22 | Chapter 2 for maximum use of available data without excessive reliance on imputation procedures (105). A frailty score was calculated for each participant by dividing the sum of the health deficit scores by the total number of health deficits assessed. This resulted in a score between 0 (no deficits present) and 1 (all deficits present). Participants were considered to be non-frail if they had a frailty index score < 0.25 and were classified as frail when having a frailty index score ≥ 0.25 (106). Statistical analysis Data were analysed using IBM SPSS Statistics version 27 (IBM Corp. Armonk, NY) and RStudio version 1.3.1073 (RStudio Team. Boston, MA). P-values were based on two-sided tests and were a priori considered statistically significant at p < 0.05 a priori and not less than or equal to 0.05. Descriptive statistics To describe the study population at baseline, descriptive statistics (mean, median, SD, IQR) were calculated while stratifying for frailty status. Differences in baseline characteristics between non-frail and frail participants were analysed using Chi-squared tests and Mann-Whitney U tests since all continuous variables were skewed. Differences in the daily duration of physical activity between fallers and non-fallers, participants who experienced a fracture and participants who experienced no fracture, and frail and non-frail participants were estimated by Mann-Whitney U tests since physical activity was non-normally distributed. Generalized estimating equations To examine the associations between physical activity on the one hand and falls and fall-related fractures on the other hand, we used generalized estimating equations (GEE’s) with longitudinal fall and fracture data over the period of 3 years. In these analyses, all data available were included in the models to prevent a healthy survivor effect. The GEE models take into account the dependency between repeated measures within a subject (107). The GEE analyses were estimated using an exchangeable correlation matrix. For both falls and fall-related fractures as outcome measures, we analysed four models. In the first model, the association between physical activity and falls or fractures was examined. In the second model, the association between frailty and falls or fractures was determined. In the third model, physical activity, frailty and an interaction term of physical activity and frailty were included to determine whether frailty status was an effect modifier. This was done by checking whether the interaction term was statistically significant. In the fourth model, age and sex were added as covariates to model 3. We tested for a non-linear association between physical activity and falls or fractures by adding a quadratic term for physical activity, but this was not statistically significant and therefore not included in the final models. Odds ratios were estimated as well as 95% confidence intervals and p-values.
RkJQdWJsaXNoZXIy MjY0ODMw