Multi-factor analysis in language production: Sequential sampling models mimic and extend regression results

ABSTRACT

For multi-factor analyses of response times, descriptive models (e.g., linear regression) arguably constitute the dominant approach in psycholinguistics. In contrast empirical cognitive models (e.g., sequential sampling models, SSMs) may fit fewer factors simultaneously, but decompose the data into several dependent variables (a multivariate result), offering more information to analyze. While SSMs are notably popular in the behavioural sciences, they are not significantly developed in language production research. To contribute to the development of this modelling in language, we (i) examine SSMs as a measurement modelling approach for spoken word activation dynamics, and (ii) formally compare SSMs to the default method, regression. SSMs model response activation or selection mechanisms in time, and calculate how they are affected by conditions, persons, and items. While regression procedures also model condition effects, it is only in respect to the mean RT, and little work has been previously done to compare these approaches. Through analyses of two language production experiments, we show that SSMs reproduce regression predictors, and further extend these effects through a multivariate decomposition (cognitive parameters). We also examine a combined regression-SSM approach that is hierarchical Bayesian, which can jointly model more conditions than classic SSMs, and importantly, achieve by-item modelling with other conditions. In this analysis, we found that spoken words principally differed from one another by their activation rates and production times, but not their thresholds to be activated.

KEYWORDS:

• Language research
• model-based analysis
• regression
• sequential sampling
• Bayesian modeling

Data availability statement

The data that support the findings of this study are available from the corresponding author, F.X.A., upon reasonable request.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 Note that previously in some other domains, the ex-Gaussian was considered an interesting potential model as a convolution of a decision process (exponential distribution) and a motor response process (Gaussian distribution), see Burbeck and Luce (1982); Dawson (1988).

2 For example, see Miletić et al. (2017) in which to fit just a three-alternative racing model with inhibition on a single experimental manipulation, 10,000 trials were needed and parameter recovery was arguably poor to serve as a pragmatic empirical modelling approach.

3 Rouder et al. (2015) have developed an important basis for racing modelling in the case of 3 known alternatives, and this could be a promising 3framework to explore whether it could be pragmatic for language production studies. Though, it is still likely that much larger experiments than current practice would be needed to model even at least 6 word alternatives.

4 See also Van Maanen and Van Rijn (2007); Van Maanen, Van Rijn, and Taatgen (2012) for similar accumulation mechanisms, though as richer theoretical models, they are less adapted for the type of empirical modelling that we explore in this work.

5 See also recent work by Oganian et al. (2016) for opposing effects in lexical decision revealed by the DDM.

6 Median splits are here used just for illustration of modelling experimental manipulations with 2 levels each. Note that for continuous variables, there are critiques in dichotomizing, such as an outcome of having less statistical power in significance tests (Cohen, 1983, 1990). Though in this exercise, we are rather interested in exploring whether the effects, and their sizes, between the models correspond when modelling the same experimental conditions.

7 The downside to maximal information pooling is confounds from unspecified interactions.

8 Furthermore while not shown here, that these models also provide decomposition effects for the individual participants as well.

9 With reference to personal (unpublished) data, two studies involving electromyography to measure motor time, also resulted in faster motor responses occurring during slower decision time (with slow drift/higher threshold) as compensatory behaviour in experimental conditions emphasizing speed.

10 Note that the scales of the plots from Experiment 1 and 2 are different due to the different lambdas from the Box-Cox transformation.

11 Note that certain language network models perform well in deciphering certain dynamics leading to errors (Dell, Schwartz, Martin, Saffran, & Gagnon, 1997), and these may be interpreted in conjunction with empirical cognitive models (SSMs) fit to the latency data of correct productions.

Additional information

Funding

This work, carried out within the Labex BLRI (ANR-11- LABX-0036) and the Institut Convergence ILCB (ANR-16- CONV-0002), has benefited from support from the French government, managed by the French National Agency for Research (ANR) and the Excellence Initiative of Aix-Marseille University (A*MIDEX). It was likewise supported by funding from the European Research Council under the European Community’s Seventh Framework Program (FP7/2007- 2013 Grant agreement n^o 263575).