Subjects who missed scheduled follow-up visits were contacted by site investigators and if they choose to discontinue participation in the study were asked to return for one final cognitive exam. Clinical personnel obtained additional information at the final visit or via the telephone if the subject refused to return. The additional information included predefined questions regarding the reasons for study discontinuation. The predefined reasons queried were concern over placebo condition, length of protocol, travel requirements, frequency of assessments, potential adverse effects of medication, concern about medication effectiveness, alternative treatment chosen by subject, alternative treatment chosen by caregiver, or caregiver unwilling to continue.

The analysis plan for the present investigation consisted of univariate assessments of possible predictors of study discontinuation followed by a model-building exercise to derive a single multivariate predictive model. The outcome variable was time to discontinuation of the closed-label portion of the trial, defined as the time from baseline to the earliest of the date of progression to dementia or the date of last completed cognitive diagnostic assessment. Subjects who did not complete the closed-label portion of the trial were considered events, subjects who converted to dementia were censored at the date of dementia diagnosis, and subjects who completed the trial without converting to dementia were censored at the date of their month 36 assessment.

Univariate assessments included unweighted log-rank tests of overall association of the covariate with time to study discontinuation, descriptive summaries of the proportion discontinuing by level of the covariate, and Cox proportional hazards model estimates of the relative risk of discontinuation by level of the covariate. Variables significantly related to study discontinuation in univariate log-rank tests at the p<0.1 level were selected for inclusion in the final model; variables related at the p=0.10–0.15 level were subsequently tested for inclusion in the final model and selected for inclusion if likelihood ratio tests to add the term were significant at the p<0.05 level or if there was evidence of confounding by the variable. A covariate was considered to have a mediating or confounding effect if its inclusion in the model altered the model coefficients for other covariate terms by greater than 15% (Hosmer et al.¹⁶). Interactions with treatment arm were assessed for all variables in the final model and tested with likelihood ratio tests with a Bonferonni adjusted p-value for an overall significance level of 0.05. Interactions significant at the p<0.05 level were included in the final model. The proportional hazards assumption was checked for all term in the final model by Schoenfeld residuals tests using the log transform of time¹⁷. To investigate the extent to which observed associations were driven by deaths and other serious adverse events, we performed two sensitivity analyses, first refitting the final model with censoring on death, and again refitting the final model with censoring on both death and discontinuation due to other serious adverse events. The distribution of continuous variables was compared using the Wilcoxon signed rank test and the frequency of categorical variables was compared using Fisher’s exact test, as indicated. All analyses were conducted using the R statistical programming language¹⁸, with log-rank tests and Cox modeling performed using the R library ‘survival’.

Baseline characteristics by discontinuation status are presented in Table 1. Subjects who dropped were different from those who continued in the study on several demographic measures. In univariate analyses, risk of study discontinuation was higher among females (HR: 1.36, p=0.020), among non-white subjects compared to white, non-Hispanic subjects (HR: 2.12, p<0.001), and among those with less than high school education (HR: 1.51, p=0.047). Rates of study discontinuation were highest at commercial sites (43.8%), and lowest at ADCs (23.0%, p<0.001). University-affiliated sites had slightly higher rates of study discontinuation compared to ADCs (28.5%, p=0.150). Subjects who were married had lower rates of study discontinuation than their unmarried counterparts (27.0% vs 39.8%, p=0.001). Only one clinical measure was a statistically significant predictor of study discontinuation in the univariate analysis: subjects with a Ham-D score of ≥6 to 12 were more likely to discontinue than subjects scoring < 6 (HR: 1.6, p=0.004). Subjects in the donepezil arm were more likely to discontinue than subjects in the placebo arm (HR: 2.2, p<0.001).

Per our pre-specified model building algorithm, covariates selected for the final multivariate model based on a univariate log-rank test p-value < 0.1 were treatment arm, site classification, marital status/informant relationship, education, gender, race/ethnicity, and HAM-D score. Age was not significantly related to study discontinuation at the p<0.1 level, but was forced into the model. Of potential pairwise interactions, only the gender by marital status term was statistically significant (likelihood ratio test p-value < 0.0001); this interaction was added to the model using married men as the referent gender by marital status category. The two measures with univariate p-values between 0.1 and 0.15, the Beck Informant Depression Scale and the primary language indicator, were added separately to the model and tested for significance. Neither significantly improved the model (likelihood ratio test p-values > 0.3), nor did their inclusion modify coefficients of covariates already in the model by more that 15%, and therefore these terms were not included in the final model. Proportional hazards assumptions were not rejected for any terms in the model (p>0.3).

The final multivariate model is presented in Table 2. Hazard ratios and p-values in the full multivariate model are largely consist with those observe in the univariate analyses, although in the final model we observe that the effect of marital status on discontinuation is present only within men (HR=2.24, p=0.001 for unmarried versus married men). The gender effect remains, but individual terms in the final model comparing married and unmarried women to the referent married men category are not statistically significant because of the reduced per comparison sample size. If these two female categories are pooled, the gender effect is statistically significant (HR=1.34, p=0.044 for women versus married men).

When we reran the model censoring subjects who died rather than counting them as study discontinuation events, the coefficients and p-values for all of the statistically significant predictors in Table 2 were not appreciably changed except for the HAM-D term, which was attenuated and no longer a significant predictor of study discontinuation (HR 1.21, p=0.27). When we reran the model further censoring both deaths and other serious adverse events, the coefficients for all of the significant predictors were not appreciably changed, although in this model the p-value for the education less than high school term became only marginally statistically significant (p-value changing from 0.02 to 0.07 after censoring of deaths and other serious adverse events).

We also found that subjects with more depression symptoms reported on the HAM-D were more likely to discontinue the trial. While we did not perform a formal competing risk analysis, when we refit the final model with censoring of those subjects who discontinued due to death, the HAM-D association with study termination was attenuated; this suggests that the association was mainly due to depressive symptoms predicting death during the trial. In fact, subjects with a HAM-D greater than or equal to six were more likely to die during the trial (5.6% died versus 1.6% among subjects with a HAM-D less than six, Fisher exact test p-value = 0.016).

We also found that subjects with less educational attainment and non-white subjects had a higher rate of study discontinuation. Although minority race or ethnicity is a commonly perceived barrier to recruitment and retention^6,7, there are little published data on clinical trial retention rates as a function of race or ethnicity, and we know of no published data on this issue within the elderly.

We found that commercial recruitment sites had twice the study discontinuation rate of non-commercial sites. All sites attended the same protocol training meetings and followed the same protocol. However, the commercial sites differ from the non-commercial sites in a number of ways. There are some indications the commercial sites recruited from a different pool of study subjects. While there was no difference between commercial and non-commercial sites in terms of the distribution of age, gender, and ethnicity, the subjects recruited by commercial sites had slightly less education (mean 14.1 years vs 14.8 years, Wilcoxon sign rank test p-value = 0.003). Lower education did predict a higher dropout rate in our sample (Table 1). However, in the multivariate analysis with both education and site in the model the HR for commercial sites was unchanged and remained highly statistically significant (Table 2), indicating that the site effect was not mediated by education in any appreciable way. We also observed that, among subjects who withdrew consent, caregivers from commercial sites were more likely to endorse as a reason for study discontinuation “Concern about effectiveness of medications” (26.0% versus 14.5% for non-commercial sites). It is possible that recruitment methods at commercial sites increased the proportion of subjects seeking active treatment via enrollment in clinical trials as opposed to subjects participating for more altruistic motivations. Subjects who seek active treatment via enrollment in clinical trials may be more likely to discontinue early if they perceive they are randomized to an arm with inactive or ineffective treatment (as reviewed in Vozdolska et al.²³). Subjects from commercial sites were also more likely to die during the trial or to be discontinued because of other serious adverse events (Table 4). This does not entirely explain the higher loss rate in commercial sites, however. When the multivariate survival analysis was repeated with censoring of the deaths and adverse event-related discontinuations, the HR for commercial sites remained statistically significant (p = 0.0002) and was attenuated only modestly, from HR = 2.21 to HR = 2.09.

Beyond study subject characteristics, we also suspect that staff members at ADCs and university-affiliated sites are more familiar with the scientific research process and more sensitive to the importance of complete data for the validity of a trial. Staff members at ADCs and university-affiliated sites often have developed skills and standards for maintaining study subject continuity in the course of their own research. These skills and standards may transfer to their performance of recruitment and retention for multicenter clinical trials. Furthermore, academic sites are more likely to recruit subjects from their own clinic populations, while commercial sites tend to rely more heavily on advertising; longstanding relationships among staff and subjects at academic sites may improve retention at these sites.

Early study discontinuation at commercial sites did reduce the statistical power of the ADCS MCI trial. Average person-years of observation for subjects recruited by commercial sites was 1.75 person-years per subject recruited, compared to 2.01 person-years per subject recruited by ADC sites and 1.95 person-years per subject recruited by non-ADC university-affiliated sites. Compounding the high study discontinuation rate was a low incident dementia rate among subjects recruited by commercial sites (Table 3). There were only 10.4 incident dementia cases per 100 person-years of observation at the commercial sites compared to 16.6 and 14.4 incident dementia cases per 100 person-years of observation at the ADC and university-affiliated sites. We speculate that this is can be attributed to a lack of specificity of diagnosis for amnestic MCI by commercial sites when screening subjects for inclusion in the ADCS MCI trial (the inclusion criteria specified that non-amnestic subtypes of MCI should not be enrolled). Clinicians at commercial sites may have been less sensitive than their ADC and university-affiliated research counterparts to those inclusion criteria intended to narrow recruitment to the amnestic subtype of MCI that is most consistent with prodromal AD.