Descriptive and Bivariate Associations
Fifteen percent of RTCI participants (N = 1,459) who transitioned from the NH to the community fell within 30 days of discharge. An overview of RTCI participant characteristics is presented in Table 1. The mean age of participants was 79.6 years, 59.5% were female and 57.4% were married. The majority (92.8%) had been admitted to the NH from an acute-care hospital and 16.6% had experienced a fall in the NH prior to discharge. Based on the BIMS, 11.7% of participants had moderate-to-severe cognitive impairment. In terms of assistance with ADLs, 69.8% needed some assistance with walking, 12% needed some assistance with toileting, and 9.7% needed some assistance with moving within bed. A majority (54.1%) of participants were using psychotropic medications and almost 40% were using analgesics prior to NH discharge. In terms of medication management, 54.7% indicated some level of assistance needed (somewhat dependent or dependent). Three quarters of participants (75.9%) had caregiver support daily or multiple times a week, and most were living with a spouse, other relative or significant other (69.8%). At discharge, 48.3% accepted skilled nursing services and 50% accepted home health aide services.
Bivariate analyses indicated statistically significant associations (alpha < .10) between falls and several of the variables considered in our conceptual framework, including health and functional variables as well as medication-related variables and caregiver support (Table 1).
EFA and CFA for the Fall-related Risk Factors
EFA with the fall-related risk variables, suggested three latent constructs: fall concerns/fall history, ADLs impairment, and use of high-risk medications. CFA was used to fit the three latent variables for the SEM models. The first item of each latent variable was set to 1 to allow for model estimation. For the fall concerns/fall history latent variable, four indicators loaded significantly, including concern with home safety (1.00), previous NH fall (1.03), fear of falling in the community (2.17), and concerns with vertigo/balance (3.10). For the ADLs impairment latent variable, three indicators loaded significantly including needing some assistance within bed movement (1.00), with toileting (0.40), and walking (−0.31). For the use of high-risk medications latent variable, three indicator variables loaded significantly, including use of analgesics (1.00), anticholinergics (0.90), and psychotropics (1.13).
Each of the three latent variable models had good fit to the data individually. We combined the three latent variables into one model to examine how well they fit the data collectively. Figure 1 shows standardized parameter estimates and correlations between latent variables in the combined model. The combined model also had good model fit with a χ 2(32) = 78.44, p < .01; CFI = .94, RMSEA = .03 (90% confidence interval [CI] = 0.02, 0.04).
Confirmatory factor analysis for fall-related risk factor latent variables. Standardized coefficients are presented. Correlations between latent variables are also presented. χ 2(32) = 78.44, p < .01; CFI = .94, RMSEA = .03 (90% confidence interval = 0.02, 0.04). ***p < .001. CFI = comparative fit index; RMSEA = root mean-square error of approximation.
EFA and CFA for the Support and Services Construct
In the EFA, we found that five indicators of supports and services loaded significantly onto a single-latent construct. We conducted CFA for these variables and the latent variable supports and services (Figure 2). The first item (receiving at least one HCBS) was set to 1. The loadings were 0.96 for use of durable medical equipment, 2.46 for caregiver support frequency, 2.61 for medication management assistance, and 3.04 for discharge location. The latent variable was influenced mainly by discharge location, medication management assistance, and caregiver support. This latent variable had good model fit with a χ 2(5) = 10.36, p = .07; CFI = .99; RMSEA = .03 (90% CI = 0.00, 0.05).
Confirmatory factor analysis for supports and services latent variable. Standardized coefficients presented. χ 2(5) = 10.36, p = .07; CFI = .99; RMSEA = .03 (90% confidence interval = 0.00, 0.05), ***p < .001. CFI = comparative fit index; HCBS = home and community-based services; RMSEA = root mean-square error of approximation.
We estimated two SEM models: one model without interaction terms and the second model with the addition of interaction terms. The initial model indicated that constructs of ADL impairment and use of high-risk medication were positively related to supports and services, while supports and services had no significant effect on falls at 30 days (Supplementary Figure 2 and Supplementary Table 1). Next, we tested the same conceptual model with interaction terms between supports and services and each of the fall-related risk constructs. Nonsignificant interaction terms were then dropped and a more parsimonious model was tested. This final model was similar to the initial model except for the inclusion of an interaction term between the latent variables of supports and services with use of high-risk medications (Figure 3). To evaluate the two SEM models (with and without the interaction term), we compared C-statistics from the ROC for each model. Based on the ROC, the model with the interaction term predicted falling within 30 days of post-discharge more accurately than the model with no interactions, with a C-statistic of 0.757 versus 0.719, respectively.
Falls SEM. The three fall risk factor latent variables are correlated (not shown for simplicity). Standardized coefficients presented. Because of the interaction term, Mplus did not provide fit statistics. Significance bolded. **p < .05, ***p < .001. SEM = structural equation model.
Results from the final model indicated a significant positive effect of the falls concern/falls history latent variable on falls, a significant positive effect of use of high-risk medications on falls, and a significant positive effect of ADL impairments on receiving supports and services (Figure 3). In addition, the interaction between supports and services and use of high-risk medications was negatively associated with falling (p = .03). Given a specific level of high-risk medication use, as receipt of supports and services increase, the risk of falling decreases. In addition, being female was negatively associated with falling while having at least one neurological condition had a significant positive effect on falling. Unstandardized and standardized model coefficients for direct effects are presented in Table 2.
Gerontologist. 2018;58(6):1075-1084. © 2018 Oxford University Press