Continuous developmental changes in word recognition and language learning across early childhood
- Department of Psychology, Stanford University, United States
- Department of Linguistics, Stanford University, United States
- Department of Psychology, University of Wisconsin, United States
- Department of Linguistics, University of California, Berkeley, United States
- Department of Communication, Stanford University, United States
- Department of Psychology, Ludwig-Maximilians-Universität München, Germany
- Department of Cognitive Science, University of California, San Diego, United States
Figures
Figure 1
Time course of word recognition at different ages.
The x-axis shows time (in ms) from the onset of the target label (vertical solid line). Colored lines show the average proportion target looking as a function of time from pre- and post-label onset at each age bin (in months). Age bins are larger for older children due to decreased data density. The dashed horizontal line represents chance looking. Error bands represent standard errors of the mean. Gray backgrounds highlight the short and long time windows used in subsequent analyses. The data within the figure is filtered such that (a) participants are required to contribute at least 5 observations and (b) there must be at least 50 participants contributing to each time bin within an age group.
Figure 2
Participant-level target looking and reaction time (log), plotted by age (log).
Longitudinal datapoints are connected by lines. The solid blue line shows a linear fit and associated confidence interval. Thin colored lines show linear fits for those datasets spanning six or more months of age. The dashed line for accuracy shows chance-level looking (0.5).
Figure 3
Participant-level variability in target looking and reaction time (log RT), plotted by age (log).
Plotting conventions are as in Figure 1.
Figure 4
Structural equation model showing the three-factor factor analysis with a regression of each latent variable on the logarithm of age.
Observed variables are notated as squares and latent variables are notated as circles. Factor loadings and regression coefficients are shown with straight, solid lines; covariances are shown with dashed lines; residual variances are shown as solid circular connections. Stars show conventional levels of statistical significance, e.g., * indicates p < 0.05, ** indicates p < 0.01, and *** indicates p < 0.001. Covariances reflect age-residualized correlations between variables.
Figure 5
Growth curves from a logistic growth model showing predicted productive vocabulary growth for children with initial reaction times one SD faster than the mean (blue), at the mean (red), and one SD slower than the mean (green).
Individual longitudinal trajectories are shown in light gray. Solid lines show global model estimates, and colored regions indicate 95% credible intervals.
Figure 6
Structural equation model showing longitudinal couplings between growth parameters.
Appendix 1—figure 1
Age distribution of unique participants for each dataset, using 3-month bins.
Appendix 1—figure 2
Distribution of retest administrations across datasets with repeated measurements, colored by dataset.
Each count indicates a retest administration (initial administrations are excluded). Administrations listed with a retest interval of 0 indicate retests within a month of the initial administration.
Appendix 2—figure 1
Correlation between reaction times on all trials and reaction times on trials where the child pointed to the correct target.
Data from Creel, 2024.
Appendix 3—figure 1
Goodness of fit for different distributional models for RT, split by age.
Appendix 3—figure 2
Distribution of RT overlaid with a log-normal distribution, split by age.
Appendix 3—figure 3
Goodness of fit for different distributional models of accuracy, split by age.
Appendix 3—figure 4
Distribution of accuracies overlaid with normal distribution, split by age.
Appendix 9—figure 1
Parallel analysis scree plot showing the eigenvalues for each factor, for actual, simulated, and resampled data.
Appendix 12—figure 1
Growth curves from a logistic growth model showing predicted productive vocabulary growth for children based on their age-residualized initial reaction times.
Predictions are shown for children with initial reaction times one SD faster than the mean for their age (blue), at the mean for their age (red), and one SD slower than the mean for their age (green). Individual longitudinal trajectories are shown in light gray. Solid lines show global model estimates and colored regions indicate 95% credible intervals.
Tables
Table 1
Characteristics of included datasets from Peekbank.
‘Admins’ denotes separate experimental sessions. ‘CDIs’ refers to whether the dataset contains parent report vocabulary data from the MacArthur–Bates Communicative Development Inventory.
| Dataset name | Pct trials (%) | N subjects | N admins | Mean age | Min age | Max age | Avg trials | Avg RT trials | CDIs | Longitudinal | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Adams et al., 2018 | 24.1 | 69 | 711 | 23.58 | 13.00 | 38.00 | 18.65 | 7.92 | x | x |
| 2 | Fernald and Marchman, 2012 | 20.1 | 122 | 679 | 23.91 | 17.00 | 32.00 | 16.23 | 6.91 | x | x |
| 3 | Weaver et al., 2024 | 8.0 | 141 | 247 | 15.74 | 13.50 | 23.60 | 18.21 | 6.78 | x | |
| 4 | Fernald et al., 2013 | 7.4 | 80 | 178 | 20.04 | 17.00 | 26.00 | 23.17 | 9.16 | x | x |
| 5 | Fernald et al., 2006 | 6.4 | 63 | 229 | 19.68 | 15.00 | 25.00 | 15.28 | 6.06 | x | x |
| 6 | Bergelson and Swingley, 2012 | 2.9 | 84 | 84 | 11.76 | 5.98 | 20.83 | 18.96 | 6.74 | x | |
| 7 | Yurovsky et al., 2013 | 2.8 | 385 | 385 | 33.77 | 12.20 | 59.51 | 5.89 | 2.63 | ||
| 8 | Borovsky and Peters, 2019 | 2.8 | 79 | 79 | 18.27 | 17.00 | 20.00 | 19.24 | 0.00 | x | |
| 9 | Potter and Lew-Williams, 2024 | 2.7 | 67 | 67 | 23.76 | 21.00 | 27.00 | 21.69 | 7.92 | ||
| 10 | Yurovsky et al., 2017a | 2.6 | 315 | 315 | 36.40 | 8.40 | 60.00 | 6.27 | 2.87 | ||
| 11 | Yurovsky and Frank, 2017b | 2.6 | 282 | 282 | 25.64 | 12.59 | 58.65 | 5.91 | 2.79 | ||
| 12 | Weaver and Saffran, 2026 | 2.4 | 64 | 64 | 18.82 | 18.10 | 20.10 | 21.82 | 8.19 | x | |
| 13 | Yoon et al., 2015 | 2.0 | 194 | 194 | 42.30 | 13.20 | 60.00 | 6.72 | 2.98 | ||
| 14 | Mahr et al., 2015 | 2.0 | 29 | 29 | 20.83 | 18.10 | 23.80 | 37.00 | 13.62 | x | |
| 15 | Garrison et al., 2020 | 1.7 | 35 | 35 | 14.46 | 12.00 | 18.00 | 27.76 | 9.41 | x | |
| 16 | Ronfard et al., 2022 | 1.3 | 40 | 40 | 19.95 | 18.00 | 24.00 | 18.56 | 7.62 | x | |
| 17 | Bacon and Saffran, 2022 | 1.3 | 38 | 38 | 22.87 | 22.00 | 24.00 | 18.08 | 8.00 | x | |
| 18 | Perry and Saffran, 2017 | 1.2 | 42 | 42 | 20.45 | 19.00 | 22.00 | 15.45 | 5.43 | x | |
| 19 | Pomper, R. & Saffran, J. R. (2017). Do infants learn to associate diminutive forms with animates? [unpublished raw data]. University of Wisconsin-Madison. | 1.1 | 76 | 76 | 16.70 | 14.00 | 19.00 | 8.71 | 3.38 | ||
| 20 | Swingley and Aslin, 2002 | 1.1 | 50 | 50 | 15.09 | 14.13 | 16.00 | 11.70 | 3.79 | x | |
| 21 | Frank et al., 2016 | 0.8 | 105 | 105 | 33.89 | 12.13 | 59.84 | 6.15 | 2.69 | ||
| 22 | Pomper and Saffran, 2016 | 0.8 | 60 | 60 | 44.27 | 41.00 | 47.00 | 7.62 | 3.30 | ||
| 23 | Moore and Bergelson, 2022 | 0.7 | 29 | 29 | 18.11 | 16.12 | 20.03 | 12.97 | 4.89 | x | |
| 24 | Pomper, R. & Saffran, JR. (2015). Modulating attention to different features of objects during word learning [Unpublished Raw Data]. University of Wisconsin-Madison. | 0.6 | 25 | 25 | 40.04 | 38.10 | 42.00 | 13.96 | 5.17 | ||
| 25 | Pomper and Saffran, 2019 | 0.4 | 44 | 44 | 40.11 | 38.00 | 43.00 | 5.32 | 2.34 | ||
| 26 | Pomper and Saffran, 2018 | 0.4 | 37 | 37 | 39.46 | 37.80 | 43.00 | 5.47 | 2.79 | ||
| Total | 100 | 2555 | 4124 | 25.38 | 5.98 | 60.00 | 14.88 | 5.51 | 14 | 5 |
Appendix 1—table 1
Characteristics of included datasets from Peekbank, sorted by what percent of the data they represent.
Percent trials refers to what percent of the trials used came from that dataset; total is the number of trials used from that dataset; and included is what percent of all trials had data that was included (based on criteria about missingness, distractor to target transition, minimum RT). LW = long-window accuracy, SW = short-window accuracy, and RT = reaction time.
| Dataset name | Pct trials (LW) | Pct trials (RT) | Total trials (LW) | Total trials (RT) | Included (LW) | Included (SW) | Included (RT) | |
|---|---|---|---|---|---|---|---|---|
| 1 | Adams et al., 2018 | 24.1% | 27.1% | 13146 | 5470 | 86.2% | 89.3% | 35.9% |
| 2 | Fernald and Marchman, 2012 | 20.1% | 22.4% | 10954 | 4526 | 91.6% | 94.8% | 38.0% |
| 3 | Weaver et al., 2024 | 8.0% | 7.6% | 4371 | 1540 | 88.6% | 91.2% | 31.3% |
| 4 | Fernald et al., 2013 | 7.4% | 7.9% | 4055 | 1603 | 83.1% | 85.7% | 32.9% |
| 5 | Fernald et al., 2006 | 6.4% | 6.2% | 3484 | 1254 | 93.1% | 93.4% | 34.0% |
| 6 | Bergelson and Swingley, 2012 | 2.9% | 2.7% | 1593 | 539 | 76.7% | 83.5% | 26.1% |
| 7 | Yurovsky et al., 2013 | 2.8% | 2.2% | 1544 | 450 | 53.2% | 55.7% | 17.2% |
| 8 | Borovsky and Peters, 2019 | 2.8% | 0.0% | 1520 | 0 | 88.9% | 88.9% | 0.0% |
| 9 | Potter and Lew-Williams, 2024 | 2.7% | 2.6% | 1453 | 523 | 89.3% | 87.7% | 32.1% |
| 10 | Yurovsky et al., 2017a | 2.6% | 2.4% | 1411 | 493 | 60.4% | 68.2% | 21.9% |
| 11 | Yurovsky and Frank, 2017b | 2.6% | 1.9% | 1401 | 393 | 65.3% | 66.6% | 20.9% |
| 12 | Weaver and Saffran, 2026 | 2.4% | 2.4% | 1309 | 475 | 67.5% | 69.2% | 24.5% |
| 13 | Yoon et al., 2015 | 2.0% | 1.7% | 1089 | 337 | 74.4% | 77.1% | 24.9% |
| 14 | Mahr et al., 2015 | 2.0% | 2.0% | 1073 | 395 | 82.4% | 82.4% | 30.3% |
| 15 | Garrison et al., 2020 | 1.7% | 1.6% | 944 | 320 | 89.8% | 89.9% | 30.4% |
| 16 | Ronfard et al., 2022 | 1.3% | 1.4% | 724 | 282 | 98.5% | 98.5% | 38.5% |
| 17 | Bacon and Saffran, 2022 | 1.3% | 1.5% | 687 | 304 | 90.4% | 89.2% | 40.0% |
| 18 | Perry and Saffran, 2017 | 1.2% | 1.1% | 649 | 228 | 89.6% | 89.6% | 31.5% |
| 19 | Pomper, R. & Saffran, J. R. (2017). Do infants learn to associate diminutive forms with animates? [unpublished raw data]. University of Wisconsin-Madison. | 1.1% | 0.9% | 592 | 179 | 75.0% | 75.8% | 24.5% |
| 20 | Swingley and Aslin, 2002 | 1.1% | 0.8% | 585 | 159 | 98.2% | 97.8% | 28.0% |
| 21 | Frank et al., 2016 | 0.8% | 0.7% | 449 | 140 | 59.5% | 66.2% | 20.3% |
| 22 | Pomper and Saffran, 2016 | 0.8% | 0.9% | 457 | 175 | 95.6% | 95.6% | 37.9% |
| 23 | Moore and Bergelson, 2022 | 0.7% | 0.7% | 376 | 132 | 97.9% | 99.2% | 34.9% |
| 24 | Pomper, R. & Saffran, JR. (2015). Modulating attention to different features of objects during word learning [Unpublished Raw Data]. University of Wisconsin-Madison. | 0.6% | 0.6% | 335 | 124 | 85.7% | 89.0% | 31.9% |
| 25 | Pomper and Saffran, 2019 | 0.4% | 0.4% | 213 | 75 | 85.9% | 89.3% | 30.5% |
| 26 | Pomper and Saffran, 2018 | 0.4% | 0.3% | 197 | 67 | 90.5% | 91.9% | 34.4% |
Appendix 5—table 1
Pairwise correlations between primary variables of interest.
| age | log age | rt | log rt | long acc | short acc | prod | comp | |
|---|---|---|---|---|---|---|---|---|
| age | 1.00 | |||||||
| log age | 0.98 | 1.00 | ||||||
| rt | –0.33 | –0.35 | 1.00 | |||||
| log rt | –0.34 | –0.36 | 0.96 | 1.00 | ||||
| long window accuracy | 0.44 | 0.48 | –0.48 | –0.46 | 1.00 | |||
| short window accuracy | 0.38 | 0.43 | –0.62 | –0.61 | 0.82 | 1.00 | ||
| production vocabulary | 0.72 | 0.70 | –0.31 | –0.33 | 0.51 | 0.45 | 1.00 | |
| comprehension vocabulary | 0.42 | 0.42 | –0.25 | –0.24 | 0.24 | 0.24 | 0.59 | 1.00 |
Appendix 6—table 1
Model comparison metrics for different functional forms of the relationship between accuracy and age.
| Model | n. obs | sigma | logLik | AIC | BIC | REMLcrit | df.residual | r2 |
|---|---|---|---|---|---|---|---|---|
| Long window, linear age | 55,337 | 0.272 | –7012 | 14,038 | 14,100 | 14,024 | 55,330 | 0.124 |
| Long window, log age | 55,337 | 0.271 | –6972 | 13,957 | 14,020 | 13,943 | 55,330 | 0.108 |
| Short window, linear age | 57,045 | 0.309 | –14534 | 29,082 | 29,145 | 29,068 | 57,038 | 0.092 |
| Short window, log age | 57,045 | 0.309 | –14,501 | 29,016 | 29,079 | 29,002 | 57,038 | 0.077 |
Appendix 6—table 2
Model comparison metrics for different functional forms of the relationship between RT and age.
| Model | n. obs | sigma | logLik | AIC | BIC | REMLcrit | df.residual | r2 |
|---|---|---|---|---|---|---|---|---|
| Log RT, linear age | 20,689 | 0.458 | –13,844 | 27,703 | 27,758 | 27,689 | 20,682 | 0.232 |
| Log RT, log age | 20,689 | 0.457 | –13,819 | 27,651 | 27,707 | 27,637 | 20,682 | 0.212 |
| Linear RT, linear age | 20,689 | 578.837 | –161,573 | 323,159 | 323,215 | 323,145 | 20,682 | 0.197 |
| Linear RT, log age | 20,689 | 578.338 | –161,546 | 323,106 | 323,161 | 323,092 | 20,682 | 0.185 |
Appendix 7—table 1
Goodness of fit comparison between different models of the relationship between age and RT.
| Model | n. obs | sigma | logLik | AIC | BIC | REMLcrit | df.residual | r2 |
|---|---|---|---|---|---|---|---|---|
| Log RT, log age | 18,940 | 0.451 | –12,393 | 24,801 | 24,855 | 24,787 | 18,933 | 0.205 |
| Log RT, log + linear age | 18,940 | 0.449 | –12,349 | 24,719 | 24,806 | 24,697 | 18,929 | 0.216 |
| Linear RT, poly(age,2) | 18,940 | 569.500 | –147,594 | 295,204 | 295,267 | 295,188 | 18,932 | 0.180 |
| Linear RT, poly(age,3) | 18,940 | 569.005 | –147,567 | 295,151 | 295,222 | 295,133 | 18,931 | 0.188 |
Appendix 7—table 2
Fixed effects coefficients for a model predicting log RT from both log age and linear age.
| Term | Estimate | Std. error | df | t value | p-value |
|---|---|---|---|---|---|
| Intercept | 6.829 | 0.026 | 20.104 | 266.137 | 0.000 |
| log age | –0.251 | 0.072 | 12.851 | –3.515 | 0.004 |
| linear age | 0.132 | 0.078 | 13.232 | 1.697 | 0.113 |
Appendix 9—table 1
Factor loadings for the exploratory three factor solution using varimax rotation.
| F1 | F2 | F3 | |
|---|---|---|---|
| RT | –0.19 | 0.81 | –0.30 |
| RT var | –0.10 | 0.81 | –0.22 |
| long window accuracy | 0.33 | –0.31 | 0.55 |
| long window accuracy var | –0.10 | 0.26 | –0.65 |
| production vocabulary | 0.95 | –0.04 | 0.30 |
| comprehension vocabulary | 0.61 | –0.16 | –0.01 |
| age | 0.63 | –0.14 | 0.37 |
Appendix 10—table 1
Comparison of confirmatory factor analysis models on longitudinal data or first administrations only.
| Model | CFI | RMSEA | RMSEALower | RMSEAUpper | TLI | SRMR | AIC | BIC |
|---|---|---|---|---|---|---|---|---|
| No age, longitudinal | 0.980 | 0.058 | 0.047 | 0.069 | 0.949 | 0.033 | 50,821 | 50,952 |
| No age, first admin | 0.992 | 0.033 | 0.018 | 0.049 | 0.980 | 0.023 | 27,464 | 27,585 |
| Age, longitudinal | 0.993 | 0.033 | 0.024 | 0.042 | 0.984 | 0.021 | 48,565 | 48,717 |
| Age, first admin | 0.997 | 0.021 | 0.007 | 0.034 | 0.993 | 0.013 | 26,173 | 26,313 |
Appendix 11—table 1
Model comparison for alternative factor structures.
p-values show differences between adjacent models; no p-values are shown for comparisons between non-nested models.
| df | AIC | BIC | Chisq | Chisq diff | RMSEA | df diff | Pr(>Chisq) | |
|---|---|---|---|---|---|---|---|---|
| Three-factor | 6 | 46,346.69 | 46,475.43 | 91.96 | ||||
| Two-factor (vocab) | 8 | 46,397.74 | 46,514.22 | 147.01 | 55.05 | 0.09 | 2 | <0.0001 |
| Two-factor (speed) | 8 | 46,486.96 | 46,603.44 | 236.23 | 89.22 | 0.00 | 0 | |
| Two-factor (variability) | 8 | 46,536.10 | 46,652.58 | 285.37 | 49.14 | 0 | 0 | |
| One-factor | 9 | 46,535.49 | 46,645.84 | 286.76 | 1.39 | 0.01 | 1 | 0.24 |
Appendix 12—table 1
Fixed effects estimates from logistic growth model.
| Type | Estimate | Est. error | l-95% CI | u-95% CI |
|---|---|---|---|---|
| Constant component of growth intercept | 2.850 | 0.606 | 1.680 | 4.109 |
| Effect oft0 RT on growth intercept | 3.187 | 0.604 | 1.970 | 4.369 |
| Constant component of growth scale | 1.121 | 0.058 | 1.009 | 1.242 |
| Effect oft0 RT on growth scale | –0.026 | 0.079 | –0.178 | 0.133 |
Appendix 12—table 2
Fixed effects estimates from logistic growth model using RT residualized on age as the predictor.
| Type | Estimate | Est. error | l-95% CI | u-95% CI |
|---|---|---|---|---|
| Constant component of growth intercept | 2.798 | 0.570 | 1.740 | 3.987 |
| Effect of residualizedt0 RT on growth intercept | 3.237 | 0.562 | 2.119 | 4.340 |
| Constant component of growth scale | 1.104 | 0.061 | 0.986 | 1.232 |
| Effect of residualizedt0 RT on growth scale | 0.056 | 0.079 | –0.097 | 0.214 |
Appendix 13—table 1
Fraction of data present for each measure at each time point for the longitudinal SEM.
| Time point | Log RT | Accuracy | Production | Comprehension |
|---|---|---|---|---|
| t0 | 0.685 | 0.865 | 0.279 | 0.183 |
| t1 | 0.035 | 0.039 | 0.024 | 0.013 |
| t2 | 0.093 | 0.096 | 0.095 | 0.000 |
| t3 | 0.028 | 0.029 | 0.027 | 0.000 |
| t4 | 0.031 | 0.032 | 0.031 | 0.000 |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Continuous developmental changes in word recognition and language learning across early childhood
eLife 14:RP109636.
https://doi.org/10.7554/eLife.109636.3
