The issue of differentiating transient from persistent problems in morphological development is central to the study of ‘late talkers’ (LTs). One of the most common criteria for classifying children as LTs has been total productive vocabulary at or below the 10th percentile on the CDI-WS or the CDI-UKSF (Dale et al., 2003; Heilmann et al., 2005). LTs tend to be identified on average around 24 months (with a range of 18–30 months). At a group level, LTs are at a substantially elevated risk for continuing language impairment; however many low scoring children catch up in later preschool years (Dale et al., 2003; Thal, Tobias, & Morrison, 1991; Thal et al., 1999). For clinical purposes group level analysis is not sufficient. Clinicians want information on individual ‘caseness’. They want to know the likelihood that children who do poorly on parent reports are the same children whose performance is low on later standardised or accepted measures that are used to define language or cognitive delay (‘gold standards’).
Westerlund, Berglund and Eriksson (2006) argued that if the prevalence of a disorder is low and the level of severity is high, as is the case with severe language disability, it is important to find as many cases as possible. Accordingly, priority should be given to the high sensitivity of a measure (proportion of true positives identified), even at the cost of low specificity (proportion of true negatives identified). There are only a handful of studies that have looked at the sensitivity of the CDI, and overall the findings are not encouraging. For example, Westerlund et al. (2006) used the Swedish Communication Screening (SCS18), a short screening version of the Swedish CDI (SECDI), in a population sample of toddlers. They found that productive vocabulary was the best predictor of the three SCS18 variables, but its sensitivity (50%) was too low to be clinically valid. They concluded that 18 months was too early to predict severe language impairment in 3-year-olds. However, 6 months later, the picture looks similar. Dale et al. (2003), using the CDI-SF in a large scale genetic study of 8386 2-year-old twins in the UK, found the predictive value of the CDI was adequate, but its sensitivity was poor. They concluded that ‘vocabulary at 2 is a predictor of poor language at 3 and 4’, but ‘too poor to be of practical utility in discriminating persistent and transient difficulties’ (p.555). Both the SCS18 and the CDI-SF are shortened versions of the CDIs, and as such are particularly suitable for use as population screens. But as Law et al. (2000) argued, although short screening measures are appealing options, the evidence to date is that they do not deliver the type of data that is really useful in a clinical context. One possibility is that the longer version, the CDI-WS, may prove more sensitive. In many respects, the results of a follow-up study at 30 months of 100 toddlers (38 LTs and 62 with a history of normal language development, according to reported productive vocabulary on the CDI-WS at 24 months) look more promising (Heilmann et al., 2005). However, despite excellent specificity (98%), sensitivity of the CDI-WS at the optimal cut-off (11th percentile) although substantially higher than the short versions, in clinical terms remained modest (68%). Although sensitivity is a key measure, tests rarely report these figures. A recent review of 43 tests of early language development revealed that measures of sensitivity and specificity were reported in 9, of which 5 only reached acceptable levels of sensitivity (Spaulding, Plante, & Farinella, 2006).