Lexical innovations are rarely passed on during one’s lifetime: Epidemiological perspectives on estimating the basic reproductive ratio of words Authors: Andreas Baumann, Søren Wichmann, Søren Wichmann, Søren Wichmann Published: 2024 ABSTRACT Lexical dynamics, just as epidemiological dynamics, represent spreading phenomena. In both domains, constituents (words, pathogens) are transmitted within populations of individuals. In linguistics, such dynamics have been modeled by drawing on mathematical models originating from epidemiology. The basic reproductive ratio is a quantity that figures centrally in epidemiological research but not so much in linguistics. It is defined as the average number of individuals that acquire a constituent (infectious pathogen) from a single individual carrying it. In this contribution, we examine a set of lexical innovations, i.e., words that have spread recently, in four different languages (English, German, Spanish, and Italian). We use and compare different ways of estimating the basic reproductive ratio in the lexical domain. Our results show that the basic reproductive ratio can be somewhat reliably estimated by exploiting estimates of lexical age of acquisition and prevalence but that the derivation based on diachronic corpus data comes with certain challenges. Based on our empirical results, we argue that the basic reproductive ratio can inform about the stability of newly emerging words and about how often such words are successfully propagated in linguistic contact events. Our analysis shows that an average lexical innovation that has spread in the previous two centuries has been passed on by each individual only to a handful of contacts. 1. Introduction Language change and epidemiological dynamics share some commonalities. This insight is not particularly new: already in the 1980s, quantitative laws of linguistic change have been inspired by epidemiological considerations [1, 2] and epidemiological models were transferred to the domain of cultural evolution [3]. Conceptually, what both domains have in common is that constituents, be they words (or other parts of linguistic knowledge) or pathogens, are transmitted within populations of individuals. Both, language change and epidemiological dynamics represent spreading phenomena [4–8]. This conceptual similarity has led to fruitful methodological transfer. So, logistic (S-shaped) dynamics, that are abundantly employed in epidemiology, function as a standard model in the study of language change [9–15]; and on a more general level, concepts from evolutionary theory (mutation, selection, drift) found various applications in the study of language change and evolution [12, 16–20]. Interestingly, one concept that is central to the field of epidemiology, the basic reproductive ratio, has received less attention in the study of language than it might deserve. In epidemiology, the basic reproductive ratio (or also: basic reproduction number, basic reproductive number), in short R0 (‘R nought’), is defined as the expected number of individuals in a population of individuals susceptible to a disease that are infected by a single infected individual [21–24]. That is, it measures how many secondary infections there are per newly introduced infected person. A central observation is that if R0>1 then a pathogen can spread, while it will, in the long run, go extinct, if R0<1. The basic reproductive ratio was explicitly applied to and formally derived for the lexical domain by Nowak and colleagues in their contributions in the Journal of Theoretical Biology [25]; with surprisingly low reception in the field of linguistics) and Nature [26], and later by Solé and colleagues [27], as well as in research on phonological change of our own [28]. In the linguistic domain, R0 measures the expected number of individuals that learn a newly introduced linguistic constituent (like a new word) from a single individual that already knows and uses that constituent, and the result about the consequences of R0 being greater or less than one applies–mutatis mutandis–as above. However, theoretical epidemiology offers more than that. First, it yields a threshold for herd immunity, i.e., a minimum fraction of individuals that cannot pass on a disease so that the disease cannot spread even if there are susceptible individuals [29–32]. This concept has a linguistic analogue. The herd immunity threshold can be interpreted as a minimum fraction of speakers resisting some linguistic change which is necessary for that change not to be successful in the end. Put differently, it can help to address the following question: how many speakers really need to reject, say, a certain word in order for the word to fall out of linguistic usage in the whole population? Second, and more fundamentally, the basic reproductive ratio, since it is defined as the expected number of successful secondary infections per primary infection during the whole infectious period, implicitly informs about contact dynamics. Again, the matter is relevant for the linguistic domain as well. If we know about the average size of personal contact networks, R0 informs us about the fraction of contacts in that network that really adopt a given linguistic innovation, such as a new word, from a single individual during the period of usage of that word. The latter restriction is crucial. It entails that R0 is dimensionless in the sense that the quantity is always put into relation with the duration of the period during which the innovation can be passed on in the first place. Put differently, R0 is not a rate and it does not tell us anything about the timescale of spreading events, nor does it inform us about how often such innovations are actuated in the first place. Obviously, one needs empirical estimates of R0 to derive such information. Theorems about how R0 can be estimated belong to the standard results of theoretical epidemiology. These results relate R0 to epidemiological and ecological concepts, in particular (i) age of first infection, (ii) equilibrium prevalence, and (iii) growth data [22, 23, 33, 34]. The aim of this contribution is to transfer these results to the study of language and to test them empirically in the lexical domain by estimating the basic reproductive ratio of words based on a simple model of the spread of linguistic constituents. It will be shown how the basic reproductive ratio of words relates to lexical prevalence data, age-of-acquisition ratings, and diachronic trajectories. The paper is structured as follows. In the following two subsections, we will first review fundamental epidemiological concepts and standard results (section 1.1), and then transfer them to the linguistic domain (section 1.2). After that, we will test predictions against diverse diachronic and synchronic linguistic data from four languages (English, Spanish, German, Italian; sections 2–3). In particular, we will see that deriving the basic reproductive ratio from diachronic data does not come without problems. Mindful of the limitations that the adopted model and the empirical approach have (section 4.1), we will relate our findings (a) to the stability of lexical constituents, (b) to the role that individuals and their personal networks play in lexical diffusion, and (c) to the relationship between linguistic prevalence, acquisition and change (section 4.2). In doing so, we do not test the analyzed model directly. Rather, the model functions as a vehicle from which different estimation methods are derived. In that sense, the underlying model is only tested indirectly by comparing its predictions against observed data. Based on our empirical analysis, we will conclude that an average newly emerging word is passed on by each individual only rarely, i.e., to a relatively small number of contacts not yet knowing the word that the individual encounters during their lifetime. 1.1 Key concepts and results in mathematical epidemiology This section provides a very brief overview of the formal derivation of the basic reproductive ratio in mathematical epidemiology. For an excellent and certainly more exhaustive review see Heffernan et al. [23] (2005) (as well as [35] for criticism). Epidemiologists are interested in the growth of the number of infected individuals within a population. Let us assume that we can divide the whole population into several pairwise disjoint subpopulations, or compartments, some of which consist of infected individuals. These compartments can reflect, e.g., social structure, age structure, or stages of a disease. Assuming that there are n compartments, let x≥0 be a n-dimensional vector that lists the number of individuals in each of the compartments. Let I and S denote the set of infected and not yet infected (susceptible) compartments, respectively. In what follows, we describe a way of deriving the basic reproductive ratio referred to as the ‘next generation method’ [36]. Let Fi(x) denote the rate of new infections that enter an infected compartment i∈I and let Ti(x) measure the rate of net-transitions that go out of compartment i. The rate of change of xi then is given by x˙i=Fi(x)−Ti(x). Given all rates Fi(x) and Ti(x), we can compute the ‘next generation matrix’ FT−1, where F=[∂Fi(xo)∂xj]i,j∈I and T=[∂Ti(xo)∂xj]i,j∈I. Here, x0 is the population dynamic equilibrium if there is no disease. The basic reproductive ratio then is given by the dominant eigenvalue ρ of the next generation matrix, i.e., R0 = ρ(FT−1). The intuition behind this definition is that reproduction depends (i) on the number of newly infected individuals and (ii) on how long individuals are infectious. The former is captured by F. The latter is captured by T, which collects all net-transition rates. Then, T−1 can be interpreted as measuring all waiting times before an individual exits all infected compartments. The derivation becomes easy and intuitively clear if there is only a single infected compartment. In the ‘susceptible-infectious-susceptible’ (short: SIS) model of infectious diseases the population is split into two compartments, susceptible and infected individuals [32]; for a review of the basic reproductive ratio in other model setups see van den Driessche (2017). In that case, S and I consist of one compartment each. Individuals can become infected or recover, in which case they switch back to the susceptible compartment. With a slight abuse of notation, let S and I denote the respective fractions of individuals. We assume that both subpopulations are homogeneously mixed. Then, SIS-dynamics can be formulated by means of a pair of ordinary differential equations S˙=−βSI+γI−μS+μI˙=βSI−γI−μI (1) where we assume I+S = 1, i.e., overall population size is normalized and kept constant. Here, β is the (per capita) rate of disease transmission and γ is the rate at which infected individuals recover. The demographic parameter μ denotes mortality rate (i.e., the duration 1/μ corresponds to one generation). Note that the term +μ in the first equation ensures that population size is constant. The disease-free equilibrium is (S,I) = (1,0), so that F=βS|S=1=β. The transition rate is T = γ+μ and the inverse T−1 = 1(γ+μ) is the expected waiting time until recovery takes place. Hence, R0 = β/(γ+μ). There are several ways of estimating the basic reproductive ratio empirically [23, 37, 38]. To begin with, it could be determined by estimating the model parameters it is defined by. Clearly, in the SIS model, estimates of β,γ, and μ suffice to determine R0. In practical terms, however, it is not always easy to obtain reliable estimates for β and γ, and this more generally holds true for more complex models [23]. There are three other interesting ways for obtaining R0 estimates. First, R0 can be estimated based on the prevalence of the disease at its population-dynamic equilibrium [32]. The argument goes like this: Let pI denote prevalence, i.e., the fraction of all infected individuals. Then for an infected individual, 1−pI is the probability of interacting with a susceptible individual, so that the number of infection events induced by that individual is R0(1−pI). At population-dynamic equilibrium when there is no change in the number of infected individuals, however, this number is exactly one and hence R0 = 1/(1−pI). We can retrace this derivation in the SIS-model by computing prevalence pI = I* (the star denoting a population-dynamic equilibrium). Thus, setting I˙=0 we find that pI=I*=1−γ+μβ=1−1/R0, which implies the above formula. Second, we can use age of first infection AoI to estimate R0 [22, 38]. Assume that in the population, all ages are homogeneously mixed and equally frequent. That is, the population follows a rectangular age distribution. Let LE denote expected lifetime (life expectancy). If we assume that, once infected, individuals cannot become susceptible again, then at population-dynamics equilibrium, AoI/LE individuals are not yet infected, and pI = 1−AoI/LE are already infected. The above considerations about equilibrium prevalence entail that R0 = LE/AoI. Third, the basic reproductive ratio can be estimated based on the intrinsic (exponential) growth rate r of a pathogen [23, 39]. In the early phase of an outbreak, intrinsic growth rate is implicitly given by I˙=rI. That is, given a trajectory of prevalence, one can estimate r. The basic reproductive ratio then is R0 = 1+rL, where L is the duration of the infected period. Irrespective of how the basic reproductive ratio is estimated, the herd immunity threshold can be computed as H=R0−1R0=1−1/R0, which turns out to equal equilibrium prevalence [29–32]. If the fraction of individuals in the population that cannot transmit the disease exceeds this threshold, the disease cannot spread and the fraction of infected individuals will approach zero. 1.2 Transferring concepts and results to linguistics The epidemiological concepts and results outlined above can be transferred to the linguistic domain. Here, in place of pathogens, there are linguistic constituents like words, phonemes, or constructions that are passed from one individual to another [4, 6, 8, 20]. Crucially, this is done through multiple interactions among individuals. Instead of infected and susceptible individuals, let us divide the speaker population into heterogeneously mixed users and non-users of a linguistic constituent [25, 28, 40]. Users already know and use the constituent of interest while non-users still have to acquire it. Let us define the respective fractions as U (for user) and N (for non-user). We can then define the system N˙=−βUN+γU−μN+μU˙=βUN−(μ+γ)U. (2) Here, β defines the per capita rate at which the constituent is transmitted from a user to a non-user. Let us refer to such events as instances of propagation. Upon successful propagation, the non-user switches to the user class. Note that β measures propagation rate under the condition that users and non-users interact, and not that a user utters the constituent. In fact, Nowak et al. (2000) [26] define this rate as β = bqϕ, where b is the number of contact events, i.e., linguistic interactions, in which learning can take place, q is the probability to adopt a constituent if encountered once, and ϕ is relative utterance frequency of the constituent. Hence, β can expected to be a function of utterance frequency, learnability, and the number of contact events. Note that the number of contact events can be much higher in linguistic propagation than in the transmission of pathogens since linguistic communication is restricted to physical contacts to a lesser extent. Due to the assumption of heterogenous mixing, the rate at which non-users switch to the user class is the product of propagation rate, the fraction of users, and the fraction of non-users (mass-action transmission; [23, 35]). The parameter γ defines the rate at which users stop using the constituent (for instance, because they forget it). It does not depend on contacts among individuals. As above, μ defines per capita mortality rate. Note that we assume all users subject to death to be added to the non-user class. This represents that new-born individuals are not yet equipped with the linguistic constituent. Apart from μ, age structure is not reflected in the model. In particular, heterogeneous mixing implies that interaction events are not conditioned by age. Propagation can occur in interactions with young individuals (reflecting the scenario of L1 acquisition in children that interact with their caretakers) as well as in interactions among old individuals (reflecting the adoption of words in everyday speech). All model parameters are summarized in Table 1. The model is schematically represented in Fig 1 (left). Compartment model. Left: schematic representation of the compartment model. Right: trajectories of logistic growth (green) and decline (magenta). Parameters: U(0) = 0.25 and γ+μ = 1 in both cases; β = 4 and β = 0.1 for the growing (green) and declining (magenta) case, respectively. 10.1371/journal.pone.0312336.t001Table 1Epidemiological and linguistic model parameters.ParameterEpidemiological readingLinguistic reading β Rate of successful disease transmissionRate of successful propagation of a constituent, depending on the average number of contacts, learnability, and utterance frequency γ Rate of recoveryRate of abandoning a constituent μ Mortality rateMortality rate Epidemiological and linguistic model parameters. As above, the basic reproductive ratio of this system is R0 = β/(γ+μ), and if R0>1 then it asymptotically approaches a stable population-dynamic equilibrium with U*=1−γ+μβ. That is, the basic reproductive ratio and the number of users at equilibrium increase as propagation rate increases, and as the rate of abandoning a constituent and mortality rate decrease. The ODE above models logistic, i.e., sigmoid, growth. Since N+U = 1, the system can be reformulated as U˙=U(β−γ−μ)(1−1U*). (3) It can be solved analytically to yield U(t)=U*1+(U*U0−1)e−rt (4) where U0 = U(t)>0 is the initial fraction of users [40]. Here, the intrinsic growth rate is given by r: = β−γ−μ, and U*=limt→∞U(t)=1−γ+μβ defines the carrying capacity of the system. The logistic equation was used abundantly to model sigmoid (S-shaped) growth in linguistics on various levels of linguistic analysis [2, 10, 13]. Clearly, if U0>U* and r>0 then the solution approaches the equilibrium from above, i.e., the curve is falling, and if U00 then the curve is growing. If r<0 then the solution approaches U = 0 (Fig 1, right). The considerations about transferring epidemiological concepts above suggest that standard results from mathematical epidemiology can be transferred to the linguistic domain as well [41]. More specifically, this means the following: First, the basic reproductive ratio of linguistic constituents can be estimated given estimates of β,γ and μ, or intrinsic growth rate r and any two of the three model parameters. Mortality rate μ can be estimated from demographic data (as μ = 1/LE). Adoption rate β and abandoning rate γ, however, are much harder to determine. Second, for linguistic constituents at their population dynamic equilibrium, the basic reproductive ratio can be estimated by means of prevalence, i.e., the fraction of users. That is, R0 = 1/(1−pU), where pU = U* is prevalence at population dynamic equilibrium. Third, for linguistic constituents at their population dynamic equilibrium, the basic reproductive ratio can be estimated by means of age of acquisition, i.e., the age at which the constituent was learned, so that R0 = LE/AoA, where AoA is the age of acquisition of the constituent. Fourth, we can use intrinsic growth rate to estimate the basic reproductive ratio as R0 = 1+r∙L where L = LE−AoA approximates the period during which an individual uses a constituent. This approximation only holds well if the constituent repertoire of individuals does not shrink substantially as they age. Note that, theoretically, all ways of estimating the basic reproductive ratio should yield identical estimates of R0. In particular, this should hold true for estimates based on age of acquisition and prevalence, since both share the same assumption, namely that the population dynamics rest at their equilibrium. Table 2 juxtaposes epidemiological and linguistic variables relevant for estimating the basic reproductive ratio. 10.1371/journal.pone.0312336.t002Table 2Epidemiological and linguistic variables relevant to the estimation of R0.VariableDescription AoI Age of first infection of a disease (epidemiological) AoA Age of acquisition of a constituent (linguistic)pI = I*Equilibrium prevalence, i.e., fraction of infected individuals at endemic equilibrium (epidemiological)pU = U*Equilibrium prevalence, i.e., fraction of users at population dynamic equilibrium (linguistic) r Intrinsic growth rate of a disease (epidemiological) or constituent (linguistic) R 0 Basic reproductive ratio:Expected number of individuals infected by an infected individual introduced into a population of susceptible individuals (epidemiological).Expected number of non-users that adopt a constituent from a single user of that individual in a population of non-users (linguistic). Epidemiological and linguistic variables relevant to the estimation of R0. Like the herd immunity threshold, we can derive, once R0 is estimated, the stability threshold H = 1−1/R0. In linguistic terms it denotes the minimum fraction of individuals which could potentially acquire a linguistic constituent that need to be excluded from adoption for that constituent to vanish in the long run. If there were, for instance, hypothetical political measures for banning a certain word, one would need to ensure that at least H individuals are not able to adopt that word. Conversely, if, in this hypothetical scenario, 1−H individuals would refuse to adhere to these measures, the spread of the word could not be prevented. Note that this does not consider social structure in any way. We will come back to this point in the discussion section. 2. Materials and methods All computations described in this section were done with R (4.4.1). Data and code for this study are available at https://gitlab.com/andreas.baumann/basic_reproductive_ratio. 2.1 Data We test the predictions outlined above on the lexical level, i.e., by inspecting dynamics of words. We restrict our diachronic analysis to words that have increased in frequency during the past two centuries, and we will refer to such words as lexical innovations. We do so irrespective of whether these words have already been present initially at low frequency, whether they are associated with new concepts or are involved in onomasiological competition, whether they are new word formations or have been adopted from a different language. Moreover, we only consider periphery words, which we define, somewhat simplistically, as words that are not used by everyone, based on prevalence estimates. For obtaining estimates of the basic reproductive ratio for these words, we combine different data sets in different languages: English (eng), Spanish (spa), German (ger), and Italian (ita). We use AoA estimates for a large set of English from Kuperman et al. [42]. AoA norms for about 4600 Spanish verbs were taken from Alonso et al. [43]. For German, we used data from Birchenough et al. [44]; 3,200 words), and for Italian, data from Montefinese et al. [45] was employed. All of these estimates refer to word forms (there are sometimes separate AoA norms for inflectional and/or derivational forms of a base form such as accelerate, accelerated, acceletation, etc. in [42]) and they were obtained by averaging subjective age-of-acquisition ratings collected via (crowdsourcing) surveys. Although ratings are subjective, estimates obtained by means of this methodology were shown to correlate with age-of-acquisition estimates obtained under controlled conditions [42, 46]. Lexical prevalence was proposed as a relevant covariate for psycholinguistic research [47], so that prevalence estimates are already available for a set of languages. Prevalence estimates for English word forms were taken from Brysbaert et al. (2019). These estimates were aggregated via crowdsourcing as well. Note that prevalence is not probit transformed as in Brysbaert et al. [48]. Rather, for a word, the prevalence estimate measures the fraction of individuals participating in the survey knowing the word. Baumann and Sekanina (2022) show that there is a slight difference between the fraction of individuals knowing a word (only passive knowledge) and the fraction of individuals using a word (active knowledge). The latter operationalization, as a matter of fact, fits better to prevalence as introduced in the previous section as the fraction of users. Overall, however, both estimates correlate strongly so that the estimates from Brysbaert et al. (2019) will serve as good enough proxies for the present study. For the other three languages (ger, spa, ita), we approximated prevalence by computing, for each word, the fraction of participants reporting a subjective AoA rating [43–45]. If no rating was given or if participants reported that the word was unknown this was interpreted as neither knowing nor using the word. Note that the accuracy of the prevalence estimates substantially depends on sample size, i.e., on the number of participants they are based on. The respective sample sizes are about 300 for English, about 23 for German, about 50 for Spanish, and 25 for Italian. Hence, maximal margins of error (for p = 0.5 at a 95% confidence level) range from about ±0.03 (for English) to slightly over ±0.1 (for German). For measuring growth, we use word frequencies derived from diachronic corpus data. For English, we derive frequency trajectories from the Corpus of Historical American English (COHA; Davies 2010) for all words in the intersection of Kuperman et al. [42] and Brysbaert et al. (2019). This corpus spans the period from 1820 to 2000. Trajectories have a resolution of one frequency estimate per decade. Frequencies are normalized with respect to the decade-wise sub-corpus size. The main advantage of COHA clearly is that it is balanced with respect to genre. For German, Spanish, and Italian, we needed to resort to normalized frequency trajectories extracted from Google books (ngrams; https://books.google.com/ngrams/). For the sake of consistency, exactly those word forms represented in the respective AoA lists were chosen in all queries for both COHA and Google books. This is, because the AoA list in [42] features estimates for separate inflectional/derivational word forms and does not differentiate between word classes. Thus, there are, e.g., separate AoA estimates for abbreviate and abbreviated, but it is not clear whether the latter corresponds to the past-tense form of the verb or to the adjectival participle. Hence, we did not differentiate between word classes in the corpus data either and aggregated frequencies of all word forms matching the entries in the AoA list per decade in COHA. In Google books, frequency trajectories have a resolution of one estimate per year. Frequencies were aggregated per decade, however, so that the resolution matches that of COHA. Table 3 shows a summary of all data sources used in this study. Fig 2 displays prevalence estimates, age-of-acquisition estimates, and historical frequency trajectories of four lexical examples. Lexical innovations. Examples for all considered languages: blasé (eng), europeizar (spa, ‘europeanize’), Charta (ger, ‘charter’), and cracker (ita, ‘cracker’): trajectories of normalized frequency (cf. 2.2.1) and pie charts illustrating age of acquisition relative to life expectation (top) and prevalence (bottom). 10.1371/journal.pone.0312336.t003Table 3Data summary for all languages (eng, spa, ger, ita).VariableLanguageCommentSourceAoA & prevalenceeng30,000 words[42, 48]ger3,200 words[44]spa4,640 words[43]ita1,957 words[45]Frequency (diachronic)spa, ger, itaTrajectories (years aggregated per decade)Google ngrams accessed with ngramr [50]engTrajectories (decades)COHA [49] Data summary for all languages (eng, spa, ger, ita). 2.2 Methods 2.2.1 Operationalization The formulae in Section 1.2 are used to estimate the basic reproductive ratio based on (i) age of acquisition, (ii) prevalence, and (iii) diachronic trajectories. We refer to (i-iii) as estimation methods. For (i-iii) we restrict ourselves to words with a prevalence smaller than 0.95, i.e., we exclude the core lexicon. In addition, for (iii), we focus on lexical innovations, which we conceptualize as words that have been increasing significantly in frequency throughout the observation period (see below for more details). We use the collected age-of-acquisition estimates, an estimate of life expectancy based on demographic data (https://data.worldbank.org/), and estimate the basic reproductive ratio as (i) R0AoA=LE/AoA. Note that life expectancy has increased through the past two centuries and that women have, on average, a higher life expectation than men, so that the assumption of a single point estimate is clearly simplistic. In the UK and the US, for instance, life expectancy has increased from about 71 in the early 1960s to about 80 in the 2000s (Fig 3, left). Notably, UK/US displays the smallest increase in life expectancy. For all populations, mean LE is roughly 75 with an average margin of error of 1 year (95% confidence interval) in this time period. This margin of error will be considered in the error analysis in 2.2.2. Furthermore, the estimation of the basic reproductive ratio depends on the age distribution being roughly rectangular (i.e., relatively evenly distributed; see above). Such a distribution is typically adopted in high-income countries (Fig 3, right) [31, 32]. In particular, this assumption seems to be met for the English-speaking population. More straight forwardly, we can use prevalence estimates pU to derive (ii) R0p=1/(1−pU). Demographic information for the relevant speaker populations. Left: life expectancy (LE) depending on time with mean LE (from left to right: eng, spa, ger, ita). Data taken from https://data.worldbank.org/. Right: age distributions for UK and US combined (eng), Spain (spa), Germany (ger), and Italy (ita) as representatives of the respective speaker populations. Data taken from https://data.un.org/. Both of these estimates rest on the assumption that the population dynamics are at their equilibrium, i.e., that the fraction of users neither grows nor falls substantially. For that reason, we compute for each word the absolute value of the Spearman correlation coefficient of time and frequency, ρ≥0 from 1950 to 2000. We use S = 1−ρ as a measure of contemporary stability. If S it is close to 1 for a word this indicates that the dynamics of that word are relatively stable and rest at their dynamic equilibrium. We will use this to examine the impact of S on the quality of the estimates of the basic reproductive ratio. Using frequency data to estimate the basic reproductive ratio requires a bit more work. The goal is to use the approximation R0≌1+r∙L where r and L denote intrinsic growth rate and the duration of the prolific period, respectively. As a matter of fact, to get r we need a prevalence trajectory, i.e., a trajectory of the number of users U(t) over time t∈T, as opposed to a frequency trajectory. Here, T={t0,t1,…,t18} is the observation period divided into 18 decades (from 1820 to 2000). It is not at all easy to obtain word prevalence based on corpus data. This is because corpora often feature only a small number of authors and because the amount of text per author is likely not representative of their linguistic knowledge (let alone the fact that the author population in linguistic corpora is not necessarily representative of the whole population, in particular, as far as the social dimension is concerned). See Johns et al. [51] for an exploration of ways to estimate prevalence based on corpus data. In the present study, however, we adopt a different approach to reconstruct prevalence trajectories for a subset of words. For each word, we take its frequency trajectory f(t) and we restrict our set of words to those that have a present-day prevalence pU<0.95. For each word we check if it has increased significantly based on a Spearman’s correlation test (correlating U(t) and t at α = 0.05). Furthermore, we derive the stability coefficient S as above. Subsequently, we normalize each frequency trajectory with respect to its maximum, i.e., f˜(t)=f(t)/maxf(t), and we define a word’s prevalence trajectory as U(t)=pU∙f˜(t). That is, we do as if the frequency trajectory would mirror the trajectory of prevalence. We only considered words with U(t)≤1 for all t and significantly increasing prevalence trajectories (i.e., words that we consider lexical innovations in this study). In a final step, we fit a logistic model U^(t)=a/(1−b∙e−rt), a,b>0, as defined in Section 1.2 to the trajectory U(t) in order to estimate intrinsic growth rate r and define (iii) R0r=1+r∙L, where we take L = LE−AoA as the prolific period of the word. This estimate was not computed for words with 1+r∙L<0, since the basic reproductive ratio is conceptually non-negative. For each model, we compute R2=1−∑t∈T(U(t)−U^(t))2∑t∈T(U(t)−U(t)¯)2 (5) to assess goodness of fit, where U(t)¯ denotes mean prevalence across all t∈T. Finally, the inflection point tI is extracted from each model. This point divides the trajectory into an initial (accelerating) and final (decelerating) period and helps to identify when the spreading phenomenon took place, at least approximately. Of course, this approximation is not unproblematic. First, it is likely that the prevalence trajectory of a word in fact precedes that of frequency because a word can also spread, language internally, through various possible linguistic contexts. That is, r and hence R0r is potentially underestimated (because r should be greater than frequency developments suggest). On another account, life expectancy has changed considerably over the past two centuries so that LE and hence R0r is in fact lower than assumed. Perhaps these biases balance each other, but this still needs to be tested. Finally, age of acquisition is likely not constant over time. See Section 4.1 for discussion. Table 4 lists the sample sizes used for all ways of measuring the basic reproductive ratio in the four languages, considering all restrictions mentioned above (periphery words; lexical innovations). 10.1371/journal.pone.0312336.t004Table 4Sample sizes for all languages (eng, spa, ger, ita) and all estimation methods (i-iii).Language R0AoA R0p R0r totalEnglish1249412512177412512German78781978Spanish136113613301361Italian3232132 Sample sizes for all languages (eng, spa, ger, ita) and all estimation methods (i-iii). 2.2.2 Statistical analysis We inspect the quality of the obtained estimates for the basic reproductive ratio, and hence the quality of the estimation methods, in three ways. First, we investigate, for each language, to what extent estimates gained via (i) age of acquisition, (ii) prevalence, and (iii) growth fall into a similar range. Second, we assess to what extent estimates provided by methods (i-iii) correlate. Finally, a derivation and analysis of the margin of error is provided for all estimates of the basic reproductive ratio. In the first part, we compute medians of R0 estimates for all methods (i-iii) in all languages, and assess if they differ from each other (Wilcoxon signed-rank test with ES=Z/N as effect size, where N is sample size [52, 53], as well as their pairwise correlations (note that correlations with R0r cannot be computed for Italian due to the lack of words with the required diachronic profiles). Second, we model pairwise associations between R0r,R0p, and R0AoA, controlling for stability S. For each language, we fit a Gaussian model (i+ii) R0AoA=b0+bpR0p+bpSR0pS+ϵ, and for all languages except for Italian, we fit two additional Gaussian models, (ii+iii) R0r=b0+bpR0p+bpSR0pS+ϵ and (i+iii) R0r=b0+bAoAR0AoA+bAoASR0AoAS+ϵ. In the latter two models, data points are weighted by goodness of fit (R2) of the temporal trajectory the respective growth rates were estimated from. This is to demote the impact of unreliable R0r estimates when assessing the relationship with the other two measures, R0p and R0AoA. Stability is controlled for because estimates of R0p and R0AoA rely on the assumption of the respective population dynamics resting at their equilibrium. Hence, correlations among measures of R0 are expected to be higher when this assumption is met, i.e., when stability S is high (close to 1). Italian was excluded in models (i+ii) and (ii+iii) involving R0r as outcome variable, because there was only a single valid estimate derived from diachronic frequency data (see above). The formulas used for estimating the basic reproductive ratio, (i) R0AoA=LE/AoA, (ii) R0p=1/(1−p), and (iii) R0r=1+r∙L=1+r∙(LE−AoA), involve nonlinearities and multiple parameter estimates that need to be considered when providing an estimate of the margin of error of R0AoA,R0p, and R0r. This is important since life expectancy, age of acquisition, and prevalence represent estimates themselves. We employ error propagation through the first-order Taylor expansion of R0AoA,R0p, and R0r to obtain the respective margins of error, ΔR0AoA,ΔR0p, and ΔR0r, i.e., one half of the corresponding 95% confidence interval [54, 55]. More specifically, we have ΔR0AoA=(∂R0AoA∂LE∙ΔLE)2+(∂R0AoA∂AoA∙ΔAoA)2=(ΔLEAoA∙)2+(LE∙ΔAoAAoA2)2, (6) ΔR0p=(∂R0p∂p∙Δp)2=Δp(1−p)2 (7) and ΔR0r=(∂R0r∂r∙Δr)2+(∂R0r∂LE∙ΔLE)2+(∂R0r∂AoA∙ΔAoA)2=((LE−AoA)Δr)2+r2(ΔLE2+ΔAoA2) (8) where ΔLE,ΔAoA,Δp, and Δr are the respective margins of error of life expectancy, age of acquisition, prevalence, and growth rate. For life expectancy, we assume a margin of error of 1 (see 2.2.1). For age of acquisition, the standard error is computed as SE=σ/NAoA, where for each word standard deviation σ and sample size NAoA are directly taken from the four age-of-acquisition datasets. For prevalence, the standard error is computed as SE=p(1−p)/Np. Here, sample size Np is again inferred from the employed datasets. Note that in the case of English, NAoA and Np differ from each other since measures were taken from two different resources [42, 48]. For age of acquisition and prevalence, the margin of error is then computed as 1.96∙SE. Obtaining the margin of error Δr requires a little more work since growth rate r is estimated from a logistic model fitted to a trajectory and not derived from a closed formula. Moreover, the trajectory U(t)=p∙f˜(t) itself depends on estimates of (contemporary) prevalence weighted by occurrence frequency in diachronic corpora, as described above. We opted for bootstrapping the margin of error for each word-specific growth rate r in the following way. For each time step t, we sampled n = 300 proportions from a binomial distribution B(Np,U(t)). This resulted in n trajectories U˜(t) randomly sampled around U(t). Each of these trajectories was in turn used to fit a logistic growth model as described above in 2.2.1, from which a logistic growth rate r˜ was derived. Subsequently, the distribution of all logistic growth rates r˜ obtained in this manner was used to compute the bootstrapped 95% confidence interval of the estimate r. The margin of error Δr was finally taken as one half of the width of this confidence interval. This was done for each word for which growth rate was estimated in 2.2.1. Note that this procedure still ignores the fact that occurrence frequencies used for constructing trajectories represent point estimates as well. Put differently, we need to assume (a) that all employed text corpora are so large that the corresponding errors vanish (which is plausible), and (b) that the employed text corpora are representative samples of the examined languages (which is arguably debatable). The obtained margins of error were then used together with Eqs (6–8) to compute the margins of error ΔR0AoA,ΔR0p, and ΔR0r for each word. Finally, the distribution of inflection points tI was examined descriptively for each language in order to assess when the observed lexical developments took place. 3. Results The analysis of expected values for the basic reproductive ratio (for all languages and methods) shows that estimates typically range from about 3 to about 14 (Fig 4; inter-quartile ranges go from at least about 2 (ger: R0r) to at most about 17 (ger: R0p)). Medians are shown in Table 5. There are no clear biases visible as far as the methods are concerned. Using prevalence yields higher estimates for the basic reproductive ratio than using age of acquisition does. This is the case for English (small effect) and Italian/German (large effects), but not for Spanish. Using growth rates to estimate the basic reproductive ratio yields smaller estimates for three languages (Spanish, German, Italian), but not for English. Moreover, the distributions of estimates based on growth for English and Spanish display a long right tail with R0r going well above 20 for a set of words. Most estimates fall below about 14 in both languages, however. German prevalence-based estimates are higher than this is the case for all other languages, and the assessment of the range of R0r in Italian is obviously imprecise given that there is only a single point estimate. Is worth to point out that R0AoA estimates seem to be relatively consistent in the sense that average estimates for all four languages fall into a small range (from 5.7 to 7.2). Of all four languages, English produces the most consistent estimates ranging between 5.7 and 6.2 with only small differences in terms of normalized effect size. Overall, there seems to be a positive relationship between consistency and sample size, English and Spanish yielding more consistent estimates across the three methods than the other two languages with considerably smaller sample sizes. Distributions of all R0 estimates in all languages. White circle: median; black bar: inter-quartile range. Pairwise differences (normalized effect size) are displayed to the left of each violin chart (‘*’ indicates significantly non-zero differences at α = 0.05). 10.1371/journal.pone.0312336.t005Table 5Median estimates for all methods (i-iii) of estimating the basic reproductive ratio in all languages.Languagemedian R0AoAmedian R0pmedian R0rEnglish5.75.96.2German7.214.02.6Spanish6.05.03.2Italian7.012.55.4 Median estimates for all methods (i-iii) of estimating the basic reproductive ratio in all languages. Fig 5 shows pairwise correlations among the three measures for English, Spanish, and German, but without taking differences in stability into account. Relationships among R0p and R0AoA are positive in all three cases (mildly positive correlation coefficients between 0.19 and 0.30). Distributions of inflection points and correlations among all R0 estimates in three languages (eng, spa, ger). Left: boxes show interquartile ranges of inflection points tI per language. Dashed lines represent ranges of participant age (displayed as birth dates) in each of the three age-of-acquisition surveys (see Table 3). Right: Pearson correlation coefficients among all R0 measures per language (‘*’ indicates significantly non-zero differences at α = 0.05). The results of the models accounting for pairwise relationships among R0p and R0AoA including an interaction with stability are shown in Fig 6. The relationships are positive for English, Spanish, and German, which is what one would expect, given the theoretical considerations. In the case of Italian, no clear relationship can be found. Interestingly, the association between R0p and R0AoA is not substantially modulated by the stability measure S. Words that are learned early are also those that are used by a large fraction of individuals, irrespective of whether they have been used stably in the past decades. However, at least for English and Spanish, a significantly positive interaction coefficient bpS can be determined. That is, the positive relationship between R0p and R0AoA is slightly strengthened as stability S increases. So, stability seems to have a small but positive effect on the quality of the estimates. This is what one would expect given the assumptions of the formulas for deriving R0p and R0AoA. In both cases, S needs to be sufficiently high. In other words, a lacking positive relationship between ways of estimating the basic reproductive ratio for low-stability words, e.g., for words that contemporarily show an increasing trajectory, is in fact not at all surprising. The models assessing the relationship between R0r and R0AoA and R0r and R0p reveal a different picture (Fig 7). In fact, weakly positive associations can only be found for English, and only for high-stability words. For words that are either in the process of growing or declining (low S), the relationship is even reversed. For the other languages, no statistically robust effects can be found. All model coefficients are listed in Tables 6–8. Models (i+ii). Linear models (i+ii) of R0AoA depending on R0p controlled by stability S with 95% confidence regions. For all languages but Italian there are statistically robust positive associations between R0AoA and R0p. In each model visualization, the region close to S = 1 corresponds to those words that are contemporarily stable (R0AoA and R0p more reliable). Models (i+iii) and (ii+iii) for English. Linear models (i+iii) of R0r depending on R0AoA (top panel) and (ii+iii) of R0r depending on R0p (bottom panel), controlling for stability S, with 95% confidence regions. Each visualization is split into three panels, reflecting increasing stability (from left to right). In the case of English, there are significantly non-trivial interactions: positive effects for high S, and negative effects for low S, as expected. 10.1371/journal.pone.0312336.t006Table 6Models (i+ii) with R0AoA as dependent variable.LanguageCoefficientEstimateSEp-valueeng b 0 5.4590.028< 0.001* b p 0.0810.004< 0.001* b pS 0.0140.0060.009*spa b 0 5.7570.028< 0.001* b p 0.0600.004< 0.001* b pS 0.0370.0140.004*ger b 0 5.2560.764< 0.001* b p 0.1660.061< 0.01* b pS 0.0550.0620.373ita b 0 6.9870.605< 0.001* b p 0.0510.0590.396 b pS -0.1140.0870.203 Models (i+ii) with R0AoA as dependent variable. 10.1371/journal.pone.0312336.t007Table 7Models (i+iii) with R0r as dependent variable.LanguageCoefficientEstimateSEp-valueeng b 0 11.8591.026< 0.001* b AoA -0.3610.1660.030* b AoAS 1.1720.154< 0.001*spa b 0 7.2551.813< 0.001* b AoA -0.2710.3030.373 b AoAS 0.2430.1620.136ger b 0 10.5214.9570.050* b AoA -0.8730.7550.265 b AoAS -0.2000.5230.707 Models (i+iii) with R0r as dependent variable. 10.1371/journal.pone.0312336.t008Table 8Models (ii+iii) with R0r as dependent variable.LanguageCoefficientEstimateSEp-valueeng b 0 13.2760.489< 0.001* b p -0.3400.056< 0.001* b pS 0.5890.111< 0.001*spa b 0 6.2300.493< 0.001* b p -0.0590.0570.301 b pS 0.0990.1830.585ger b 0 2.0811.7270.246 b p 0.1770.1390.221 b pS 0.3930.3080.220 Models (ii+iii) with R0r as dependent variable. Fig 8 shows the distribution of margins of error across all languages and estimation methods. In the case of English and Spanish, margins are relatively small (with median scores not exceeding 1.28 for English and 2.77 for Spanish), irrespective of how it was estimated. For example, the confidence interval for the basic reproductive ratio of an average English word in the dataset goes from about 4 to about 8. The German and Italian data, however, reveal large margins of error when the basic reproductive ratio is derived from prevalence. This is a direct consequence of the small sample sizes in the German (23 data points per prevalence estimate) and Italian data (25 data points per prevalence estimate). Across all languages, acquisition-based estimates are most accurate (median scores of the margin of error not exceeding 1.17). Distributions of the margin of error of the basic reproductive ratio across all languages and estimation methods. Margins of error based on a 95% confidence level. In the Italian data, only a single growth rate was derived, hence no distribution is shown for this configuration. Displayed numbers and white circles represent median scores. Thick black bars indicate inter-quartile ranges. 4. Discussion 4.1 Generalizability and limitations The observation that R0 estimates derived through AoA and prevalence correlate for all languages (except Italian) is reassuring but not particularly surprising, for that matter. We will discuss the theoretical implications of this relationship below. What is surprising, however, is that the positive relationship between both estimates is not substantially modulated by the stability of the recent development. This is interesting because contemporarily unstable, i.e., growing or declining words, are not expected to yield reliable R0p and R0AoA estimates. Contemporarily unstable words do not fulfil a crucial assumption for deriving these measures (cf. 1.2 and 2.2.1). However, at least for English and German the interaction with stability points into the theoretically expected direction (albeit at a low effect size). The general lack of a positive correlation seems surprising given that R0r was also defined by means of prevalence (because trajectories were normalized to fit present-day prevalence) and age of acquisition (because the prolific period is constrained by age of acquisition) in the first place. The observation that most average R0 estimates fall into a similar range, across languages as well as across estimation methods, is reassuring as well and potentially more revealing. While it can be expected that, across languages, scores derived with the same method fall into similar ranges, it is potentially more surprising that empirically estimated ranges indeed roughly coincide when making comparisons across methods. It can be seen that different estimation methods produce more similar results for English, followed by Spanish. They show certain discrepancies for German and Italian where prevalence-based estimates are higher than the other two estimates (large effect sizes in all cases for German and Italian). It is plausible that this can be attributed to the size and quality of the German and Italian data. Both datasets are relatively small, perhaps biased towards the core lexicon, and prevalence estimates were derived only indirectly (by considering missing AoA scores) and based on a smaller number of participants per word (about 25). Indeed, prevalence-based estimates for German and Italian are almost rendered useless by their, on average, large margins of error. We learn from this that using prevalence for the estimation of the basic reproductive ratio seems to only make sense if prevalence estimates are based a sufficiently large number of participants (about at least 50 as in the case of Spanish). The approach of estimating R0r employed here suffers from a multitude of problems. First, it is very likely that age of acquisition, just like life expectancy, was not constant through the past two centuries. This is so, because age of acquisition is a function of frequency even if one controls for prevalence [48]. It could be that individuals living, say, one century ago were exposed much more rarely to words that are nowadays used frequently; not because these words were employed by fewer individuals but because they displayed low utterance frequency per se (for instance because there was a relatively more frequent competing synonym available). By consequence, historical age of acquisition would be higher, hence lowering R0r for low-frequency words (because the prolific period is shortened in this way). That is, our approach potentially overestimates R0r for low-frequency words. The only way of mitigating this issue is to obtain reliable historical age-of-acquisition estimates. Evidently, subjective ratings are not available for early periods so that one needs to rely on reconstructed estimates, ideally working with evidence independent from frequency. An approach that might prove useful in this regard is that of estimating age of acquisition through simulating semantic learning. Botarleanu et al. [56] have exploited word embeddings trained on incrementally increasing corpora to approximate age of acquisition through semantic trajectories. Since this approach could be applied to historical corpora as well it could be used to approximate historical age of acquisition norms. Note, though, that word-embedding similarities were shown to depend on frequency as well [57] so that this factor needs to be considered carefully. It is work pointing out, however, that judging from the range of inflection points of the estimated diachronic trajectories (Fig 5, left), the majority of all innovations covered in this study seem to represent relatively recent phenomena, mostly confined to the 20th century. The interquartile range of inflection points largely overlaps with the range of birth dates of the participants in the three age-of-acquisition surveys. In the case of English and German, participants seem to have witnessed the emergence of the examined innovations (in the case of Spanish, inflection points largely precede participant birth dates, though). Note that the actual inflection points could be overestimated due to the time lag between writing and publishing texts [58]. This affects Google Books (spa, ger) and, to a lesser extent, COHA (eng). Second, and related to that, we used normalized frequency as a proxy of prevalence for reconstructing prevalence trajectories. While it was shown that lexical prevalence and frequency are correlated significantly [47, 59], the two measures do not match (for instance, computing Pearson’s r for frequency and prevalence for all words in Brysbaert et al. [48] known by less than 95% of all individuals yields a correlation of R = 0.28. For prevalence thresholds of 80% and 50%, one yields R = 0.16 and Rr = 0.12, respectively). Notably, this relationship becomes weaker for lower prevalence values so that using frequency as a proxy does not seem to be a very reliable strategy when examining dynamics of lexical innovations that are initially rare by definition. More reliable historical prevalence estimates could be obtained by inspecting the fraction of authors in large-scale corpora that employ a particular word, an approach already pursued by Johns et al. [51] and Feltgen [60]. Although promising, this approach evidently suffers from data scarcity in earlier periods: for estimating prevalence in this way, one ideally needs both a large number of authors and a large and representative selection of texts per author, and prevalence will be necessarily underestimated [60].There is an interesting mismatch between the lacking correlation between R0r and R0AoA on the lexical level, and that on the phonological level that we have demonstrated to hold for the dynamics for English word-final consonant clusters [28]. The obvious advantage of phonological diachronic analyses is that prevalence estimates are more accurate: before the onset of certain sound changes in the history of English, prevalence of (most) English word-final consonant clusters was exactly zero, and upon completion prevalence can be safely assumed to be close to one for the majority of such sound sequences. Estimating intrinsic growth then amounts to inspecting the steepness of the growth curve; a task involving fewer degrees of freedom. Third, the nature of the employed historical corpus data could have affected the R0r estimates. In contrast to COHA, which consists of a genre-stratified sample of texts throughout (most of) the observation period, the Google Books corpus was shown to be increasingly biased towards scientific texts [58]. As a result, growth rates r and hence R0r could be overestimated for scientific words and, consequently, underestimated for non-scientific words for Spanish and German (Italian was not included in the diachronic analysis, anyway). This also entails, that some non-scientific words were missed out in the present analysis due to not displaying significantly increasing trajectories. On another account, this bias might also explain why R0r is significantly lower than R0p and R0AoA in the case of Spanish and German. A post-hoc comparison of English frequency trajectories in COHA and Google Books (see attached code for details) suggests that the great majority of trajectory pairs (78.8%) are relatively similar (strong and significant correlation between the pair of trajectories) while only a few cases (1.6%) yield opposing trends (significant negative correlation between the pair of trajectories). However, biases towards scientific vocabulary could still affect the estimates. With that in mind, conclusions drawn from English corpus data (i.e., COHA) seem to be more robust. Fourth, the approximation of the prolific duration as L = LE−AoA rests on the assumption that individuals do not forget words. Estimates of vocabulary size depending on age could make the approximation more robust (and, at the same time, inform us about the rate γ in the model given in Eq (2)). Extant research on the relationship between age and vocabulary size, however, suggests that the number of known words increases at least up to the age of 70 [61, 62] so that the assumption made, although certainly simplifying, does have some empirical support. The mathematical model is evidently simplistic. First, it assumes constant homogeneous mixing, and hence the three ways of estimating the basic reproductive ratio rest on this assumption as well. Problems that come along with this assumption are reviewed in [35]. It was shown that the invasion boundary in models that assume a more realistic population structure (scale-free networks) is lowered, so that R0 estimates could be effectively underestimated [7, 63]. This is because in real-world social networks there are individuals that have a disproportionally large number of contacts [64, 65], an observation that was suggested to be responsible for why cultural evolution progresses faster than biological evolution [66]. Mean-field models provide a good way of integrating a more realistic network structure [7, 63]. Meyers [67] shows that in an epidemiological contact network, R0=T(〈k2〉−〈k〉〈k〉), (9) where T is transmissibility (i.e., mean probability of transmission given a contact event), and 〈k〉 and 〈k2〉 are mean degree and mean squared degree of the contact network. Since degree variance is given by 〈k2〉−〈k〉2, more heterogeneous contact networks are expected to display larger R0 values. Network models could be informed by social-media data [68] or by extrapolating from contact surveys conducted for epidemiological purposes [69]. It would be interesting to see if a more realistic population structure would also lead to stronger correlations among the three R0 estimates. The model does not feature age structure. That is, adoption is assumed to be independent from age, which does not reflect how cognitive-semantic plasticity changes over time [70, 71]; clearly, the lexicon grows as age increases which in turn affects learning of new words. Likewise lexical processing and retrieval changes with age [72, 73]. Implementing several age-classes with differential adoption, abandoning, as well as mortality rates could mitigate this issue. Related to that, the procedure of estimating the basic reproductive ratio from age of acquisition assumes a rectangular age distribution. While this approximately applies to the age distributions of the present speaker populations considered in this study, this might not be the case for other speaker populations. On that note, the fact that English data mostly yield expected outcomes in this study (in contrast to the other three languages) might not necessarily be a reflex of larger sample size. It could also result from the fact that the age-distribution of the English-speaking population comes closer to a rectangular (i.e., equal) distribution than that of the other three languages (Fig 3). Likewise, life expectancy for the English-speaking population shows the smallest increase of all languages covered in this study. Both is in favor of the model assumptions. Hence, it could be that the higher quality of R0 estimates (as is evident from the smaller differences between estimation methods and higher correlations among the respective estimated values) is grounded in demographic properties of the speaker population as well. The model does not feature social structure, nor does it reflect geo-political patterns or other population demographics. Evidently, this can be problematic given that the investigated languages are spoken in several countries. However, compartment models are flexible enough to integrate information like this if contact rates among social and geo-political compartments are known. Likewise, the simplistic assumption of constant population size can be relaxed (note that work by May and Anderson [32] implies that similar approximations of the basic reproductive ratio based on age of acquisition can be used for increasing populations). Finally, the model assumes constant demographic and transmission rates. In particular, it excludes any type of stochasticity [16, 74]. Based on analytical work on stochastic SIS-models [75], we have shown earlier [76] that stochasticity in transmission decreases the basic reproductive ratio. This could be examined empirically based on diachronic corpus data. Finally, when it comes to language-internal dynamics, the model is simplistic in that it assumes the transmission of one word to be independent from the rest of the lexicon. That is, it ignores effects of the semantic neighborhood [77] and semantic competition [78]. Such interactions would require a version of the model in which multiple words can explicitly interact with each other. All the limitations discussed in this section point at avenues for improvement of the model and empirical analysis. Note, though, that similarly abstract models and concepts (in particular, the basic reproductive ratio) have been employed fruitfully in the field of epidemiology. This is by no means to be interpreted as an excuse for further improvement, but it tells us that we can learn something about contact-driven dynamics even from abstract models. 4.2 Implications Given the limitations discussed in the previous section, does the approach adopted here represent an accurate model of lexical spread? Certainly not–as it depends on simplistic assumptions. If the model assumptions are met, however, our analysis suggests that periphery words, on average, show a basic reproductive ratio of about R0≅6, with inter-quartile ranges going from about 2 to 17, depending on the language and estimation method. This puts spreading periphery words–in principle–roughly on par with infectious diseases like Rubella (‘three day measles’, [79]) Poliomyelitis [24, 32], or the COVID-19 Delta variant [80, 81] (‘in principle’ because the basic reproductive ratio is dimensionless; see discussion below). This figure defines a threshold that an average lexical innovation needs to pass to spread successfully [40]. That not every linguistic variant will spread is plausible, and the issue that new variants require a certain frequency of users in order to be of communicative value (which is rarely the case for new variants to begin with) was dubbed the “threshold problem” by Nettle [82]. Importantly, though, the epidemiological model adopted here does not feature any mechanism of positive frequency dependence modeling increased communicative value for higher user frequency. Rather, the model entails that a threshold for successful spread exists even if such frequency-dependent pressures are absent. Note that while the three estimation methods are expected to yield similar estimates, differences between languages might in fact reflect differential social, geographic, and demographic structures of the respective speaker populations. A larger set of languages is required, though, to investigate such effects. The estimate also implies that the herd-immunity threshold of an average lexical innovation equals about 83% (H=1−1/R0=1−1/6=0.83; minimal and maximal bounds of observed inter-quartile ranges yield estimates for this threshold that go from 50% to 94%). That is, 83% of all individuals would need to be prevented from adopting that word for it to vanish in the long run. Whichever hypothetical measures of language policy a ruling body would implement, they would need to be fierce to show an effect. Likewise, only a fraction of 17% resisting said measures would suffice to maintain that word, even in the lack of social or geographic structure that would enforce stable usage in a smaller subpopulation. The observation has interesting consequences for the study of language policy and censorship ([83]; see also [84] for a recent example). What the model entails is that, from this admittedly simplistic epidemiological point of view, top-down prohibitive measures of language policy are likely not to have good chances of success (and note that in this argument we do not even consider the preserving effect of written language). Integrating social structure into the model (and potentially differential language policies) would be crucial for a detailed analysis of the matter. Finally, the estimate R0≅6 also tells us about the role that individuals play in the propagation of new words through contact networks. Mean personal network size was shown to equal about 600 [85]. This reflects all individuals that one gets to know during one’s life time. This means that a single individual that picks up a lexical innovation passes it on to only about 1% of their personal contacts–notably, throughout the whole period during which that individual knows and uses that word. If one considers an average offspring count of about 1 (per individual), the impact that a single individual has on the spread of words decreases even further: literally only a handful of other people will on average pick up the lexical innovation due to interactions with that individual. That is, the basic reproductive ratio of words can shed light on how information is transmitted through speaker networks [86, 87]. A notable difference between linguistics and epidemiology in this context is that the prolific, i.e., infectious, period of words L is much longer than that of the above-mentioned infectious diseases (once acquired, words are kept for a much longer time span than, say, COVID-19 pathogens that get dealt with by the immune system over the course of a few weeks, on average). It follows from the approximation R0≅1+r∙L that words (with lower r but higher L) can plausibly display similar R0 values as the mentioned diseases (with higher r but lower L). Low growth rates do not entail that lexical change per se is a rare phenomenon (we look at each individual word separately, and there can by multiple lexical innovations that spread in parallel), nor that lexical change is inert (the basic reproductive ratio, as a dimensionless quantity, does not measure rate of growth). However, the finding informs us about the relative impact that each individual has on such change phenomena. Importantly, R0 is defined in such a way that it only considers propagation events in which one individual already knows and uses a certain constituent while another individual does not do so. It does not capture events in which both individuals already know a word and where the usage of the word is promoted in one individual as a reaction to the other one using it, say, through psychological priming effects or socio-pragmatic factors like conformism and prestige [82, 88, 89]. See [90] for a socio-pragmatically more elaborated version of the model including conformity and non-conformity biases. On a more general level, it is evident that the model, if taken for granted, together with the three ways for deriving the basic reproductive ratio provides links between three different domains: (i) lexical acquisition, (ii) synchronic lexical distribution, and (iii) lexical change. In particular, it provides a mechanistic link relevant to three lines of research. First, the link between language acquisition and change was subject to multiple studies in the field of evolutionary linguistics [91]. Generally speaking, words that are learned early are diachronically more stable in that they are less likely to change their form [28, 71, 92] or meaning [93]. One mechanism that was suggested to explain this link is that of cognitive entrenchment [94–96]. Words that are learned early display a higher degree of routinization (e.g., in their production, retrieval, or perception), remain stable in linguistic usage, and are therefore less likely to get ousted. In fact, the epidemiological model discussed in this contribution is agnostic with respect to such mechanisms: it predicts that early acquired words are diachronically more stable simply because early acquisition increases the length of the prolific (i.e., infectious) period in individuals, thus promoting successful propagation. In that sense, the model can also function as a–let us call it: population dynamic–base line against which hypotheses about the effects of cognitive entrenchment need to be tested. Second, and connected to this, the basic reproductive ratio links synchronic lexical distribution (prevalence) and diachronic lexical change (growth). In short, the prediction is that what is contemporarily prevalent has been successfully growing in the past. While this might sound trivial at first sight, the mathematical intricacies involved in that link provide more information. The idea is that a word’s prevalence at its population dynamic equilibrium (i.e., neither when it is growing nor declining) entails how fast it has been growing in the past, and vice versa. Quantitative research on diachronic lexical change largely draws on investigating frequency trajectories, or (trajectories of) dispersion measures based on token frequency and document frequency [97–99]. As discussed before, it is not trivial to derive lexical prevalence from corpus data [51], so that using a combination of frequency trajectories and contemporary prevalence estimates (as in this study) or conducting apparent-time diachronic analyses [100, 101] (Boberg 2004; Bailey 2004) remain the only feasible options. Both approaches come with their own problems. Most notably, they can only be reasonably applied to recent phenomena (words that have been prevalent a century ago but got ousted afterwards cannot be captured by contemporary surveys). Likewise, both approaches are challenged by speaker internal changes [102]. We conclude that more research is needed to reliably test the link between equilibrium prevalence and growth suggested by the model. Third, and finally, the basic reproductive ratio links acquisition with synchronic lexical distributions: words that are acquired early are used by a large share of the population [48], a link that was subject to research on language learning. It was shown that learning from multiple informants via distributed exposure promotes acquisition, both in infants [103, 104] and adults [105]. Again, the epidemiological model does not necessarily need to rely on mechanisms like this. Even if all individuals learn words at a constant rate, independent of their age, it trivially predicts that early acquired words are more widely spread and vice versa because (a) early acquisition increases the prolific period and hence the number of users of a word and (b) because words are more quickly adopted by a non-users if there are many users around. What is potentially more interesting than the correlation between ease of acquisition and prevalence is how exactly they are linked to each other through the basic reproductive ratio (more specifically: 1/(1−pU) = R0 = LE/AoA and hence pU = 1−AoA/LE): we have seen that straightforward transformations of age of acquisition and prevalence lead to estimates of the basic reproductive ratio that fall into similar ranges. Crucially, these transformations are theoretically motivated and follow from relatively basic epidemiological considerations (as opposed to assumptions about how age, learning, and cognition hang together). That is, the basic reproductive ratio functions as a theoretically grounded way of normalizing data from diverse domains to map them onto the same scale. 5. Conclusion In this contribution, we have explored ways of estimating the basic reproductive ratio in the linguistic context [25]. For this, we have revisited a relatively simple population dynamic model of the spread of single words, originally established in the field of mathematical epidemiology. Under the assumption that the model roughly captures how words spread through speaker populations, our study allows for the following conclusions. First, estimates derived from diachronic frequency trajectories only fit to those obtained from prevalence and age of acquisition when words do not change substantially in frequency. Even in that case, however, the correlation is far from strong. Second, and more optimistically, judging from the good fit of the estimates derived from prevalence and age of acquisition, it can be concluded that one does not necessarily need knowledge of the past development of words in order to determine the basic reproductive ratio. Our analysis shows that periphery words that have emerged relatively recently, on average, show a basic reproductive ratio of about R0 = 6. This not only lets us put numbers on how stable lexical innovations are (viz. the herd-immunity threshold of about 83%), but it also provides insight into the question of how often successful propagation of a single lexical innovation among individuals takes place at all. The answer is: only a handful of such events during the whole lifetime. Arguably, assertions like this cannot be derived by investigating token frequencies drawn from diachronic corpus data. Hence, we think that epidemiological models when they are applied to linguistics [3, 25, 26, 40], despite their high degree of abstraction, can provide interesting insights into how language evolves. 1 Aug 2024 PONE-D-24-25073Lexical innovations are rarely passed on during one’s lifetime: epidemiological perspectives on estimating the basic reproductive ratio of wordsPLOS ONE Dear Dr. Baumann, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please see comments from myself and two reviewers below. Please submit your revised manuscript by Sep 15 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Søren Wichmann, PhD Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. We note that you have indicated that there are restrictions to data sharing for this study. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For more information on unacceptable data access restrictions, please see http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Before we proceed with your manuscript, please address the following prompts: a) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., a Research Ethics Committee or Institutional Review Board, etc.). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of recommended repositories, please see https://journals.plos.org/plosone/s/recommended-repositories. You also have the option of uploading the data as Supporting Information files, but we would recommend depositing data directly to a data repository if possible. We will update your Data Availability statement on your behalf to reflect the information you provide. 3. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: http://journals.plos.org/plosone/s/supporting-information. Additional Editor Comments: The two reviewers raise many questions, which you should try to take into account to all possible extent. In a few cases a bit more is asked for than is possible to deliver in terms of data available so the best response in these cases would be to briefly indicate the limitations and perhaps (again briefly) speculating on how results might be different given more information. One of such "what if" questions of my own relates to how things would look if we know more about differences in the amount of contacts among individuals. The estimated R0 looks pretty high for an average among all individuals, but one would expect that there is a non-normal distribution of the amount of contact individuals have when contacts include readers/listeners of printed and audio media, so an estimated mean might not be very meaningful. Perhaps the distribution of contacts is something like a power law, with the vast majority of individuals belonging to a fat tail contributing even less to the spread of innovations than the R0 = 6 would suggest. In whatever way this is done (just mentioning, speculating, trying to find data...), I think it needs to be taken into account to at least some minimal extent that lexical diffusion is different from the spread of diseases in that a single individual can reach a large part of a population in one instance. There are a few typos etc.: l. 90: an average -> on average l. 91-92: perhaps move "not yet knowing the word" up to just after "contacts" l. 96-97: Heffernan et al. (2005) -> the paper of Heffernan et al. (23) l. 118: see van den Driessche (2017) -> see (24) l. 524: the lexicon growths -> the lexicon grows l. 530: data, mostly -> data mostly l. 615: delete "(Boberg 2004; Bailey 2004)" [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Partly Reviewer #2: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: No Reviewer #2: Yes ********** 3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes Reviewer #2: Yes ********** 5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: The paper proposes an interesting method transfer from epidemiology to linguistics, but suffers from simplification of language data and diachronic change factors to the point that I am not entirely sure it contributes much to linguistics in its current form, where the predictions are not tested yet rather strong claims about language change are made in the end. The unciritical usage of Google Ngrams for lexical data is also questionable. I would suggest the author to either: a) repurpose this as a method proposal paper, acknowledge the limitations, and remove the sweeping claims in the end; or b) figure out a way to actually test these predictions on a test set of lexical change cases, and work out how to estimate a confidence interval around the proposed R values and resulting claims. Don't get me wrong, it is an interesting idea, but the massive simplifications and ignorance of the nature of written corpus data will, I fear, only lead to backlash from linguists. Whereas this has potential to be actually useful, if presented carefully and tested appropriately. I will comment on a selection of passages below and attempt to explain my reasoning (the passages are copied from the pdf so line breaks and line numbers intersect with text, don't mind that). After that, we will test predictions against diverse diachronic and synchronic linguistic data from four languages (English, Spanish, German, Italian; sections 2-3). --- This is my main problem with the paper, I don't see how the model is really "tested". If there were, for instance, hypothetical political measures for banning a certain word, 232 one would need to ensure that at least individuals are not able to adopt that word. --- This does not take into account uneven language policies and societal attitudes, e.g. in earlier centuries the upper and lower classes of these countries differed in their language use, and even today usage of certain words can be restricted to certain groups who may internally or externally regulate (think of racial slurs for example). We test the predictions outlined above on the lexical level, i.e., by inspecting dynamics of words --- This is another problem I see, there is zero discussion on what counts as a word when it comes to frequency data - is it any space-separated token in a corpus, a lemma, or something else? There is no discussion on lemmatization, but the word lists used are presumably lemmas, so were only the forms corresponding to dictionary forms counted? E.g. for English nouns it's the singular, so were plurals (analogously other forms of verbs) not counted in frequencies? English is morphologically simple, but e.g. German is already more complex with each lemma having a number of forms. How does that affect results? words that have increased in frequency during the past two decades, 243 and we will refer to such words as lexical innovations. We do so irrespective of whether these words 244 have already been present initially at low frequency --- So which is it, innovations (not present before), or words that were low frequency and afterwards higher frequency? Moreover, we only consider periphery words, which we define, 247 somewhat simplistically, as words that are not used by everyone --- I am not sure what this even means. The corpora used here have no data on speakers. 250 estimates for a large set of English from Kuperman et al. (42). Prevalence estimates for English 258 were taken from Brysbaert et al --- This is my other gripe - how are age of acquisition estimates of modern-day children supposed to say anything about (presumably mostly adult) lexical innovations of literate language users in 1820? What is the connection here? Note that the accuracy of the prevalence estimates substantially depends on sample size, i.e., on 270 the number of participants they are based on. The respective sample sizes are about 300 for English, 271 about 23 for German, about 50 for Spanish, and 25 for Italian --- Am I to understand correctly that these sweeping claims about how lexical innovations work in human language are based not only on modern estimates but on a handful of e.g. 23 speaker optinions per language? For German, Spanish, and Italian, we needed 280 to resort to normalized frequency trajectories extracted from Google books --- Google Books Ngrams is a highly questionable source of language information - with some reservations it can be useful for e.g. grammatical things, but it has been shown to be questionable for lexical research purposes, as its sampling is pretty random and what may appear as "innovations" can easily be just sampling noise. Please see "Characterizing the Google Books Corpus: Strong Limits to Inferences of Socio-Cultural and Linguistic Evolution" by Pechenick et al and "The Impact of Lacking Metadata for the Measurement of Cultural and Linguistic Change Using the Google Ngram Data Sets—Reconstructing the Composition of the German Corpus in Times of WWII" by Koplenig, and consider if this is a valid corpus for your study. Also, think for a moment about the speaker communities. You are comparing US English where the speaker population went from around 8M to 300M during the COHA corpus period; Italian which is also spoken mostly in one country, German spoken in 4+ countries, and Spanish which by now has half a billion native speakers and is the official language in 20+ countries (in varieties which the Google Books ngram data does not differentiate between). How are these even remotely comparable in a model that relies so much on the individual speaker, and how can such a model reliably applied to such volatile and heterogenous populations? We use the collected age-of-acquisition estimates, assume a life expectancy of 75 years (https://data.worldbank.org/), and estimate the basic reproductive ratio as (i) 0 300 = /. Note 301 that life expectancy has increased through the past two centuries and that women have, on average, a 302 higher life expectation than men, so that the assumption of a single point estimate is clearly simplistic. --- Life expectancy in 1850 in the US was around 40 years, almost half your assumption. I don't see why this simplification is made, instead of working known expectancy estimates into your model. 303 In the UK, for instance, life expectancy has increased from about 71 in the early 1960s to about 80 in 304 the 2000s (Fig 3, left). Notably, the UK displays the smallest increase in life expectancy. --- The COHA is US American English. I don't see how the UK is relevant here. It is not at 328 all easy to obtain word prevalence based on corpus data. This is because corpora often feature only a 329 small number of authors and because the amount of text per author is likely not representative of their 330 linguistic knowledge (let alone the fact that the author population in linguistic corpora is not 331 necessarily representative of the whole population --- This is a massive problem on its own, acknowledged but not taken into account - but the related problem is differential number of speakers over time - more people simply became literate over 200 years in these countries, so if the corpus is randomly sampled, the number of speakers in it should be significantly different, and ignoring this surely introduces bias in a model so heavily reliant on the unit of individual. Just a note on 3D plots, and tables and figures with abbreviations - these are just very hard to read. Would be much easier if the terms were written out on the axes and in tables. 3D plots like these are almost unreadable; could just use color for the 3rd dimension where needed. 382 The analysis of expected values for the basic reproductive ratio (for all languages and methods) shows 383 that estimates typically range from about 3 to about 14 --- Here's my third main issue with the paper: what do we learn from this, and how do we know if these numbers are anywhere near true values? There is some post-hoc discussion later, but what use is this number for linguistics and advancing our understanding of how language works and changes? 545 Given the limitations discussed in the previous section, does the approach adopted here represent an 546 accurate model of lexical spread? Certainly not – as it depends on simplistic assumptions. If the model 547 assumptions are met, however, our analysis suggests that periphery words, on average, show a basic 548 reproductive ratio of about 0 ≅ 6, with inter-quartile ranges going from about 2 to 17, depending on 549 the language and estimation method. This puts spreading periphery words – in certain settings – 550 roughly on par with infectious diseases like Rubella --- Again, rhetorically, so what? The number that may or may not be correct is like a number if a completely unrelated domain. What is the confidence/margin of error here, given all the (considerable) limitations here pertaining to corpus data, population data and the model, are we talking 6±1 or 6±6? 571 Finally, the estimate 0 ≅ 6 also tells us about the role that individuals play in the propagation of new 572 words through contact networks. Mean personal network size was shown to equal about 600 (80). This 573 reflects all individuals that one gets to know during one’s life time. This means that a single individual 574 that picks up a lexical innovation passes it on to only about 1% of their personal contacts – notably, --- Exactly the same critique as above. 575 throughout the whole period during which that individual knows and uses that word. If one considers 576 an average offspring count of about 1 (per individual), --- This is below rate of replacement, and well below the average number of children historically. Making this assumption equates to assuming these four languages or their speaker populations have gone extinct by now. As a final comment, if such a model would be tested against known case, I expect there would need to be some way to account for an error term or variance left undescribed. These current estimates assume a singular process of lexical innovation, whereas language change is complex and affected by numerous factors mostly or entirely ignored here. To provide some ideas: - Age and mixing is briefly discussed in the end, but nothing about sociolinguistic prestige (tons of literature on that) - Selection and drift literature is cited but it seems like drift could be one way to conceptualize the residual/error not described by this model (for a more recent update since the Newberry paper cited here, see e.g. "Reliable identification of selection mechanisms in language change"). - Borrowings are mentioned, but how the proportion of L2 speakers affects language is ignored (see e.g. "Larger communities create more systematic languages", "Language Structure Is Partly Determined by Social Structure") - Stadler et al is cited but this potential mechanism is ignored; relatedly, other works have looked into neighborhood/topical effects ("Quantifying the dynamics of topical fluctuations in language" "Where New Words Are Born: Distributional Semantic Analysis of Neologisms and Their Semantic Neighborhoods") - The issue of semantic change is practically ignored - it is not just new forms being innovated, but new meanings arise and old words are given new meanings (maybe useful: "Algorithms in the historical emergence of word senses", "Survey of Computational Approaches to Lexical Semantic Change") I am not saying a model would need to necessarily include parameters for all these things, but some acknowledgement that it is not just neutral disease-like passing of new words would be prudent, and if the model were to be evaluated on some test set, it would make sense to reserve an error term or otherwise account for the fact that there are other things happening in language change. Reviewer #2: In this work, the author views lexical dynamics as an infection process, whereby a particular lexical item can infect a user and thus spread through a population. In particular they focus on computing the basic reproduction number R0 for four different languages through three different methods. Subject to underlying assumptions on the dynamics being correct, all three methods should give similar results and the author finds that they do indeed correlate well, with a median value (over words in language) of around 6. The author performs a careful statistical study of the results, showing that the correlations in particular, and the findings in general, are robust. The discussion section does a good job of examining the implications of the results for language dynamics, and in addressing the limitations of the study. Overall I have little to criticise, although I think the manuscript could benefit from a little more discussion of the following points: 1. The comment that "most average R0 estimates fall into a similar range, across languages ... is reassuring ... and potentially more revealing" deserves expansion. The easiest R0 to intuit is the the one obtained by dividing life expectancy by age-of-acquisition. A value of 6 is therefore indicative of the fact that the median word is typically acquired at around 1/6th of one's overall lifetime, which seems broadly reasonable even given historical changes in life expectancy. The question then is: how much scope is there for variation in values of R0 given how they are defined? If the answer is 'not much' then we should perhaps not be too surprised that we get similar ballpark figures back. (I might argue that the correlations between values from the different methods are therefore more revealing, because there it's less obvious that the variables that they depend on should be related.) 2. Following from point 1, it would help to flag up the basic assumptions behind the model defined by the equation on p8. (NB: it is very helpful to number equations, as it makes it easier for readers to refer to them.) Key among these are the chosen number of compartments, and the fact that populations are assumed to be well-mixed. These assumptions are discussed towards the end of the manuscript, but I think it is worthwhile to emphasise that the different ways one can define R0 are a consequence of these assumptions. Thus if one does not get similar values from the different methods, one possible explanation is that the underlying assumptions are not valid. 3. Related to point 2, it would help to clarify what the infection event that causes a transition from the N to U compartments is envisaged to be and what the characteristic timescale of infection is. Within the 'age of acquisition' approach, it seems implicitly the case that N to U event is a childhood language acquisition event, which one could potentially question for lexical dynamics. For example, when there is an innovation, a word could also spread through an adult population without having to wait to be learnt by the next generation of children. I think this impacts on the discussion around the relatively small number of people that we infect with our lexicon: if we have only childhood language acquisition in mind, then 6 is actually quite a big number, since most individuals tend to act as a caregiver to a small number of children. The timescale question meanwhile I think is important when making comparisons to diseases like Rubella and Covid-19. R0 doesn't say anything about timescales - just how many people you pass the infection onto, on average - but it is clear from these comparisons that the despite having different R0, the timescale over which a lexical item infects a population is much longer than that of an infectious disease. It is not immediately obvious to me what sets the timescale in the linguistic case, especially if infection can happen through use as well as acquisition. It would be interesting if the authors could add any insights they have on this issue to the manuscript. ********** 6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: Yes: Richard Blythe ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 26 Sep 2024 All responses can be found in the 'Response to reviewers' document attached at the bottom. Submitted filename: Response to reviewers.pdf 7 Oct 2024 Lexical innovations are rarely passed on during one’s lifetime: epidemiological perspectives on estimating the basic reproductive ratio of words PONE-D-24-25073R1 Dear Dr. Baumann, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. If you have any questions relating to publication charges, please contact our Author Billing department directly at authorbilling@plos.org. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Søren Wichmann, PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 18 Oct 2024 PONE-D-24-25073R1 PLOS ONE Dear Dr. Baumann, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset If revisions are needed, the production department will contact you directly to resolve them. If no revisions are needed, you will receive an email when the publication date has been set. At this time, we do not offer pre-publication proofs to authors during production of the accepted work. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few weeks to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Søren Wichmann Academic Editor PLOS ONE