaped curve for the spread of the innovative variant of possessive have. Graphs for the other contexts reveal similar sigmoidal trajectories.

However, there are some outlier years between the late 1840s and c. 1890 with a greater-than-expected number of innovative forms also visible in the inversion and ellipsis contexts. These divergences can be explained as a result of register fluctuations in the corpus. COHA predominantly includes texts of standard American English, which is typically used by the upper, educated social classes. However, linguistic forms from other varieties intrude into this register. Fictional texts may quote or parody discourses that are not actually representative of the grammar of the authors, but of one of their literary characters’ manner of speaking.

A) The innovative variant of possessive have is often recorded in 19th century fictional, direct speeches of lower-class characters (Warner 1995: 547). This intrusion of innovative possessive have from low-prestige varieties into high-register writings substantially elevates the overall frequency of the innovative variant in the fiction genre during the early stages of the change. The data distribution is compatible with a change ‘from below’ (e.g., Labov 2007: 346).

To illustrate, the innovative negative possessive have construction in (14a), uttered by a slave hunter in Uncle Tom’s Cabin, also shows stigmatized negative concord (don’t never have no). Example (14b), uttered by a slave named ‘Cato,’ not only includes an innovative possessive have question but also distinct features of African-American Vernacular English, such as fortition of inter-dental fricatives to alveolar stops (e.g., this, the, thanks as dis, de, tanks), the reduction of final consonant clusters (e.g., and, last, expect as an’, las’, spec), and the realization of standard intervocalic v as b (heaven as hebben) (AAVE phonological features number 5, 1 and 12 in Rickford 1999: 4–5).

(14)

a.

They don’t never have no doubt o’ your meaning, Tom.

(UncleTomsCabin 1852)

b.

I is loss fum de cumpny, but dis is de ferry, and I spec dey’ll soon come. But didn’t we have a good time las’ night in Buffalo? Dem dar Buffalo gals make my heart flutter, dat dey did. But, tanks be to de Lord, I is got religion. I got it las’ night in de meetin.’ Before I got religion, I was a great sinner; […]. But now I is a conwerted man; I is bound for hebben;

(Escape 1861)

B) The conservative variant of possessive have is often found in direct speeches of late 20th century or 21st century historical fiction, where it likely serves as a depiction of a literary character’s upper socio-economic class or of a text’s ancient setting. These intrusions lower the overall frequency of the innovative variant in the fiction genre during the late stages of the change.

The following 21st century examples illustrate this effect. The conservative negative possessive have structure in (15a) is produced by a fictional upper-class mademoiselle set in 1840s New Orleans. The conservative direct possessive have question in (15b) portrays old-fashioned parlance imagined as typical of the early 19th century regency period in England.

(15)

a.

My mother has not the will – but you know her trials. I must wed without delay

(RoguesSalute 2007)

b.

“I can not feel a heartbeat.” Phaedra’s voice was surprisingly calm given the terror rising in her throat. “Have you a doctor close at hand, Mr. Kemble?” “Yes, yes, just round the corner.”

(TemptedAllNight 2009)

The influence of register fluctuations on the measurement of the change in the realization of possessive have can be related to the two reasons for deviations from monotonicity discussed in Section 2.1.5. Accidental data contamination due to inconstant inclusion of different text types might be instantiated here as an over-representation of fictional texts including direct speech in some years. Indeed, COHA’s proportion of fictional texts drops from c. 55% in 1850 to c. 40% by 1950. Real developments in the language system or culture that are adequately reflected in the textual record might manifest as shifts in the production of fiction, such as a decline in slave narratives after the abolition of slavery, or the emergence of the modern period drama.

Non-fiction texts document the change in a “purer” way in the sense that they consistently record Standard American English whereas fictional texts allow a glimpse into the wider use of the competing forms in more variable or marked styles. The imitations of, first, vernacular and, later, archaic uses of possessive have in fiction should raise and lower the base probability with which the innovative variant is encountered in fiction during the change’s early and late phase, respectively. This should result in a difference in the rate of change such that the change seemingly progresses more slowly in fiction, and more rapidly in non-fiction.

Importantly, the innovative form gains ground steadily independently of genre. Despite the register fluctuations, I therefore conclude that the change meets the fifth requirement of a largely monotonic increase as well.^[12]

Lastly, the change is attested by and large in its entirety, innovative possessive have being virtually non-existent before the 1820s and having reached a probability of ∼95% in the 2000s. The absolute end point of the change is not recorded in the data. This is true for all contexts except for VP-adjunction, where the earliest stages of the change are not recorded instead and the change has been at ceiling since the 1990s. The inception of the change may not be identified accurately since the corpus contains fewer documents for the early 19th century than later periods making it harder to pick up early instances of the innovative variant. Nevertheless, the data meets the sixth condition, complete attestation, as well since the corpus samples texts from at least the largest part of the transitional period.

4 Results

The diachronic variation in possessive have will be analyzed with mixed-effects logistic regression models (for accessible introductions, see e.g., Gries 2015; Johnson 2009). The model I propose includes the following variables:

1. The ‘Year’ predictor estimates the effect of time. It is a continuous variable based on the text dates provided by COHA. The variable has been standardized to z-scores.^[13]

2. The ‘Genre’ variable lets the intercept of the regression line vary between fiction and non-fiction (=COHA categories news, magazine, non-fiction).

3. The ‘Year:Genre’ interaction term allows for divergence in the slopes of fictional and non-fictional texts.

4. The ‘Context’ factor quantifies the difference between the grammatical contexts.

5. The ‘Year:Context’ interaction allows each context to change at different rates.

6. A random ‘Text’ variable controls for differences between individual texts.

The resulting model is presented in Table 1.

When the ‘Year’ predictor is transformed back on its original scale, the rate of change of 0.0552 log-odds per year (95%-confidence interval [0.0533–0.0571]) in the default genre, non-fictional texts, would imply that the innovative variant progresses from 1 to 99% of use in 167 years. The ‘Genre’ coefficient suggests that fictional texts are slightly more conservative than non-fictional texts. The rate of change for fictional texts, back-transformed on the ‘Year:Genre’ interaction, is only 0.0552–0.0108 = 0.0444 (95%-confidence interval [0.0423–0.0464]), predicting a time span of 207 years for a spread from 1 to 99% of use. This finding is compatible with the effect of register fluctuations proposed in Section 3.3. The ‘Context’ variable lists the differences between negation and the other contexts. VP-adjuncts have a substantially higher base probability of occurring with the innovative variant than the others. All of these variables are significant predictors for the realization of possessive have in this model. The results remain unchanged if another context is chosen as the reference category. The random ‘Text’ variable reports 9,706 individual texts aggregating the data, and a variance of the varying intercepts of 1.181. This suggests moderate clustering within texts (ICC = 26.4%). The model performs well, displaying a good model fit (Pseudo-R²_marginal = 0.635, Pseudo-R²_conditional = 0.731) and high classification precision (C-index = 0.941, classification accuracy = 86.6% vs. baseline: 55.2%). Therefore, this study also meets the seventh desideratum for a test of the CRH, an adequate statistical model.

I can at last check for identity in the rates of change in the four contexts. The ‘Year:Context’ interaction is non-significant for any pair of negation and a second context. This is true if any other context is made the default. The data is thus consistent with a scenario in which the true difference in the rate of change between the grammatical contexts is zero.

The state-of-the-art evaluation of the CRH uses a likelihood ratio test (LRT) comparing a full model with a term for the interaction between the ‘Year’ and ‘Context’ variables (as above) to a reduced model without this interaction term. Table 2 presents the result.

Model 1: (reduced model)

Have ∼ Year + Genre + Year:Genre + Context + (1 | Text)

Model 2: (full model)

Have ∼ Year + Genre + Year:Genre + Context + Year:Context + (1 | Text)

The addition of the ‘Year:Context’ interaction term does not significantly improve the model fit (assuming a chi-square distribution of the LRT statistic). The LRT thus provides no justification for its inclusion in the model.

Figure 2 plots the rise of innovative possessive have for each of the four linguistic contexts based on independent mixed-effects models, thus further highlighting the great similarity in the rates of change between them. The range between the highest and the lowest point estimates of the ‘Year’ coefficients returned by these models indicates that the time needed for the change to proceed from 1 to 99% should vary by 20 years in non-fiction and 18 years in fiction. To repeat, this effect size was not returned as significant with the sample size of this study and it bears repeating that the paradigm of null hypothesis significance testing cannot establish that an effect does not exist. However, assuming that this magnitude is in fact realistic, how could one assess whether it would be meaningful? An interpretation of this effect size cannot be fully objective but must rely on some reference frame within the research context. My view is that divergences of 20 or 18 years should be regarded as small because there are many complexities in measuring linguistic changes that could plausibly create noise of this magnitude, such as uneven distribution of tokens across contexts by different authors, especially where those are paired with uncontrolled social variables or sub-genres, editorial or prescriptive changes of texts, or dependencies of grammatical contexts on other factors.

Figure 2:

Data points, point estimates and 95% confidence intervals for the rise in innovative possessive have from independent mixed-effects regressions for each context.

In summary, the evidence adduced from the change in the realization of possessive have is fully compatible with the CRH.

5 Discussion

The findings of this paper have implications for several research questions in historical syntax. The most important consideration is this: Previous affirmations of the CRH could quite plausibly be rejected on the basis of small or low-quality data, inexplicit analyses or other complications (see 2.1.1–2.1.7). The present study avoids such problems by identifying and abiding by seven essential methodological desiderata for a test of the CRH. Even though the method of null hypothesis significance testing cannot determine how likely the data are to emerge if there is an effect of time–context interaction, the compatibility of the data with the absence of such an effect should lead to a strengthening of one’s confidence in the correctness of the CRH. In fact, it may be time to abandon the special status ascribed to the CRH as prima facie less plausible than its alternative and instead to treat it as a fallback position in the absence of evidence like any other null hypothesis. Once the CRH is viewed more widely as an ordinary null hypothesis, linguists should provide explanations when grammatical contexts do not develop at identical rates (language–internal interactions, cultural shifts etc.) rather than presuppose that rates of change may and will vary by default.

Another contribution of this study comes from the demonstration of a visibly S-shaped propagation of a linguistic innovation (see e.g., Figure 1). Such S-shaped patterns support models of diffusion that incorporate some kind of advantage of an innovative form such as certain cognitive constraints (selection) (e.g., Newberry et al. 2017) over random or purely interactional processes such as social diffusion (drift) (e.g., Blythe et al. 2009). This leads to the question of what such an advantage or “fitness” of a form may be. Perhaps the general problem can be approached from specific cases like the one discussed here. My own idea is that the advantage of an innovative string like don’t have the may lie in reduced syntactic ambiguity over a competitor string such as haven’t the. Other causes for the change have been proposed. Warner (1995) links the innovation to a re-set parameter after the loss of thou and the rise of do-support, which also resulted in other, roughly contemporaneous changes in the auxiliary system, such as the emergence of is being or the disappearance of elided verb phrases headed by be. Rohrbacher (1999: 196) believes that the change was motivated by the fact that the conservative position of possessive have does not usually assign theta-roles. Therefore, possessive have may have begun to cluster better with verbs like own or want than be or will. One could also point to traditional ideas of regularization or analogical pressure from semantically similar verbs (I own a house is to I don’t own a house as I have a house is to X). Moreover, the timing of the VP-adjunct context, changing about half a century earlier than the other environments (see Figure 2), mirrors very closely Early Modern English verb placement facts, where the loss of post-verbal VP-adjuncts likewise precedes the rise of do-support (e.g., Haeberli and Ihsane 2016). This cyclic behavior is curious and may be suggestive of a causal influence of the former on the latter development (some related discussion in Roberts 1985). Trudgill et al. (2002) argue that the adoption of do-support with have in American English is linked to a relatively low number of dynamic uses of have (e.g., preference for take a look over have a look in that variety). One causal explanation that is not easily compatible with the findings of this paper is the ‘Rich Agreement Hypothesis’ (e.g., Koeneman and Zeijlstra 2014 and references therein). It states that English verbs are placed structurally low as a consequence of the loss of verbal inflections. Possessive have did not undergo any notable changes in its inflections close to or during the time it came to appear structurally low. More work is needed to establish the specific advantage of innovative possessive have and the general nature of selection pressures giving rise to S-shaped diffusions in linguistic changes.

There are numerous further lines of inquiry that might be stimulated by the results of this paper. Some grammatical frameworks may be better equipped to deal with the specifics of the change in possessive have than others. The VP-adjunct context in particular might be predicted to be a context that should change in tandem with the others by some models of English clause structure but not by others. For instance, Culicover (2008) might find it difficult to accommodate the VP-adjunct facts. One could control for additional determinants that may influence the realization of possessive have and check if they affect the outcome for the CRH. Relevant additional factors include, for instance, semantic distinctions between concrete and abstract possession or between dynamic and stative situations (e.g., Lorenz 2016; Tagliamonte 2003). The discussion of the genre fluctuations (see Section 3.3) could serve as an impetus to investigate the life cycle of competitors in greater detail. It might be quite common for a conservative form to acquire formal, archaic or otherwise marked connotations towards the end of a change. It might be worthwhile to attempt a fresh reflection on the relation between lexical diffusion and the CRH. Since most syntactic changes concern abstract patterns instantiated by concrete lexical items, they may involve some degree of lexical diffusion. It seems important for those cases to come with explicit assumptions and analysis techniques. All of these questions seem to me worthy of pursuit in the future.

6 Conclusion

This study subjected the CRH to a new falsification attempt using an innovation in American English possessive have. It should, in my estimation, count as the most demanding one ever to have done so. The result showed that the data is compatible with the CRH. It would therefore be rational to update one’s prior belief about the CRH and move towards accepting it as an ordinary null hypothesis.