Susan Collins and colleagues  report a sophisticated analysis of a five-wave longitudinal data set examining the relationship between readiness to change and drinking behaviours and problems in a large sample of college drinkers. They used the 12-item Readiness to Change Questionnaire (RTCQ) , and computed a single readiness to change (RTC) score for each participant by adding the total scores for pre-contemplation, contemplation and action (after first reverse-coding the pre-contemplation items), and then dividing by 12 to produce a mean item score between 1 and 5. They used this score as both a predictor and an outcome variable in their analysis and found positive predictive effects of RTC on subsequent drinking as well as positive predictive effects of drinking on subsequent RTC. In this commentary I highlight several key issues around using RTC to predict drinking behaviour which may help to inform future research on this topic.
Collins et al.  do not offer an explanation of the unanticipated positive relationship between RTC and subsequent drinking. One possibility that may be worth exploring is that it is partly a consequence of measurement error. In a multivariable analysis, the presence of random measurement error in the observed variables may lead to particular effects being under- or overestimated. Single items were used to measure the drinking variables. If the amount of random measurement error in these measures was greater than the amount of random measurement error in the RTC score (which was based on 12 items), then the effect of RTC on subsequent drinking controlling for concurrent drinking may have been overestimated, because the association between RTC and concurrent drinking was underestimated and therefore the effect of concurrent drinking was ‘under-controlled’. This could make a true zero effect of RTC on subsequent drinking appear to be a small positive effect. It might be informative to conduct sensitivity analyses in which the variables are represented as latent variables and the amount of measurement error is varied by fixing the error variances at different non-zero values .
Prediction can be maximized by ensuring that predictor and criterion are matched . In Collins et al. , predictor and criterion did not conform to what Ajzen calls the principle of compatibility . The RTCQ was completed 6 months before the drinking behaviour measures and the latter referred to ‘the past month’, whereas the RTCQ items do not specify a time-frame. The RTCQ items use phrases such as ‘doing something about it’, ‘changing my drinking’ and ‘drinking less alcohol’, which do not correspond to any of the specific drinking behaviour measures used as criteria. If a participant expresses strong agreement with the item ‘I am trying to drink less than I used to’, we do not know whether they are trying to reduce the frequency with which they drink, the quantity they drink on a typical occasion or some other aspect of their drinking behaviour.
A measure that predicts may be useful even if it is unclear what precisely it is measuring. However, if the aim is to draw causal inferences, other considerations become important , including the meaning and validity of the predictor variables. Computing a single RTC score assumes that all 12 items on the RTCQ reflect a single dimension of RTC. For example, someone who scores low on contemplation and high on action may have the same RTC score as someone who scores high on contemplation and low on action. The evidence that the RTCQ items reflect a single dimension is weak. The strongest evidence for unidimensionality would be to show in a confirmatory factor analysis that a single-factor model (with all 12 items reflecting one RTC factor) provided an excellent fit to the data. I am aware of only one study that has examined such a model . The findings supported a model with three correlated factors; the single-factor model showed a poor fit. Budd & Rollnick  also found support for a model with three correlated factors (although only after some items were allowed to load on more than one factor). A model with a single second-order factor (labelled readiness to change) also fitted the data well. However, this should not be interpreted as strong evidence for unidimensionality. A second-order factor model may provide a good fit even when the correlations between the first-order factors are relatively low (indeed, in the case of one second-order factor and three first-order factors, the correlations between the first-order factors will be explained perfectly by the second-order factor). In the absence of evidence for unidimensionality, the meaning of the RTC score remains unclear. Using their existing data, Collins and colleagues could test the dimensionality of the RTCQ and also examine whether the separate components of the RTCQ predict drinking behaviour differentially. This would be a valuable contribution to the literature.