To examine the likelihood that an obstacle-crossing task at hospital discharge will estimate fall status at 3-months post hospital discharge.
Binary logistic regression was used to examine the relationship between obstacle-crossing ability (pass/fail) at the time of hospital discharge and fall status (faller/non-faller) at 3 months post hospital discharge. The participants were dichotomized into 2 groups based on the classified fall status (faller or non-faller) at 3 months post hospital discharge. The participants were also dichotomized into 2 groups based on their obstacle-crossing ability (pass or fail) during the obstacle-crossing task (exposure) at hospital discharge.
First, between-group differences were assessed for faller/non-faller and obstacle fail/pass using chi-square tests for categorical variables and one-way ANOVA for ordinal variables or continuous variables. Alpha levels were not adjusted for multiple independent variables in order to maintain statistical power, test specific associations of interest, and inform the multivariate analysis based on recommendations by Savitz and Olsanin.169 For this study, our primary interest in analyzing group differences was to inform our multivariate analysis.
Second, an unadjusted binary logistic regression was used to examine the association between the fall status (faller/non-faller) outcome and each of the independent variables and the association between the main exposure of obstacle-crossing performance (pass/fail) and each of the independent variables. Emphasis was placed on the odds ratio (OR) and confidence intervals (e.g. width) as measures of effect and prevision. In addition, the relative risk (RR) was reported because of the high frequency of falls post stroke,63,170 with the potential for the OR to be inflated.171 Sensitivity, specificity, positive and negative predictive values, and positive and negative likelihood ratios were calculated to examine the discriminative value of obstacle-crossing ability for fall status.
Third, Spearman correlation matrices were used to determine the relationship between the continuous explanatory variables (covariates). The covariates chosen for analysis in the correlation
matrices were variables that were: (1) significantly different for those classified as fallers and non- fallers, (2) significantly different for those who failed and passed the obstacle-crossing task, and (3) those with significant bivariate associations with the exposure, obstacle-crossing, and the outcome, fall status. The correlation matrices were grouped by common assessment domains to assist in determining the covariates used in the logistic regression models. If variables were strongly correlated (rs>0.59), only one of the variables was used in the model to address possible collinearity and to maintain a parsimonious model.
Stroke severity and lower extremity motor function were considered as covariates based on previous research demonstrating a relationship between (1) stroke severity measured by the NIHSS and falls63,170 and (2) lower extremity function and balance post stroke as measured by the Step Test and falls.35 Schmid et al.63 found in a secondary data analysis (n = 1269) that the NIHSS was independently associated with post-stroke falls in an acute hospital setting (OR 3.63; 95% CI 1.46, 9.00). In addition, the same group found that increased stroke severity (NIHSS ≥ 4) was independently associated with higher fall risk (OR 5.73; 95% CI 1.65, 19.94) for a cohort of chronic stroke survivors living in the community (n = 52).170 Mackintosh et al.35 found that the combination of a fall in the hospital or during rehabilitation, combined with poor balance, provided strong predictive power for recurrent falls with sensitivity and specificity values greater than 80% for a cohort of individuals less than 6 months post stroke (n = 55).
Next, binary logistic regression models using a backward deletion method were used to examine the association between obstacle-crossing performance and fall status. For this backward deletion method, covariates with Wald P values > 0.05 were removed from the model. The models were performed with covariates found to have: (1) significant differences for those who passed and failed the obstacle-crossing task at hospital discharge (exposure) or who were fallers and non-
fallers within the first 3 months post hospital discharge (outcome), and (2) bivariate associations with the exposure, obstacle-crossing, and the primary outcome, fall status. Satisfying both an association with the outcome and main exposure would indicate that this covariate was a potential confounder. To assess for model fit, we used the overall statistical significance of the model for how well the model predicts the categories of interest (fall status and obstacle-crossing ability) compared to no independent variables and the Hosmer and Lemeshow goodness of fit test to assess how poorly the model does at predicting the categorical outcome. To assess for cases that do not fit the model well (outliers), we determined cases (participants) with standardized residuals greater than ±2.5 standard deviations. Cases with standardized residual values greater than 2.5 were inspected in further detail to determine why the cases were outliers and to determine if removal from the analysis was necessary. In addition, we also examined other patient-specific, mobility, and self-report variables for inclusion in the model as explanatory variables using the same criteria as outlined. Furthermore, we considered including variables not found to be significant in our model if they were significant factors in previous studies.In order to ensure that the multivariable model provided a robust description of the association between the predictor variables and the outcome without overfitting, an event-per-variable ratio of 10 (patients at risk of falls):1 (predictor variables) was maintained for all of the multivariable models. For this analyses, we carefully analyzed the results considering that the study was powered based on the primary analysis (unadjusted model) described above.
Sample size justification for Aim 1:
Forty-five percent of people with stroke fall within 6 months of discharge from rehabilitation35 and one in three people with stroke fail one or more trials on an obstacle-crossing
task.15 The obstacle-crossing task, in our opinion, needs to be able to identify fallers based on failing to successfully cross the obstacle with a minimum of 75 % sensitivity to be a meaningful clinical measure for detecting risk for falls. Based upon pilot data from 10 participants, where 50% of participants had fallen during the 3-month follow-up and 29% of participants who passed the obstacle test had experienced a fall, a sample size of 50 would have 80% power to detect a significant (2-tailed, alpha=.05) minimum detectable odds ratio of 5.4. Based on our pilot data, we over-recruited by 12% (target n = 56) to account for potential missing data and participants lost to follow-up. This odds ratio fits within the range of odds ratios and confidence intervals between the exposures or explanatory variables and outcome found for similar sample sizes in previous literature (Table 3.6) . These do not include obstacle-crossing ability as an exposure as the data do not exist for testing at hospital discharge post stroke. In addition, an odds ratio lower than 5.4 is likely not meaningful in a clinical setting.
Table 3.6. Summary of previous reported exposure and outcome odds ratios and confidence intervals
Author, Year Participants Time post Stroke
Exposure Outcome OR CI
Schmid et al, 2010 1269 Acute NIHSS Falls 3.63 1.46 – 9.00
Schmid et al, 2010 52 > 6 months NIHSS Falls 5.73 1.65 – 19.94
Mackintosh et al, 2006 55 < 6 months Step Test Falls 19.80 2.30 – 168.70