Statistical Analysis

Baseline (ie, Round 1) characteristics were summarized using percentages and means (SE); and these values were compared between men and women.

We calculated 4-year mortality (95% confidence intervals [CI]) for each of the seven mobility and self-care activities at Round 1 according to the different categorical scales. Before summing scores across the seven activities to form a composite score, consistent gradients in mortality were required for each activity after accounting for the precision of the estimates. An increase in mortality with a lower score (ie, better function) based on the lower bound of the 95% CI was considered a mortality reversal. These calculations were repeated for 3-year mortality using the functional activities at Round 2. Because several mortality reversals were observed for the 5-category and 4-category scales, composite disability scores were calculated only for the two 3-category scales by summing scores across the seven activities. Based on the mortality gradients observed, each activity was scored as: 1: fully able, 2: vulnerable, and 4: assistance for version c; and 1: independent, 2: difficulty, and 4: assistance for version d.

A valid functional scale should be strongly associated with relevant outcomes such as mortality and change in physical performance. We evaluated predictive accuracy of the two 3-category scales when assessed at a single time point and for changes over one year.

For each of the 3-category scales, 4-year mortality was calculated based on the composite disability scores at Round 1. The multivariable associations between each of these scores and mortality were evaluated using logistic regression models for 4-year mortality and proportional hazards models for time to death over 4 years. These models adjusted for age (in years), sex, race/ethnicity, and education. The C-statistic was used to assess the predictive accuracy of the logistic regression models, whereas the Akaike Information Criterion (AIC) value was used to assess the fit of the proportional hazards models. The model with the lowest AIC value has the best fit (or relative predictive accuracy). These analyses were repeated separately for men and women.

Changes in the composite disability scores were calculated by subtracting the Round 1 from Round 2 values. The multivariable associations between these changes and the two longitudinal outcomes were evaluated using proportional hazards models for time to death over 3 years and regression models for concurrent change in SPPB scores, subtracting the Round 1 from Round 2 values. Because the latter analysis was restricted to participants having SPPB scores at both time points, changes in the composite disability scores were recalculated for this sample. The multivariable models were adjusted for age, sex, race/ethnicity, and education. To assess model fit, the AIC value was used for the proportional hazards models, whereas the R-square was used for the regression models. These analyses were repeated separately for men and women, and the associations were evaluated within subgroups according to the Round 1 composite disability scores.

The regression coefficients from the multivariable models were used to calculate the changes in composite disability scores corresponding to a clinically meaningful change in SPPB score. The results from the proportional hazard models were then used to estimate the expected increases in mortality over 3 years corresponding to clinically meaningful changes in the composite disability scores.

All analyses were performed using the SAS Survey Procedures and account for the complex sample design (SAS, version 9.4; SAS Institute, Cary, NC). Hence, all values other than group sizes are weighted.