5.7 Other modelling actions
5.7.1 Instrument fit
Some instruments fit better than others. Figure 5.6 shows the box plots of outfit per instrument. Instruments
vin generally fit well, whereas discrepancies between model and data are larger for
sbi. Through additional modelling, we found that it was extremely difficult to get enough high-quality bridge items that could link
bat (Battelle Development Inventory) to the other instruments. We also found that models without the Battelle were able to better discriminate children in the upper range of the D-score scale. We therefore opted to remove
bat from the model, even though this meant that one cohort (
GCDG-BRA-2) had to be dropped from the analysis.
It is not clear why
bat does not fit. Perhaps the scoring system of the Battelle in three categories invokes scoring behaviour that is different from the PASS/FAIL scoring used by most other instruments, even though this appears to be less of a troublesome aspect in
aqi, which also uses three response categories.
5.7.2 Splitting, combining and selecting equate groups
Most of the modelling effort went into finding a set of high-quality equate groups that link the instruments. For example, we tried to bridge the South-African study placing
vinxxc016 (uses a short sentence) into
EXP26 (two-word sentences) and
EXP36 (sentences of 3 or more words), but neither option led to a reasonable model. On the surface, milestone
by3gmd060 (balances on right foot, 2 seconds) appears to fit within
GM60 (balances on foot), but the analysis showed large discrepancies with the other items in the groups, so it had to be taken out.
Subject-matter experts identified 38 items that were thought to be cross-culturally incompatible. Table 5.1 provides an overview. Many of such milestones involve a specific language concept (such as a pronoun), refer to stairs (less common in rural settings), help in house or clothing behaviour. These items have different meanings in different contexts, so they were not used to bridge instruments.