Our primary goal here is to develop and test via simulation a bank of CDI items and IRT parameters that we can recommend to those wanting to develop and conduct CDI computerized adaptive tests (CATs). Our approach is as follows: We first fit basic IRT models (1-parameter logistic (1PL; i.e. Rasch), 2PL, and 3PL) to CDI data (French (French), WS form), and perform a model comparison. For the favored model, we then identify candidate items for removal based on low total item information, and then use the (full) item bank in a variety of computerized adaptive test (CAT) simulations on wordbank data. We provide recommendations for CAT algorithms and stopping rules to be passed on to CAT developers, and benchmark CAT performance compared to random baselines tests of a similar length.
We use the combined French (French) production data from Words & Gestures (WG, who are already producing words) and Words & Sentences (WS) CDI form, for a total of 1782 participants. From this data we remove 121 children <12 months of age, who should not be producing any words yet. We also remove an additional 56 children 12+ months of age who are not yet producing any words, as these children cannot be used to fit the IRT models. Finally, we remove 522 administrations from the Kern dataset, which show oddly high ability estimates (despite normal production sumscores). The production sumscores by age for the remaining children are shown below.
Age | N |
---|---|
12 | 35 |
13 | 91 |
14 | 39 |
15 | 26 |
16 | 59 |
17 | 35 |
18 | 37 |
19 | 50 |
20 | 28 |
21 | 39 |
22 | 63 |
23 | 76 |
24 | 74 |
25 | 52 |
26 | 27 |
27 | 40 |
28 | 26 |
29 | 41 |
30 | 38 |
31 | 29 |
32 | 49 |
33 | 42 |
34 | 42 |
35 | 33 |
36 | 10 |
37 | 1 |
39 | 1 |
We fit each type of basic IRT model (Rasch, 2PL, and 3PL) using the
mirt
package.
Compared to the Rasch model, the 2PL model fits better and is preferred by both AIC and BIC.
Model | AIC | BIC | logLik | df |
---|---|---|---|---|
Rasch | 400054.7 | 403516.0 | -199333.4 | NA |
2PL | 386243.7 | 393156.3 | -191735.8 | 692 |
The 2PL is favored over the 3PL model by both AIC and BIC.
Model | AIC | BIC | logLik | df |
---|---|---|---|---|
2PL | 386243.7 | 393156.3 | -191735.8 | NA |
3PL | 387614.7 | 397983.7 | -191728.4 | 693 |
The 2PL is preferred over both the Rasch (1PL) model and the 3PL model, so we do the rest of our analyses using the 2PL model as the basis for the CAT. Next we look for linear dependencies (LD) among the items, and also check for ill-fitting items. We will remove any items that show both strong LD and poor fit.
We examined each item for pairwise linear dependencies (LD) with other items using \(\chi^{2}\) (Chen & Thissen, 1997), and found that 3 items show strong LD (Cramer’s \(V \geq 0.5\)): “attends”, “veux”, and “dedans”.
Our next goal is to determine if all items should be included in the item bank. Items that have very bad properties should probably be dropped. We will prune any ill-fitting items (\(\chi^{2}\) \(p<.001\)) from the full 2PL model that also showed strong LD.
62 items did not fit well in the full 2PL model, and these items are shown below.
## Warning in matrix(itfit2pl$item[bad_items2pl], ncol = 4): data length [62] is
## not a sub-multiple or multiple of the number of rows [16]
non | body | vieux.vieille | les.courses |
bébé.chat | chausson.pantoufle | vilain.e | à.côté.de |
bébé.chien | pantalon | vite | vers |
hibou | pyjama | aller..verb. | ceux |
mouton | fourchette | arrêter | et |
nounours | médicaments | donner | faire |
pingouin | savon | faire.un.bisou | vert.e |
eau..drink. | téléphone | montrer | acheter |
crayon | fenêtre | prendre | courir.après |
bouche | salle.de.bain | prendre.dans.ses.bras | aller.bien.avec |
fille | eau | recevoir | tenir |
frère | avion | regarder | écouter |
gens | bon.ne | pourquoi | attendre |
grand.mère | fatigué | quand | aller |
personne | mignon.ne | quoi | non |
sœur | rouge | jeu | bébé.chat |
Now we re-fit the 2PL model without the 200 items showing strong LD and poor fit removed. (Note: should look removed/remaining items by category, difficulty, discrimination, and RMSEA.)
Next, we examine the coefficients of the 2PL model. Items that are estimated to be very easy (e.g., maman, papan) or very difficult (au sujet de, au sommet de) are highlighted, as well as those at the extremes of discrimination (a1).
Next, we will run simulated CATs on the data from the 1083 12-36 month-olds. However, since many of these participants’ data are from the CDI:WG form, there are many missing responses (compared to the CDI:WS). In order to run the simulated CATs, we impute the missing data using the participants’ estimated ability and the 2PL model. Overall, 7.3% of the data was missing, and will be imputed.
For each wordbank subject, we simulate a CAT using a maximum of 25, 50, 100, 200, 300, or 400 items, with the termination criterion that it reach an estimated SEM of .1. For each of these simulations, we examine 1) which items were never used, 2) the median and mean number of items used, 3) the correlation of ability scores estimated from the CAT and from the full CDI, and 4) the mean standard error of the CATs.
Maximum Qs | Median Qs Asked | Mean Qs Asked | r with full CDI | Mean SE | Reliability | Items Never Used |
---|---|---|---|---|---|---|
25 | 17 | 18.127 | 0.988 | 0.133 | 0.982 | 468 |
50 | 17 | 25.267 | 0.991 | 0.123 | 0.985 | 410 |
75 | 17 | 30.873 | 0.991 | 0.121 | 0.985 | 372 |
100 | 17 | 35.997 | 0.992 | 0.120 | 0.986 | 329 |
200 | 17 | 54.827 | 0.993 | 0.118 | 0.986 | 195 |
300 | 17 | 72.705 | 0.993 | 0.117 | 0.986 | 84 |
400 | 17 | 90.393 | 0.994 | 0.117 | 0.986 | 21 |
Finally, following Makransky et al. (2016), we run a series of fixed-length CAT simulations and again compare the thetas from these CATs to the ability estimates from the full CDI. The results are quite good even for 25- and 50-item tests, but note that we add a comparison to tests of randomly-selected questions (per subject), and find that ability estimates from these tests are also strongly correlated with thetas from the full CDI. The mean standard error of the random tests shows more of a difference.
Test Length | r with full CDI | Mean SE | Reliability | Items Never Used | Random Test r with full CDI | Random Test Mean SE |
---|---|---|---|---|---|---|
25 | 0.990 | 0.119 | 0.986 | 424 | 0.950 | 0.236 |
50 | 0.994 | 0.095 | 0.991 | 279 | 0.970 | 0.183 |
75 | 0.996 | 0.087 | 0.992 | 198 | 0.979 | 0.157 |
100 | 0.997 | 0.082 | 0.993 | 116 | 0.983 | 0.143 |
200 | 0.999 | 0.072 | 0.995 | 0 | 0.992 | 0.108 |
300 | 0.999 | 0.069 | 0.995 | 0 | 0.995 | 0.093 |
400 | 1.000 | 0.068 | 0.995 | 0 | 0.998 | 0.081 |
Testing with a minimum of 25 items, a maximum of 50, and termination at SE = .1, and ML scoring. First we’ll do it using the MI start item, and then we’ll try choosing an age-based starting item per subject (based on mean theta for each age).
We select a starting item with a difficulty just below the average ability (theta) for each age (in months). The mean theta per age is shown below, along with the selected starting item.
age | theta | sd | n | definition | index | item_info |
---|---|---|---|---|---|---|
12 | -1.38 | 0.77 | 35 | au.revoir | 16 | 0.98 |
13 | -1.39 | 0.67 | 91 | au.revoir | 16 | 0.96 |
14 | -1.21 | 0.66 | 39 | au.revoir | 16 | 1.25 |
15 | -0.52 | 0.53 | 26 | bain | 17 | 2.72 |
16 | -1.06 | 0.73 | 59 | au.revoir | 16 | 1.44 |
17 | -1.24 | 0.54 | 35 | au.revoir | 16 | 1.21 |
18 | -0.96 | 0.48 | 37 | bravo | 20 | 1.61 |
19 | -0.85 | 0.58 | 50 | bravo | 20 | 1.89 |
20 | -0.50 | 0.47 | 28 | chaussettes | 142 | 2.88 |
21 | -0.44 | 0.44 | 39 | chaussettes | 142 | 3.56 |
22 | -0.26 | 0.44 | 63 | couche | 148 | 6.06 |
23 | -0.20 | 0.52 | 76 | couche | 148 | 7.56 |
24 | -0.07 | 0.38 | 74 | couche | 148 | 9.68 |
25 | 0.00 | 0.46 | 52 | yeux | 122 | 11.55 |
26 | 0.02 | 0.55 | 27 | yeux | 122 | 12.16 |
27 | 0.29 | 0.47 | 40 | assiette | 166 | 13.93 |
28 | 0.65 | 0.46 | 26 | avec | 479 | 12.34 |
29 | 0.50 | 0.40 | 41 | cuisine | 440 | 15.58 |
30 | 0.53 | 0.61 | 38 | cuisine | 440 | 14.63 |
31 | 0.74 | 0.44 | 29 | parler | 622 | 14.56 |
32 | 0.79 | 0.57 | 49 | parler | 622 | 14.73 |
33 | 0.84 | 0.45 | 42 | parler | 622 | 13.78 |
34 | 0.97 | 0.44 | 42 | se.réveiller | 626 | 9.49 |
35 | 0.98 | 0.53 | 33 | se.réveiller | 626 | 9.39 |
36 | 0.96 | 0.66 | 10 | se.réveiller | 626 | 9.80 |
37 | 0.91 | NA | 1 | parler | 622 | 11.23 |
39 | 1.05 | NA | 1 | renverser | 607 | 9.39 |
Scoring / Start Item | Median Qs Asked | Mean Qs Asked | r with full CDI | Mean SE | Reliability | Items Never Used |
---|---|---|---|---|---|---|
ML / MI | 25 | 29.261 | 0.989 | 0.123 | 0.985 | 381 |
MAP / MI | 25 | 29.068 | 0.992 | 0.113 | 0.987 | 383 |
ML / age-based | 25 | 29.213 | 0.989 | 0.123 | 0.985 | 373 |
MAP / age-based | 25 | 29.018 | 0.991 | 0.113 | 0.987 | 376 |
Does the CAT show systematic errors with children of different ages? The table below shows correlations between ability estimates from the full CDI compared to the estimated ability from each fixed-length CAT split by age (139 11-13 month-olds, 129 14-16 mos, 114 17-19 mos, 139 20-22 mos, 198 23-25 mos, 102 26-28 mos, 109 29-31 mos, 124 32-35 mos, and 28 35-38 mos). This is comparable to Table 3 of Makransky et al. (2016), and the correlations here are consistently high across age groups.
Test Length | [11,14) mos | [14,17) mos | [17,20) mos | [20,23) mos | [23,26) mos | [26,29) mos | [29,32) mos | [32,35) mos | [35,38] mos |
---|---|---|---|---|---|---|---|---|---|
25 | 0.959 | 0.977 | 0.975 | 0.982 | 0.976 | 0.982 | 0.962 | 0.960 | 0.991 |
50 | 0.982 | 0.992 | 0.985 | 0.989 | 0.985 | 0.992 | 0.971 | 0.976 | 0.993 |
75 | 0.986 | 0.994 | 0.988 | 0.993 | 0.989 | 0.995 | 0.975 | 0.984 | 0.994 |
100 | 0.988 | 0.995 | 0.992 | 0.995 | 0.991 | 0.996 | 0.985 | 0.988 | 0.995 |
200 | 0.995 | 0.998 | 0.997 | 0.999 | 0.996 | 0.998 | 0.996 | 0.997 | 0.999 |
300 | 0.995 | 0.999 | 0.998 | 0.999 | 0.998 | 0.999 | 0.998 | 0.999 | 0.999 |
400 | 0.997 | 0.999 | 0.999 | 1.000 | 0.999 | 1.000 | 1.000 | 1.000 | 1.000 |
We further look at the correlations with age using the preferred CAT settings (min_items=25, max_items=50, stopping at SE=.15).
Scoring / Start Item | [11,14) mos | [14,17) mos | [17,20) mos | [20,23) mos | [23,26) mos | [26,29) mos | [29,32) mos | [32,35) mos | [35,38] mos |
---|---|---|---|---|---|---|---|---|---|
ML / MI | 0.969 | 0.983 | 0.966 | 0.982 | 0.977 | 0.981 | 0.948 | 0.959 | 0.99 |
MAP / MI | 0.977 | 0.987 | 0.977 | 0.982 | 0.976 | 0.981 | 0.961 | 0.96 | 0.991 |
ML / age-based | 0.97 | 0.978 | 0.967 | 0.982 | 0.975 | 0.982 | 0.949 | 0.957 | 0.99 |
MAP / age-based | 0.977 | 0.985 | 0.977 | 0.982 | 0.975 | 0.983 | 0.961 | 0.959 | 0.99 |
Below we show the distribution of ability (theta) from the 2PL model by age.
Finally, we ask whether the fixed-length CATs work well for children of different abilities. Below are scatterplots that show the standard error estimates vs. estimated ability (theta) for each child on the different simulated fixed-length CATs. The 25-item CAT shows some visible distortion, but the 50-item CAT is already quite smooth, and the 75-item CAT indistinguishable from the 300- or 400-item CATs. Based on these plots and the above tables we may recommend that users adopt a 50-item CAT using the 2PL parameters, but suggest that they may want to administer a full CDI if the participant’s estimated theta from the CAT is <-0.5 or >2 (where the SE from CAT starts to exceed 0.1).
Of the 632 pruned CDI:WS items, 353 were selected on one or more administrations of the fixed-length 50-item CATs simulated from the wordbank data. Which items were most frequently selected for the fixed-length 50-item CAT? Shown in the table below, only 2 items were selected on more at least 50% of the tests.
Item | Proportion |
---|---|
cuisine | 1.00 |
yeux | 0.57 |
Below we show the overall distribution of how many of the 632 pruned CDI:WS items were selected on what percent of the CATs of varying length (50, 75, or 100 items). Note that we do not include in the graph the number of items that were never selected on each test: 279 items never selected on the 50-item test, 198 items on the 75-item test, and 116 items never selected on the 100-item test. The longer the test, the less skewed the distribution, but even on the 100-item CAT most of the appearing items are selected less than a third of the time.
Below we show the 84 items from the pruned CDI:WS that were never selected on the maximum 300-item CAT.
animal | fauteuil | aider | monsieur |
glace | four | chanter | copain.ine |
œuf | frigo | essuyer | peux |
orange..food. | lavabo | faire.du.vélo.de.la.moto | fait |
petits.pois | salon | mettre | laisse.moi |
raisin | balançoire | nager | panier |
soupe | ciel | nettoyer | coussins |
sucre | école | tirer | noir |
viande | jardin | toucher | fort.e |
figure | magasin | derrière | orange |
jambe | neige | sur | triste |
langue | pelle | tous.tout | blanc.he |
culotte.slip | piscine | haricots.verts | attraper |
robe | plage | machine.à.laver | réparer |
short | travail | forêt | aimer.bien |
bouteille | content.e | pour | faire.de.la.peinture |
couverture | gentil.le | un.une | glisser |
oreiller | joli | Il | travailler |
papier | méchant.e | elle | goûter..verb. |
photo | propre | herbe | cuisiner |
serviette | sec.che | rue.route | aller |
What about the items that are most selected across all of the CATs (25-400-item)? Here are the top 50:
cuisine | pied | ballon | coucou | chocolat |
yeux | chien | bébé | entrée | miaou |
parler | chaud.e | poisson | yaourt | oiseau |
couche | voiture | merci | au.revoir | meuh |
chaussettes | chaussure | pain | canard | assiette |
renverser | cuillère | gâteau | livre | allô |
main | pomme | biberon | nom.de.l.enfant | encore |
lapin | tête | banane | avec | chut |
nez | bateau | oui | vélo | ouaf.ouaf |
bain | chat | bravo | bonjour | aïe.bobo |
These are predominantly nouns, including several body parts.
We now show an example CAT for two simulated participants, one with
ability (theta) = 0, and one with theta = 1. The CAT gives a minimum of
25 questions and terminates either when SEM=0.15 or when 50 items is
reached. The theta estimates over the test for each participant is shown
below, with selected item indices on the x axis. The theta=0 participant
(left) answered 25 questions, and the theta=1 participant (right)
answered 25. The final estimated theta for the theta=0 participant was
0.009, and for the theta=1 participant was 0.828. The package
mirtCAT
can be directly used to simply generate a web
interface (Shiny app) that allows such CATs to be run on real
participants, as well as the simulations we have conducted here.
Makransky, G., Dale, P. S., Havmose, P. and Bleses, D. (2016). An Item Response Theory–Based, Computerized Adaptive Testing Version of the MacArthur–Bates Communicative Development Inventory: Words & Sentences (CDI:WS). Journal of Speech, Language, and Hearing Research. 59(2), pp. 281-289.