The difference in magnitude of the MCI-CGIC between donepezil or vitamin E treatment compared to placebo was similar to other secondary clinical outcomes - both significant and not - in a trial in which the medications did not show overall advantages compared to placebo. Based on the CGIC ratings, there was modest worsening in MCI patients over 6 and 12 months. At 12 months approximately 11% were rated improved, 50% not changed, and 40% worsened. The mean donepezil-placebo difference in CGIC of -0.13 standard deviation units (Cohen's d) at 6 months and -0.15 at 12 months was half the mean treatment-placebo differences reported in trials of cholinesterase inhibitors with mild to moderate AD patients, was similar in magnitude to that of other outcomes ratings in the MIS trial5
(), and of the same magnitude as in another MCI trial19
. Nevertheless, the lack of significant outcomes compared to placebo limits the ability to compare these effect sizes.
Treatment effect sizes at 6 and 12 months: Treatment vs. Placebo
Screening or baseline characteristics of patients, white ethnicity, higher education, and generally worse functioning on cognitive, ADLs, and depression measures, predicted worse CGIC at 6 and 12 months These predictors of worse CGIC have been identified previously as predictors of conversion to AD or more rapid cognitive decline 14,15,16
. In effect, the closer a participant was to AD the greater the likelihood for a change rating. The MCI-CGIC also predicted changes in clinical ratings scales at 6 and 12 months, thus providing evidence for its external or concurrent validity. In particular, increases in CDR-sb scores that reflect clinical progression of disease severity without reference to baseline functioning were associated with increasing CGIC ratings, providing additional evidence for the validity of this CGIC rating approach in MCI.
The CGIC is measured in ordinal categories and can be considered ordered from improved to no change to worse. Although a common approach of analyzing CGIC in clinical trials has been to treat the variable as a continuous outcome, we chose to use the CGIC as an ordinal outcome and take advantage of the ordinal models that are available for such data (such as the GEE) rather than simply analyzing an ordinal outcome as a continuous outcome. One advantage is that the logistic link function specification takes into account the ceiling and floor effects of the dependent variable, the CGIC, while the linear models do not17
. This is particularly important when the dependent variable is skewed, or when different covariate groups are compared which have widely varying skewness of the dependent variables18
We used several different regression models to evaluate the effect of the covariates on the CGIC at the 6 and 12 months assessments. The GEE modeling framework was used to assess the impact of treatment (hypothesis 1), baseline covariates (hypothesis 2) and change in clinical measures (hypothesis 3) on the CGIC at 6 and 12 months based on the analyses of change patterns over these time intervals. Fixed-effects logistic regression models were used to test the significance of the change in clinical measures at a particular time point (6 or 12 months) with the CGIC at that time. The results from the two modeling approaches had clinically similar effects, although the statistical significance differed. There are two major reasons for this: the generalization of the fixed-effects logistic model (binary or ordinal) for longitudinal data is the mixed-effects logistic regression model (binary or ordinal) and not a marginal model (GEE); and the marginal model and the mixed-effects regression model are different in the scale and interpretation of the regression coefficients. We decided to use the GEE model because the scientific interest in this study was the estimation and inference of the regression parameters and not of the variance-covariance structure of the longitudinal data. In addition, the time points in the study were fixed with little variability in terms of the timing of the measurements over time. The ease of interpretation of the odds ratio and the robustness of the model to variance-covariance misclassification makes the GEE an appropriate model for these particular analyses.
Moreover, the scientific questions that the two models answer are different. The GEE approach models the longitudinal experience of the study patients and provides information on change over a period of time, while the fixed-effects model (which models cross-sectional data) does not. Additionally, for hypothesis 3, not only does the CGIC change over time, but the values of the predictors (clinical variables) also change over time. The GEE analyses describes the dynamic relationship between the two variables in time, while the fixed-effects LR analyses describes the association between the two variables at one point in time (i.e., 6 or 12 months).
In a 6-month, randomized, placebo-controlled trial donepezil-treated MCI patients had better scores on the ADCS CGIC-MCI than did placebo patients but the differences were not significant at endpoint19
. Although mean difference data was not provided, at six-months 32.6% of donepezil patients compared to 24.3% of placebo patients showed at least minimal improvement, and 51.7% vs. 60.4%, respectively, showed no change. By comparison, in the present trial at 6 months, 19.1% vs. 11.2% showed at least minimal improvement and 49.8% vs. 58.2% showed no change (, ) yielding virtually identical rate differences. The authors of the six month trial suggest that the lack of treatment differences might be due to the impact of MCI primarily on a single domain (memory) within the CGIC. Further, they suggest that CGIC ratings are influenced by minimal impairments in other areas at baseline along with minimal changes in the primary memory impairments associated with MCI over six months. The results of the present analyses, using a larger sample size, are similar to the lack of donepezil-placebo statistically significant differences on the CGIC-MCI reported19
but suggest further that there are various influences on a CGIC rating in MCI in addition to memory20
. Thus, this study provides the first systematic evidence that CGIC ratings, at least in MCI patients, are based on more than assessments of memory.
Yet, these analyses also suggest that the CGIC, along with the other measures above, may be of more limited sensitivity and use for the more mildly cognitively-impaired MCI patients in clinical trials unless the memory domains are expanded. For the MCI amnestic subtype classification, an expanded memory domain would be appropriate. On the other hand, the lack of change in the MCI-CGIC and other outcomes may be reflecting that donepezil and vitamin E do not work in MCI patients over 6 to 12 month treatment durations.