shows the characteristics of the sample at baseline. The patients were predominantly White, female, and well-educated. The majority of patients had few or no mobility/IADL limitations at baseline (mean=3.7, median=2.0), but 21% of the sample had a score of 7 or higher.
Characteristics of the sample at baseline (n=248)
Mean trajectory model
A mean trajectory model using PROC MIXED with time and baseline as the only predictors indicated the effect of the first year was significant, but the effect of time was not, suggesting that after controlling for the immediate change after enrollment, functional status did not significantly change over time. These results remained essentially unchanged when potential confounders were added to the model. Mobility/IADL limitations were predicted by older age, fair or poor self-rated health, female sex, higher baseline MADRS score, less education, lower MMSE score, and lower levels of perceived social support. We tested two higher order polynomials. A likelihood ratio test comparing a model with cubic, squared, and linear terms for time to a model with only a linear term was not significant, suggesting that the effect of time controlling for the first interval was linear.
Latent class trajectory analysis
Using Latent Gold software, we determined the optimal number of latent classes by comparing models with 1–5 classes using time and baseline as the only predictors. Using the BIC statistic and the need for sufficient class size for model stability, a three-class model provided the best fit to the data. shows the results of the LCTA. Classes are numbered by size.
Results of the latent class trajectory analysis with time and baseline interval as predictors
At baseline, patients in class 1 (42%) had low scores on the mobility/IADL measure, patients in class 2 (37%) had higher scores, and patients in class 3 (21%) had few to no limitations. The Wald tests test if the intercepts and slopes were significantly different across the latent classes. The Wald test for the intercepts was significant. The effect of time, controlling for the baseline interval, did not vary across classes and was not significant. The overall baseline effect was significant, and differed by class. Patients in classes 1 and 2 had a decrease (improvement) in function score the first year, while patients in class 3 showed no significant change. The mean scores in the last column represent a weighted average of the class specific intercepts and slopes and what we would observe in a mean trajectory model.
Within the context of LCTA, we explored the relationship between variables previously shown to affect function and the trajectories (analysis not shown). We first added these variables in a manner that allowed them to affect class membership. No variables were significantly associated with class 1. In contrast, older age, being female, fewer years of education, higher baseline MADRS score, fair/poor self-perceived health, and lower perceived social support were associated with membership in class 2. Younger age, male sex, excellent/good self-rated health, and higher levels of subjective social support were associated with class 3. We then added these variables to the model allowing them to affect the slope instead of class membership for each trajectory. The effect on mobility/IADL function of age, being female, and years of education varied by class, while the effect of race and marital status did not. The effect of both health variables (self-perceived health and MMSE score) varied by class, while the effect of the clinical and social variables did not differ by class.
Functional outcomes by class using mixed models
LCTA utilizes maximum likelihood to estimate the course of functional status, and the trajectories are based on marginal means. We chose instead to use the estimated classes to model change by class using a mixed model approach. These results are shown in . The effect of time was not significant, while the effect of the first interval varied by class, as did the effect on functional status of MMSE score, MADRS score, self-perceived health, and being female. The classes did not appear to change at a differential rate, but differed primarily in the mobility/IADL score at the beginning of the study.
Results of the final mixed model showing the effect of class predicting the course of mobility/IADL function over four years controlling for potential confounders (882 observations from 248 patients)
We tested a series of models assessing interactions with class and time. None of the possible 3-way interactions (each potential confounder * time* class) or 2-way interactions with time (each potential confounder*time) was significant, and these terms were removed. The addition of the group of 2-way interactions with class (each potential confounder*class) was significant (χ2 =92.1, p<.0001), indicating some of these variables had differential impact on the level of function depending on class membership. We removed two nonsignificant interaction terms one at a time (years of education and age), and left the remaining product terms in the model.
shows the predicted trajectories from the mixed model for the three classes controlling for the effect of the potential confounders. The reference group is the patients assigned to class 3, who remain basically free of limitations. Patients in class 1 appear to have more limitations than those in class 3, but also appear to show little absolute or relative change over time after the first year. The patients in class 2 have more limitations at baseline than those in either class 1 or 3, appear to show a decrease in limitations the first year, and then appear to maintain a poor trajectory of function. The influence of the potential confounders in the data can be observed by comparing the slopes to those in the uncontrolled model presented in .
Predicted Mobility/IADL Score by Class Adjusted for Potential Confounders (MMSE score, MADRS score, self-rated health, and female sex set to the mean)
As a final step, to further understand the interpretation of the first interval, we examined change in MADRS score and change in function the first year. The change scores were significantly correlated (Pearson correlation coefficient=0.25, p<.0001). MADRS scores declined on average 14.3 points the first year, while mobility/IADL scores declined 1.1 units.