Comparison of Functional Status Improvements Among Patients With Stroke Receiving Postacute Care in Inpatient Rehabilitation vs Skilled Nursing Facilities

This cohort study examines the difference in improvement of functional status among patients with stroke who received postacute care in inpatient rehabilitation vs skilled nursing facilities.


Sample Selection
The study sample included Medicare beneficiaries age ≥66 years discharged from January 1, 2013 to November 30, 2014 to an inpatient rehabilitation facility (IRF) or skilled nursing facility (SNF) following an index acute stay for stroke, denoted by Medicare Severity Diagnosis Related Group codes 061-066. Additional inclusion criteria included 1) Medicare Part A coverage without enrollment in a Health Maintenance Organization in the year before and 1 month after the index stroke discharge, 2) residing in the community prior to the index stroke hospitalization and 3) with full mobility and self-care functional measures at the IRF admission and discharge or SNF admission and last follow-up (eFigure 1). In this cohort section, we removed 7,639 observations due to incomplete data for function scores in SNF. To check for differences in the baseline variables between those who remained in the analysis and those excluded because of missing outcome data, we compared the original sample and the sub-cohort which excluded the 7.1% with incomplete data for function scores in SNFs. Even though some of the Chi-square p-values are significant, none of the absolute standardized differences were > 0.1. This means these two groups are balanced. We included this information in the supplementary materials (eTable 1 and eTable 2).

Functional Measures: Mobility and Self-care
Post-acute care assessment data were used to characterize both stroke severity as well as outcomes of care using admission and discharge functional measures. The Centers for Medicare and Medicaid Services (CMS) documentation requirements for the Inpatient Rehabilitation Facility-Patient Assessment Instrument 1 and the Minimum Data Set 3.0 2 both contain functional information used to document changes in patient abilities from admission to discharge (in the Inpatient Rehabilitation Facility-Patient Assessment Instrument) or from admission to follow-up assessments (in the Minimum Data Set 3.0). In this study, we utilized only the mobility and self-care test items from these instruments (eTable 3). The Inpatient Rehabilitation Facility-Patient Assessment Instrument has five mobility items: bed/chair/wheelchair transfer, toilet transfer, tub/shower transfer, locomotion-walk/wheelchair, and locomotion-stairs; and six self-care items: eating, grooming, bathing, dressing-upper extremity, dressing-lower extremity, and toileting, scored with a 7-point rating scale (1-total assistant to 7-complete independence). Similarly, the Minimum Data Set consists of six mobility items: including bed mobility, transfer, walk in room, walk in corridor, locomotion on unit, and locomotion off unit; and five self-care items: eating, personal hygiene, bathing, dressing, and toilet use. The Minimum Data Set 3.0 has two different rating scales: a 5-point self-performance rating scale (0-independnet to 4-total dependence) and a 4-point support rating scale (0-no to 3-more than two persons' physical assistance).
Using data from the Inpatient Rehabilitation Facility-Patient Assessment Instrument and the Minimum Data Set 3.0, we used the crosswalk developed by Mallinson and colleagues (2011) to construct comparable admission and discharge functional scores for the two post-acute care settings. 3 The co-calibrated crosswalk contains mobility and self-care domains and was established using a Rasch common-person equating method. 4 This measurement method has demonstrated efficacy for equating similar activities of daily living instruments. 5,6 For instance, crosswalks have been successfully developed between the Functional Independence Measure and the Minimum Data Set for veteran patients, and between the Functional Independence Measure and the modified Barthel Index for Korean adult patients. These crosswalks also demonstrated good psychometric properties. The cocalibrated functional scores at admission and discharge for mobility and self-care were reported on a 0-100 point scale, where higher scores indicate greater functional status, which is consistent with the Rasch measurement method.

Covariates -Hierarchical Condition Categories
We used CMS Hierarchical Condition Categories for comorbidities identified through diagnoses on the inpatient claims from the previous year and the secondary diagnoses during the index stroke hospitalization. While the primary purpose of CMS Hierarchical Condition Categories is for adjusting risk in the Medicare Advantage payment plans, it can also be used to risk-adjust patient-level health conditions. 7 Among the 79 categories, we used the 30 most frequent CMS Hierarchical Condition Categories (yes, no) in the analysis. 8

Statistical analysis
We began with unadjusted bivariate analyses of all variables compared across IRF and SNF settings. We employed several analytic approaches to control for potential confounders, including multivariable analysis, instrumental variable analysis and inverse probability weighting with propensity scores. The multivariable approach used ordinary least squares adjusting for covariates. Next, we employed inverse probability treatment weighting with propensity scores with and without multilevel adjustment.
The propensity score was generated with a logistic regression model using average treatment effect estimation that incorporated all covariates listed in eTables 4 and 5. 9 To assess the balance of covariates, we report p-values from Chi-square tests with propensity scores weighting. Average treatment was estimated with the weight using the inverse probability of receiving the care delivered to patients. A normalized weight was used to avoid extreme propensity scores by dividing each individual propensity score by the mean of all propensity scores. Next, we used hierarchical general linear mixed models to account for patients nested within hospitals. In these models, we used inverse probability treatment weighting with propensity scores adjusted for the patient and hospital characteristics which were not balanced after propensity score weighting. Additionally, we used ordinary least squares models with inverse probability treatment weighting with propensity scores also adjusted for unbalanced covariates, to compare functional status outcomes (mobility and self-care) at discharge between IRFs and SNFs.
As a complement to the ordinary least squares and hierarchical general linear mixed modeling, we employed instrumental variable analysis to adjust for unmeasured confounders across patients and facilities. 10 The instrumental variables included: (1) the difference between the distance from the acute hospital to the nearest IRF and the distance from the acute hospital to the nearest SNF; (2) the difference between the distance from the beneficiary's residence to the nearest IRF and the distance from the beneficiary's residence to the nearest SNF; (3) the number of stroke patients in the hospital referral region in 2013-2014 who are discharged to an IRF, divided by the number discharged to an IRF or SNF, expressed as a percentage; and (4) the previous discharge location assignment (IRF or SNF) for patients with the same type of stroke from the same acute hospital (eTable 7). The third and fourth instrumental variables were determined using the 142,542 stroke discharges selected during the study sample selection (step 7, eFigure 1). For the fourth instrumental variable, in the case of a tie between the previous discharges to IRFs and SNFs, a missing value was assigned for the index stroke discharge. Also, for a given hospital and a given stroke type, the first index discharge in the study period would have a missing value for this variable. Therefore, 91,016 records with non-missing data remained for this instrumental variable analysis. Prior to the analyses, for each instrumental variable, we examined the balance in covariates' distribution using standardized difference. For the first three instrumental variables, we created binary variables based on the median and investigated the balance between the two groups. The standardized differences were calculated using the SAS macro %stddiff 11 and are presented in eTable 8. A difference of less than 0.1 (or 10%) is generally considered a small difference between two groups. 12 Instrumental variable modeling was done using R's ivmodel package, 13 treating as endogenous the same adjustment variables used in propensity modeling, and reporting the two-stage least square estimates.
Partial F tests were used for the null hypothesis of no association between each instrument and SNF versus IRF care. To evaluate the strength of instrumental variables, we considered the first-stage F statistic for both mobility and self-care outcomes. The lowest F statistics were for the differential distance between the beneficiary and the nearest IRF or SNF (F=642.9), followed by differential distance between the acute hospital and the nearest IRF or SNF (F=2105.4), previous IRF or SNF assignment from the acute hospital by stroke type (F=2501.2) and percentage of IRF patients within the hospital referral region (F=5032.6). In other words, all F statistics were greater than 630. An F statistic greater than 10 suggests that an instrument is not weak. We utilized the Hausman test to examine if any variables were correlated with the residuals in the ordinary least square regression which produce biased estimates. In such a situation, the need for the instrumental variable approach is supported. 14 All test results were significant (p<0.0001), indicating bias of the ordinary least square approach and justifying the instrumental variable modeling approach. Besides the previous IRF or SNF assignment, the other three instrumental variables were used as continuous measures in the instrumental variable analyses. We estimated the parameters using 2-stage least square regression. For the control outcome of 30-to 365-day mortality, the parameters were estimated from 2stage residual inclusion models because the outcome was dichotomous.  046.11, 046.19, 046.2, 046.3, 046.71, 046.72, 046.79, 046.8,  046.9, 290.0, 290.10, 290.40, 290.8, 290.9, 294.0, 294.10, 294.20,  294.8, 294.9, 330.0, 330.1, 330.2, 330.3, 330.8, 330.9, 331.0, 331.11,  331.19, 331.2, 331.6, 331.7, 331.81, 331.82, 331. An unmeasured confounder associated with mobility score change from admission to discharge and discharge destination (IRF vs SNF) by a risk ratio of 4fold each could explain away the lower confidence limit, but weaker confounding could not. An unmeasured confounder associated with death and discharge destination (IRF vs SNF) by a risk ratio of 1.6-fold each could explain away the upper confidence limit, but weaker confounding could not.