We analyzed the same publicly available data set described in Gahl (2008, 2009), Gahl and Baayen (2024), and Lohmann (2018a). With the exception of Gahl (2009), these previous analyses of the data set all used type-level models, i.e. modeling word duration averaged over word types. We refrained from collecting or proposing any novel variables or making other changes to the dataset, for the sake of comparability to earlier studies and so as to maintain the focus on methodological issues.

According to Gahl (2008), the initial list of target words for the data set contained all English word forms with at least one non-homographic homophone, based on the CELEX database (Baayen et al., 1995). Word forms with identical spelling and pronunciation (e.g. nouns and verbs spelled time) were treated as tokens of the same type. Gahl (2008)’s exclusion criteria excluded the following items: (1) spellings associated with more than one pronunciation, e.g. tear (homophonous with tier and tare); (2) pairs involving function words, such as in/inn and or/ore, and interjections, such as whoa/woe; (3) pairs, such as source/sauce, that are homophones in the (Received Pronunciation of British English) CELEX transcriptions, but unlikely to be homophones in the varieties of American English represented in the Switchboard corpus; (4) items containing transcription errors in CELEX; and (5) names of letters of the alphabet. The resulting list contains 409 target types, represented by 79,219 tokens.

The spoken duration of all tokens of these items was extracted from the time-aligned orthographic transcript (Deshmukh et al., 1998) of the Switchboard corpus (Godfrey et al., 1992), a corpus of 240 hours of telephone conversations between strangers. We initially retained tokens that were immediately followed by an unfilled pause (defined as a period of silence of 0.5 seconds or longer) or by a filled pause (e.g. a hesitation marker such as um, uh, err) and allowed Fluency, coded as a binary factor, to interact with talker age and with the relative frequency of target and homophone. However, these disfluent tokens were too unevenly distributed to allow meaningful models of these interactions. Therefore, we excluded them from all further analyses. Excluding the disfluent tokens left 56,024 tokens of 403 distinct types for analysis.

We use a published, type-based model of the data set as a baseline. The model summary, replicating Gahl and Baayen (2024), appears in Table 1, and the model’s smooth terms are visualized in Figure 1. In the type-based model, predicted (average) duration increased linearly with residualized baseline duration, and in a nearly linear fashion with the proportion of prepausal tokens. The effects of noun bias and Orthographic regularity were non-significant. Increasing phonological neighborhood density (PND) was associated with decreasing duration in a nonlinear fashion, with the steepest predicted decrease in the lower range of PND. As for the interaction of lemma frequency and frequency ratio, the contour plot in Figure 1 shows the general gradient in the regression surface to be negative. That pattern reflects the oft-observed relationship between increasing lemma frequency and decreasing predicted duration. Increasing lemma frequency was associated with decreasing variance in duration (bottom right panel of Figure 1).

Further scrutiny of the model revealed that the deviance residuals of the model were approximately Gaussian and did not co-vary with lemma frequency or with the frequency ratio (all |z| < 1, all p values > .7), suggesting that, for all other predictors, the assumptions of normality and constant variance were met to a satisfactory degree. The two variables that did not give rise to significant effects were either not well established as predictors of spoken word duration (orthographic regularity) or estimated in a manner that was likely to be highly imprecise (noun bias). In sum, the type-based model recovered a number of well established effects and was silent on variables that independently appear questionable as predictors.

The type-based model could undoubtedly be improved, but we refrain from doing so here: In the context of the present study, the model simply serves as a point of comparison for type-based vs. token-based models.

We begin our discussion of token-based models with a model that is similar to the type-based model in the choice of lexical predictors, but adds token-level information. Similar to the method usually adopted in token-based analyses using Linear Mixed Effects Regression, this first token-based model includes a by-word-form random intercept. The remaining variables are identical to those in the type-based model, with the following exceptions: First, we did not include the Pause quotient, i.e. the proportion of tokens of each type that were followed by a pause; secondly, we included three variables coding token-specific information: the talker’s sex and age, and the target’s contextual probability, given the word immediately following the target word in the utterance. The summary of the model is shown in Table 2, and visualizations of the regression smooths appear in Figure 2.

All of the predictors specific to token-based models reached significance in the expected direction, given prior corpus-based models not restricted to homophones (e.g. Bell et al., 2009; Horton et al., 2010; Yuan et al., 2006), as follows: Increasing word bigram probability was associated with shorter duration. Tokens produced by male talkers tended to be shorter than those produced by female talkers, other things being equal. Increasing talker age was associated with increasing token duration up to about age 30 and above age 55; there was no clearly discernible effect of talker age in the middle range of talker age.

With respect to the predictors that also appeared in the type-based models, the token-based model behaved similarly to the type-based model in some respects: The overall direction of the effect of PND was negative, and the effect was steepest in the lower ranges of that variable. Predicted duration increased with residualized baseline duration, as expected, although the shape of this effect was nonlinear, unlike in the type-based model. The effect of orthographic regularity was associated with a p-value exceeding our preset alpha level, as in the type-based model. Unlike in the type-based model, the effect of noun bias reached significance, such that target forms more likely representing nouns than some other part of speech had longer predicted duration. That pattern is consistent with the idea that nouns are more often phrase-final in English than other parts of speech, i.e. the rationale for including this variable in the original (type-based) analysis of the homophone data (Gahl, 2008).

The token-based model also recovered the overall association of increasing frequency with shorter predicted duration, but its predictions about the interaction of target frequency with frequency ratio were more complex than in the type-based model. In the type-based model, increasing lemma frequency was associated with decreasing predicted duration across almost the entire range of the frequency ratio. In the token-based model, the effect of lemma frequency was nearly absent for words with relative frequencies of about 2, reasserting itself for words in the highest range of the frequency ratio variable. There was also some evidence for an effect of frequency ratio, such that low-frequency targets had shorter predicted durations if they had very low frequency ratios, i.e. high frequency homophone twins. To the extent that this pattern turns out to be robust, it may reflect an effect known as frequency inheritance: Low-frequency homophones of high-frequency words have sometimes been found to behave in some respects like high-frequency words, as if “inheriting” the frequency of their high-frequency twins. Such inheritance effects, which have been found in data based on speech errors (Dell, 1990) and latencies in semantic decisions and translation (Jescheniak & Levelt, 1994; Jescheniak et al., 2003), may coexist with lemma-specific effects of frequency on duration (cf. Gahl, 2008, for discussion).

A striking prediction of the token-based model concerns the shape of the predicted variance in duration as a function of target frequency. Recall that in the type-based model, the predicted variance decreased as target frequency increased. As seen in Figure 2, the prediction seemingly goes in the opposite direction for the token-based model. The predicted variance in duration for a given target frequency increases with target frequency. That pattern, of higher variability of high-frequency targets, is one that we take up in Section 4. The confidence region around the predicted variance is wide for low values of frequency and narrows with increasing frequency, as one would expect, given that uncertainty about an estimate (here, the estimated variance) decreases as sample size increases.

The distribution of the by-word-form random effect suggests departures from normality near the extremes of the by-word-form values. Model diagnostics further revealed that the number of basis functions was overly constrained in the case of the smooth terms for bigram probability, suggesting that the shape of the smooths, as depicted in Figure 2, is underfitting the data; further exploration revealed that this issue persisted when the number of basis functions was increased to as high as 19. We return to this issue in 3.8.1 below.

Setting aside the overly constrained estimate for the effect of bigram probability, the token-based model recovers all of the expected effects of lexical factors on word duration, and it additionally reflects plausible effects of contextual probability, talker sex, and age. The by-word-form random effect did not suggest systematic departures from normality. In sum, the token-based model looks successful at first glance.

The data set for the token-based model may well support additional predictors – and omitting such predictors may distort some of the predictions of the token-based model. Anticipating that objection, we therefore fitted a token-based model including two predictors that did not reach significance in the type-based model, viz. morphological complexity and biphone probability. Model summary and partial effects plots for that model are in Appendix B. Neither morphological complexity nor biphone probability reached significance based on our preset alpha level. The effect of orthographic regularity continued to be non-significant. The shape of the relationship of the remaining terms was similar to the model without the additional terms, and the predicted variance as a function of target frequency once again increased with increasing frequency. We take these results to alleviate concerns one might have about the omission of these variables distorting the initial token-based model (Figure 2).

We have seen two token-based models that appear to be successful at capturing both type-level and token-level information. However, a problem with these models becomes apparent when we consider model concurvity. As noted above, concurvity is a measure of the degree to which each component of the model can be approximated by the other components. The consequences of high concurvity are analogous to those of high collinearity in linear models: Both interfere with model interpretability, by making it impossible to identify each predictor’s relationship to the outcome, and by rendering individual estimates uninterpretable, i.e. liable to change substantially in response even to minor changes in modeling procedure.

Concurvity values for the type-based model (shown in Table 3) indicate high concurvity for the parametric part of the model, which suggests that combinations of the factorial predictors (i.e. talker sex and/or noun bias) are not independent of the smooths, as well as for the predicted variance as a function of frequency. The remaining lexical variables in the type-based model do not give rise to high concurvity.

Matters are different for the two token-based models discussed so far: As shown in Table 4, most of the terms representing type-level variables in the token-level models are highly predictable from other smooth terms in the models. The concurvity levels for PND, orthographic regularity, baseline duration, target frequency, and (in the model containing this term) biphone probability are all near .89 or even higher, rendering these terms uninterpretable. This is in stark contrast to variables whose values can vary within lexical types, i.e. local context (word-based bigram probability, for which concurvity values range from .28 to .46) and talker age (concurvity between .01 and 02, suggesting that talker age is independent of lexical variables in our sample).

Table 4 further suggests a pattern in the interplay of the by-word-form random effect with other type-level variables: When the random effect is included, PND, orthographic regularity, baseline duration, and target frequency all give rise to high concurvity. When it is not, concurvity levels for these variables go down considerably. The interplay of the by-word-form random effect with target frequency is also worth noting. Frequency is highly predictable (concurvity > .89 or higher) in all models containing the by-word-form random effect, and much less so (concurvity < .5) in models without that term. Conversely, the random effect is least predictable (<.4 vs. >.5) from other terms when frequency is not a predictor in the model. The by-word-form random intercept, where present, gives rise to concurvity values ranging from .37 to .57, i.e. noticeably higher than those for bigram probability and talker age, but far lower than those for variables pertaining to phonological form (PND and residualized baseline duration). The intermediate degree of concurvity of the random intercept is plausible, in light of the fact that each word form (e.g. /taіm/) in the data set receives a single posterior prediction shared by two lexical items (e.g. time and thyme) with different characteristics.

To understand whether the high model concurvity might be pinned down to specific pairs of predictors, we also inspected pairwise concurvity values. The pairwise concurvity estimates for the model in Table 2 are shown in Table 5. These estimates suggest that the by-word-form random effect is by far the worst offender, giving rise to concurvity values at or near 1 nearly across the board; the only variables that are not fully predictable from the random effect are the two token-specific variables (bigram probability of the target word, given the next word, and talker age). The redundancy of the by-word-form random effect, and its apparent ability to render all other lexical variables in the model redundant, suggest that including that term may be a mistake. A logical response to that problem might to remove that term from the model. That is what we do next.

A token-based model without the by-word-form random effect, but otherwise identical to our initial token-based model 2, is summarized in Table C.1 in Appendix C. The smooth terms are visualized in Figure 3.

The model estimates resemble those of the models with the random effect in some ways: Tokens produced by male talkers tend to be shorter than those produced by female talkers. The parametric effect of noun bias was once again significant (unlike in the type-based model, but like in the token-based models with the random effect), in the same direction as before. The smooth term estimates were also similar to the model with the by-word-form random effects in several respects: Increasing word bigram probability and pnd were associated with shorter duration. Increasing talker age was associated with increasing token duration in the lower age range (up to about age 30) and, with greater uncertainty, above age 55. Another point of similarity to the model with the random effect is the increasing predicted variance as a function of target frequency: The tensor product of target lemma frequency and relative frequency yielded a prediction surface that was, overall, similar to that of the models with the random effect.

Removing the random effect also resulted in some differences, however: As one might expect, there was a noticeable narrowing of the confidence regions for the estimates of the lexical variables, i.e. pnd, orthographic regularity (which reached significance in this model), and residual baseline duration. Increasing residual baseline duration, which had shown a non-linear effect (flattening in the upper ranges of the predictor) in the token-based models with the random effect, was associated with a nearly linear effect in the model without the random effect, resembling that variable’s behavior in the type-based model. Removing the random effect resulted in a higher AIC (38,888.81, compared to 35,075.65 in the corresponding model with the random effect). By that criterion, the model with the random effect is preferable. As an aside, we note that the substantial change in the AIC suggests that the lexical variables, although considered well-established as predictors of lexical processing, are far from perfect as predictors of word duration in conversational speech. We do not consider this outcome to be cause for concern. The notion that the lexical variables are significant predictors at all is hardly threatened by the observation that the predictors are imperfect.

The rationale for removing the random effect was the high concurvity seemingly associated with that predictor. Did removing the random effect solve that problem? Table 4 suggests that the answer may be yes: The lexical predictors (PND, orthographic regularity, baseline duration, and frequency) were associated with much lower concurvity in the model without the random effect.

The only model component that continued to give rise to high concurvity after removing the random effect was the term predicting the variance as a function of target frequency. We therefore remove the variance estimate next, to give a random-effects model the chance to incur minimal redundancy. It is conceivable, after all, that the redundancy of the random effect was exacerbated by estimating the variance in duration as a function of frequency along with the mean. Indeed, the concurvity associated with the term predicting the variance may result from the fact that frequency appears to be predictive of duration mean and variance. This might explain why mean and variance vary in similar ways. Therefore, although the concurvity of the variance may not be the main cause for concern, a random effects model assuming equal variance might conceivably reduce concurvity to acceptable levels.

To understand the interplay of the by-word-form random effect with estimating the variance in duration as a function of frequency, we compare models assuming equal variance with and without the random effects term. Full model summaries and partial effects plots for all predictors in the models can be found in Tables D.1 and D.2 and Figures D.1 and D.2 in Appendix D. Here, we focus on the patterns of concurvity, and on some salient differences in the smooth terms.

Table 4 shows the effect on model concurvity of including the random effect in the two models assuming equal variance: The model including the random effect incurs high concurvity for all lexical variables; as before, the only variables that are unaffected by the problem are those capturing token-level properties, i.e. word-based bigram probability of the target, and talker age. The model without the random effect shows much more acceptable levels of concurvity. In fact, in that model, even the parametric terms are not fully predictable from the rest of the model, unlike in all other models considered so far. On the minus side, not modeling the variance resulted in a higher AIC (36,467.36, up from 35,075 for the otherwise identical model in Table 2), suggesting that modeling the variance is well justified from the point of view of balancing parsimony and goodness-of-fit.

How does dropping the assumption of equal variance affect the predicted relationship between the various predictors and duration? Figure 4 shows the model predictions for three variables for the two models without the variance term. For the sake of comparison, the top row of the figure repeats the corresponding partial effects plots of the initial token-based model, i.e. the model including the by-word-form random effect and modeling the variance (Figure 2).

As Figure 4 shows, the models differ in the degree of uncertainty around those estimates, as well as in the shapes of the estimated relationships, as follows: The confidence regions around the estimates for the smooth terms are noticeably wider for the two models that include by-word-form random effects, compared to the model without that term (the bottom row in the figure). This is to be expected, given that dropping the random effect leaves it to the remaining terms to capture more of the differences across target words.

As for the shape of the relationship between predictors and predicted token duration, the effect of residualized baseline duration is non-linear in the two models with the random effect, increasing in the lower ranges of residual baseline duration and then flattening or even changing direction (though with high uncertainty) in the upper ranges. The relationship is more closely linear in the model without the random effect and without modeling the variance – resembling the linear relationship between residual baseline duration and predicted duration in the type-based model (Figure 1). The predicted effect of PND likewise changes shape: That effect is nearly linear in the model with the random effect assuming equal variance (the middle row of Figure 4), unlike in either of the other token-based models, and unlike in the type-based model. Finally, the shape of the prediction surface for the tensor product is similarly enigmatic for the models including the random effect, regardless of the inclusion of the variance term. In the model without the random effect and assuming equal variance, the prediction surface shows twin summits in the upper left quadrant (i.e. the highest values of frequency ratio in the lower ranges of target frequency) towering over the expanse of the remaining surface, which is almost entirely flat.

In sum, the high concurvity we observed for the models with the by-word-form random effect cannot be blamed on modeling the variance as a function of target frequency, as concurvity is high for both models that include the random effect. Moreover, assuming equal variance hurts the model’s AIC, i.e. its goodness-of-fit balanced against parsimony.

There is another possible culprit for the problems associated with the random effect, and that is the decision to include target frequency as a predictor (in the tensor product of target frequency and frequency ratio). In including frequency as a predictor, we followed common practice in corpus-based research, including some of our own. But including frequency as a predictor in token-based models might well be problematic, given that the token count itself reflects lexical frequency (albeit frequency in the corpus under analysis, rather than the CELEX frequency estimates used here). To give the random effects model another chance to account for grouping structure in the data without compromising the independence of the lexical predictors, we remove frequency from the set of predictors in the model.

The model summary and partial effects plots can be found in Table D.1 in Appendix D. The pattern of significant effects is essentially unchanged in the model without frequency vs. the earlier models. The visualization of the smooths in Figure D.1 suggests that the model without frequency most closely resembles the initial token-based model, i.e. the model including the by-word-form random effect. Both PND and residual baseline duration give rise to non-linear effects (generally decreasing with increasing PND, and generally increasing with increasing residual baseline duration). As before, the distribution of the by-word-form random effect suggests some departures near the extremes.

Does removing frequency as a predictor solve the problem it was intended to solve? Table 4 suggests that the answer is no: As in each previous model that included a by-word-form random effect, all lexical predictors (PND, orthographic regularity, and residual baseline duration) are nearly perfectly predictable from the model as a whole; the only non-redundant predictors are the token-level properties, i.e. target bigram probability and talker age. What this suggests is that, while including frequency as a predictor may be problematic, removing it does not solve the problem of high concurvity.

So far, we have said very little about whether the various models we considered balanced accuracy and parsimony acceptably well or satisfied the modeling assumptions. These are the questions to which we now turn.

We begin by summarizing how our models fared vis-à-vis the four potential problems we identified in Section 1.

We have been utilizing a data set that has seen multiple previous analyses, in order to keep the focus on the methodological points we are making, rather than the properties of particular estimates of (type-level and token-level) variables. It might be asked whether the problems we are pointing out are specific to English homophones – or, for that matter, to the analysis of spoken word durations or other acoustic properties. We believe that this is not the case. Our points apply in principle to any analysis of lexical variables in unbalanced data sets, especially when the variables involved are correlated with type frequency. Such might be the case for analyses of reading times in running text, or lexical effects in typing speed in naturalistic samples. One example where similar issues arise is Chuang et al. (2026), who use GAMs to predict f0 contours, in an analysis of Mandarin tones. In that study, the research question and the nature of the data necessitate a token-level analysis: The research question concerns the existence of word-specific pitch contours. In order to address that question, it is, first of all, necessary to control for non-lexical local determinants of pitch, such as neighboring tones, position within an utterance, and syllable duration. Averages at the type level would make it impossible to control for these token-by-token properties. Secondly, Chuang et al. (2026) link token-specific pitch contours to their corresponding contextualized embeddings, i.e. another token-specific set of (semantic and form-based) properties of tokens in the context of utterances. In order to keep high-frequency words from dominating the resulting models, while ensuring that estimates of lexical (i.e. type-level) pitch signatures are based on satisfying sample sizes, Chuang et al. (2026) set an upper limit to the number of tokens per word included in the analysis, as well as a lower limit. While this solution is imperfect, for reasons we are about to discuss, it does allow lexically-specific effects to assert themselves. Importantly, the empirical domain Chuang et al. (2026) concerns neither English, nor homophones, nor word duration, illustrating that the problems to which we are drawing attention extend beyond the particular dataset under discussion in the current article.

The problems we described are bound to arise whenever per-item samples are very unequal, even in large samples. When items have very different frequencies, the random intercepts may diverge considerably from normality (Douglas Bates, p.c.). Furthermore, when item-bound (here, by-word-form) predictors are of interest, the random intercepts are confounded with the item predictors. When predictors have values that are instantiated across multiple tokens of the same word form, whether a predictor will reach significance in the presence of a by-item random effect will depend on which values of the predictor happen to recur across these tokens. For predictors with non-repeating values, on the other hand, there is a one-to-one relation between predictor values and random intercepts. As a consequence, concurvity values will be near 1 when smoothing splines are used.

Sampling a set number of tokens for any given word type (see, for an example, Pluymaekers et al., 2005), might seem to be one strategy for addressing some or all of the problems we noted: In this way, one might keep the data from being overwhelmed by high-frequency words and remove the confound of item frequency and by-item sample size. One might take this strategy one step further and sample repeatedly, so as to avoid relying on a single sampling operation. However, repeating the sampling procedure introduces a new problem: Any given token of a high-frequency word will have a smaller chance of being drawn, and tokens of infrequent words will be included with near certainty (or complete certainty, in the case of hapax legomena). The results, over many runs, still reflect type-level properties very unevenly.

In addition, creating equal sample sizes does not remove inherent properties of high-frequency words and the confounds with other lexical properties that they give rise to. We argue that this problem cannot be overcome by using larger or more evenly sampled data sets. These are strong claims, and we suspect that analyses of additional data sets will be needed to explore these claims fully. But we see reason to believe that the properties of high-frequency words systematically affect statistical models of lexical properties, in ways that have not been fully appreciated and that may necessitate complementing token-based analyses with type-based ones. One reason for this is that lexical frequency is correlated with multiple other lexical variables (see e.g. Frauenfelder et al., 1993). Therefore, if model residuals are correlated with lexical frequency, then they will also be correlated with other lexical variables.

It might be asked why random effects, whose job it is to account for grouping structure in the data, are themselves correlated with item frequency. We think that one reason for this is that high-frequency word forms tend to be massively ambiguous (or vague), i.e. to be associated with many different meanings (and/or senses: the distinction between ambiguity and vagueness unfortunately does not help here). High-frequency word forms effectively represent many different words, as noted, for example, in Baayen et al. (2006). Along similar lines, Pimentel et al. (2020), cited in Hermalin (2025), found that more contextually predictable words tended to have more meanings/senses. As a consequence, form-level estimates for high-frequency words represent attempts to characterize wildly heterogeneous groups, and increasingly so with increasing frequency. If one believes, as we do, that word meanings and word senses matter for phonetic realization, this heterogeneity will make itself felt in the model predictions: The heterogeneity of the meanings of high-frequency forms may be partly to blame for the relationship between frequency and by-item adjustments. In this connection, it is interesting to note that the predicted variance went up as frequency increased (cf. Table 2). This might seem unexpected under the usual assumption that uncertainty decreases with increasing sample size. But it is expected, given that high frequency forms are associated with many different meanings.

If this reasoning is correct, then drawing equal numbers of tokens (say, n = 5 per word form) again will not fix the issue: The five tokens of hypotenuse will still mean similar things; whereas the five tokens of corner will probably mean different things, having been drawn from such different contexts as in our corner of the world, they backed him into a corner, the store is right around the corner, the referee awarded a corner, and so on. The heterogeneity of high-frequency forms will still be present, even if one creates equally sized samples. Careful sense disambiguation based on context-specific information about meaning may mitigate this issue. It is possible to use sense-disambiguation algorithms from Natural Language Processing to assign a discrete sense to the individual word tokens. Such a procedure does not, however, address the question of “where one sense of a word ends and the next begins” (Kilgarriff, 2006, p.29). Whether this approach solves the technical issues due to semantic ambiguity and vagueness remains an open empirical question. Alternatively, contextualized embeddings can be calculated for every single token, using Large Language Models. Such contextualized embeddings estimate words’ meanings in context, but the full richness of an individual speaker’s communicative intentions for a given word token can only be approximated (for a comparison of these methods in a corpus-based study of Mandarin tone, see Chuang et al., 2026).

As stated at the outset, we are emphatically not claiming that token-level models are to be avoided under all circumstances. Setting aside the issue of model assumptions for a moment, what might be considered ballot-box stuffing in one context may amount to desirable weighting of evidence in the next. We believe caution is in order, particularly when the interest of an analysis concerns types, rather than tokens. Token-level models estimate token-level behavior, which can make for a mismatch between the statistical unit of analysis and the inferential target. This invites the question of how to choose between type-level and token-level models, or perhaps how to combine the two.

Fundamentally, the choice of analytical target depends on one’s theory of the processes that generated the data. Testing predictions about elements of a (type-based) lexicon vs. acoustic, perceptual, or situational constellations not tied to word types calls for different solutions (see e.g. Gahl & Baayen, 2024, for discussion). Even when the research question ultimately concerns type-level properties, token-level analyses may be a necessary element of the analysis, precisely because the process that generated the data is hypothesized to involve both type-level and token-level forces. For example, as mentioned above, the research question in Chuang et al. (2026) concerns the existence of word-specific f0 contours in Mandarin tones, a type-level property of words. In that study, the research question and the nature of the data necessitate a token-level analysis: Averages at the type level make it impossible to control for these token-by-token properties. Superficially, the token-level analysis is necessary to control for utterance-level local determinants of pitch. More fundamentally, the authors’ theory of the process that generated the data involves both type-level forces (the hypothesized lexically-specific tone contours) and utterance-level forces (such as local speaking rate and shades of meaning in specific contexts of use).

Again, the choice between analyses based on tokens and analyses based on types requires reflecting on the goal of the analysis. For example, if the goal is to predict word duration as accurately as possible with as few predictors as possible, then token-level models taking into account pre-pausal lengthening and local speech rate may be perfectly satisfactory. But these predictions will be blind to many processes that may, nevertheless, be operative in human language production. Token-level models can also help address certain questions about learning and development, where one may want to embrace the idea that high-frequency words are represented in the data proportionally to their frequency. During learning, word types are encountered proportional to their frequency in the learner’s experience; as a consequence, the cognitive system receives its fine-tuning predominantly from the higher-frequency words.

We conclude with a set of recommendations. Again, which of these are optimal depends on the specific (empirical and theoretical) goals of any given research project.

It is beyond the scope of the present study to evaluate these recommendations. Rather, our goal is to point out that the ability to model large sets of observations has led to a proliferation of models that may actually run counter to some of the goals of psycholinguistic research.

Type-based analyses need not forego information about contextual variables altogether. In fact, when used in combination with token-level information, they can shed light on the consequences of the accumulation of utterance-specific properties for lexical processing. For example, Seyfarth (2014) found that “Words that usually appear in predictable contexts are reduced in all contexts, even those in which they are unpredictable” (Seyfarth, 2014, p.140). Similar cumulative effects, e.g. of the positions a word most often occurs in, on its production, even in other contexts, have also been reported in Brown et al. (2021) and Sóskuthy and Hay (2017).

Ultimately, the decision whether to treat any variable, including frequency, as a lexical property depends on hypotheses about the processes one is interested in. Whether form frequency should be considered a property of tokens of, say, thyme depends on whether one believes that the frequency of the form [taɪm] was relevant to the processing of tokens of thyme.