Patient functional outcomes in skilled nursing facilities: the mediating role of declining therapy

Abstract

Objective:

To estimate changes in functional outcomes in Medicare fee-for-service skilled nursing facility (SNF) patients after declines in therapy minutes that occurred after Patient-Driven Payment Model (PDPM) implementation and during the COVID-19 pandemic. We also tested the hypothesis that declines in therapy minutes mediated changes in patient functional outcomes.

Design:

Retrospective cohort study using mediation analysis. Exposures were PDPM implementation (October 2019) and COVID-19 pandemic onset (March 2020). The mediator was average minutes of therapy per day (MTD) including all physical, occupational, and speech therapy minutes during the SNF stay.

Setting:

All US SNFs.

Participants:

3,534,928 post-acute SNF stays from January 2018 through September 2021.

Interventions:

Not applicable

Main Outcome Measure(s):

Functional outcomes based on changes in activities of daily living (ADL) scores between SNF admission and discharge, measured by a validated 28-point scale: (1) ADL decline and (2) high ADL improvement (≥4 points improvement). Mediation analysis quantified effects of declining MTD on functional outcomes, adjusting for patient, facility, and community-level confounders.

Results:

Average MTD declined from 122.2 pre-PDPM to 96.5 post-PDPM and further to 87.7 during COVID-19. After PDPM implementation, the adjusted probability of ADL decline increased by 1.7 percentage points (pp) (95% CI 1.6, 1.7); during COVID-19, ADL decline increased by 3.7pp (95% CI 3.6, 3.7). Declining MTD mediated 47.9% of the increase in ADL decline post-PDPM and 26.5% during COVID-19. The adjusted probability of high ADL improvement increased by 1.3pp (95% CI 1.1, 1.3) post-PDPM and 1.0pp (95% CI 0.8, 1.0) during COVID-19. Declining MTD negatively mediated this increase in high ADL improvement by −2.9pp post-PDPM (95% CI −2.9, −2.9) and by −3.5 pp during COVID-19 (95% CI −3.6, −3.5).

Conclusions:

Declines in SNF therapy after PDPM implementation and during COVID-19 mediated substantial worsening in patient functional outcomes. Ensuring adequate therapy provision may improve functional recovery in SNF patients.

Introduction

Each year, over 14,000 skilled nursing facilities (SNFs) provide nursing and rehabilitation – physical, occupational, and speech therapy services – during more than 1.6 million post-acute SNF stays for Medicare fee-for-service (FFS) beneficiaries.¹ After an acute hospitalization, SNF care aims to stabilize and improve health status.^1,2 Accordingly, therapy in SNFs focuses on goals of improving patients’ functional independence in activities of daily living (ADLs) to facilitate a safe transition home and prevent readmissions.^2,3 Functional improvement is an important outcome of SNF care, as low levels of mobility and ADL function are associated with poor outcomes, such as higher post-discharge support needs.^2,4–7 The assumptions that higher levels of functional improvement during a SNF stay may be achieved by providing higher volumes (i.e., more minutes) of therapy, and that providing higher therapy volumes would require more resources, were reflected in the Centers for Medicare & Medicaid Services (CMS) payment system for SNFs beginning in 1998, which included higher payments for patients receiving more therapy minutes.^8–11

The assumption that more therapy would drive improvements in function during SNF stays began to be questioned by CMS and the Medicare Payment Advisory Commission,^12–14 who raised concerns that high therapy volumes were driven by financial incentives more than by patients’ needs or therapy goals.^15–18 Further, existing evidence suggested a link between higher volumes and better functional outcomes, but the quality of evidence is low.³ In response to this potential over-provision of therapy, when CMS implemented a new SNF reimbursement model, the Patient-Driven Payment Model (PDPM) in October 2019, financial incentives for high-volume therapy were removed.¹⁹ Under PDPM, SNF reimbursement rates are now calculated based on patient clinical characteristics rather than therapy volume.

After PDPM implementation, SNF therapy department staffing levels and therapy minutes provided to patients declined.^20–22 One study found that adjusted therapy volumes across all Medicare FFS SNF stays declined by nearly 30 minutes per patient-day, for a 23.7% reduction in volume post-PDPM implementation.²² When the COVID-19 pandemic began a few months later, there were further declines in therapy volumes, even when accounting for changes in SNF case-mix during the pandemic.^1,22 There has been debate about whether decreases in therapy reflect underuse or a “right sizing” of therapy volume.^20,23–26 Ultimately, the impact of reductions in therapy volume on patient functional outcomes remains unclear. One study that was limited to SNF patients admitted after hip fracture in the early months of PDPM found no changes in ADL function at SNF discharge.²³ CMS also reported no substantial differences in readmissions or the percentage of SNF stays with serious falls after PDPM implementation.²⁶ However, no studies have specifically examined the impact of declining therapy due to PDPM and/or the pandemic on functional outcomes.

In this study, we address these gaps by estimating the association of PDPM implementation and COVID-19 onset (the exposures) with changes in functional outcomes in SNFs. We then conduct a mediation analysis to test our hypothesis that the association between the exposures and outcomes is mediated by changes in therapy volume.

Methods

Data Sources and Cohort

We received approval and waiver of informed consent from the Institutional Review Board at the University of Washington. We conducted secondary analyses of all Medicare FFS post-acute SNF stays from 2018 through September 2021. We first identified hospitalizations using Medicare Provider Analysis and Review (MedPAR) hospital claims for all FFS beneficiaries for whom we could verify three months of continuous Medicare enrollment after hospital discharge using Master Beneficiary Summary Files (MBSF). We used MedPAR SNF claims to verify SNF admission within 3 days of hospital discharge, and then identified SNF stays with Minimum Data Set (MDS) 3.0 admission and discharge assessments.²⁷ MDS assessments must be completed for all FFS SNF stays at admission, at regular intervals, and at discharge, and we included only stays with a complete FFS discharge assessment, thereby excluding patients who died during the SNF stay. We removed stays with lengths of stay under 1 day or over 100 days, after which CMS defines patients as long-stay residents rather than post-acute patients.²⁸ We excluded stays with >5 hours of therapy per day, as these were likely inaccurately coded. We excluded those patients with missing (a) covariates, which are detailed below or (b) function scores on the admission or discharge assessment. We excluded stays that started before and ended after October 1, 2019 when therapy minutes documentation changed due to PDPM. Finally, we removed all stays starting in September 2019, regardless of discharge date, to avoid including only the short stays that did not span PDPM. We merged MDS data with facility-level data from CMS public data sources including Nursing Home Compare, the Payroll-Based Journal, Provider of Services files, and LTCFocus.²⁹ See Supplement Figure S1 for a cohort flowsheet.

Outcomes

Our primary measure of ADL function, extracted from the MDS section G, was a 28-point scale of combined mobility and self-care items. These items have been validated against other measures,^30–32 used frequently in prior literature,^33–35 and had high levels of completeness across study years. The scale rates patient independence with seven mobility and self-care tasks (i.e., dressing, personal hygiene, toilet use, locomotion, transfers, bed mobility, and eating). Higher scores represent greater disability, however, no minimum clinically important difference has been established.³⁶ As such, we chose to create two indicators to categorize changes in function: decline in ADL performance (ADL decline) and high improvement in ADL performance (high ADL improvement). ADL decline was defined as any decline in ADL performance as indicated by an increase in disability score during the SNF stay. As the goal of SNF care is to improve or stabilize function for post-acute patients, any worsening, regardless of magnitude, represents a negative outcome.³⁷ High ADL improvement was defined based on the top quartile of score changes in our data (a four point reduction in disability score), identifying patients with the most substantial improvement in ADL performance.

Exposures

Our exposures were the implementation of PDPM in October 2019 and the onset of the COVID-19 pandemic in March 2020. We created indicators for three periods based on the month of SNF admission: 1) before PDPM, 2) after PDPM implementation but before pandemic onset, and 3) during COVID-19.

Mediator

Our mediator was therapy volume, defined as average minutes of therapy per day (MTD) of the SNF stay for physical, occupational, and speech therapy combined. While speech therapy is provided at much lower volumes than physical and occupational therapy,²² interventions for swallowing and communication could impact the ADL score, especially because the scale includes eating items.³⁸ We included all individual, group, and concurrent therapy minutes, reflecting all therapy received, regardless of the number of patients in each session. Due to PDPM-mandated changes in MDS reporting requirements, different methods were required to obtain MTD before and after October 2019. Before PDPM, the MDS included minutes and days of therapy for a specific lookback period prior to each required assessment (i.e., admission, 14-day, 30-day, 60-day, 90-day, and discharge assessments).²² For pre-PDPM stays, we summed all therapy minutes from each required assessment and divided by the number of days of therapy summed from each individual assessments to create an average MTD across the stay. After PDPM implementation, the MDS no longer collected minutes on interim assessments and instead included new items for total minutes and days of therapy for the whole stay on the discharge assessment, which we used to calculate MTD for post-PDPM stays.

Covariates

Patient-level variables included age, sex, marital status, partial or full dual Medicare-Medicaid eligibility during the month of hospitalization, race and ethnicity, need for an interpreter, and rurality of their home Zip Code.³⁹ We included the major diagnostic category⁴⁰ of the primary diagnosis for the hospital stay, indicators for a surgery and intensive care unit utilization during hospitalization, diagnosis of Alzheimer’s Disease or related dementias, and a weighted Elixhauser comorbidity index.⁴¹ We included the 28-point ADL score from the admission assessment to ensure we measured changes for patients with similar function at SNF admission. We included indicators for special treatments received during the SNF stay (e.g., hospice care, chemotherapy, hemodialysis, ventilator) and indicators for delirium, history of recent falls, pain, visual impairment, pressure sores, incontinence, and the agitated reactive behavior scale.⁴²

SNF-level variables from public data included: five-star quality of care rating, average hours of total registered nurse, licensed practical nurse, and certified nurse assistant staffing hours, and average SNF census during the month of SNF admission. We indicated if a SNF was part of a chain, located in an urban or rural area, and the percentage of annual Medicare and Medicaid patients. Finally, we averaged daily county-level COVID-19 case rates across the exact dates of the SNF stay.⁴³

Mediation Analysis

Mediation analysis quantifies how exposures such as policies or interventions produce their effects on outcomes by estimating the role of an intermediary factor – or mediator – that sits on the causal path between exposures and outcomes.^44,45 By proposing a theoretical causal model from the exposures to the outcomes,^44,46 specific pathways can be isolated. Mediation analysis can be used to quantify the effects of any relevant pathways. Mediation analysis is generally broken down into three components, 1) the total effect which is the overall effect of the exposure regardless of the causal pathway, 2) the direct effect, which is the isolated effect of the exposure on the outcomes that does not go through any intermediary, and 3) the indirect effect, which is the effect of the exposure on the outcome through a specific mediator of interest. The total effect can be derived by adding the direct effect to all indirect effects. Figure 1 shows the directed acyclic graph with our proposed theoretical model, highlighting the direct and indirect effect causal pathways between the exposures (PDPM implementation and COVID-19) and outcomes (changes in patient function) with the potential for mediation (MTD).⁴⁴ By assessing indirect effects compared to total effects, we can estimate how much of the effects of PDPM and COVID-19 on patient function in SNFs were driven by declines in therapy minutes.

Figure 1. Directed acyclic graph depicting proposed theoretical causal relationships between study variables.

The figure shows that the total effects of the patient-driven payment model (PDPM) and COVID-19 (exposures) on patient functional outcomes can be decomposed into direct and indirect effects of each exposure. Direct effects are the effects of PDPM implementation and COVID-19 onset on functional outcomes excluding the role of potential mediators (i.e., reduced therapy minutes). The indirect effects are the effects of PDPM implementation and COVID-19 on functional outcomes that operate through the mediator. The figure also depicts how the mediation analysis accounts for confounding variables that are associated with the exposures and/or mediator and the outcomes.

We first estimated the total effects of PDPM implementation and COVID-19 onset on our two outcomes: 1) ADL decline and 2) high ADL improvement. For each outcome, estimates were obtained from a linear mixed-effects model with facility- and beneficiary-level random intercepts to account for clustering, and independent variables for time period (i.e., before PDPM, after PDPM implementation before COVID-19 onset, and during COVID-19). Models were adjusted for all pre-specified patient, facility, and community-level covariates, described above, that were hypothesized to be potential confounders. We included calendar month fixed effects to account for seasonality. We do not explicitly model the mediator (MTD) in this estimation in order to capture the total effect of the exposures.⁴⁴ We also excluded other potential mediators, such as SNF length of stay, from the analysis because conditioning on them may introduce bias.⁴⁴

Models for each outcome were used to estimate direct and indirect effects (examples in Supplement Table S1). First, we estimated the direct effect of PDPM and COVID-19 on patient function using linear mixed-effects models for each outcome with the same independent variables for potential confounders, time period, and random intercepts as our total effects models; while controlling for the mediator (MTD) to remove the indirect effect from the estimated association between exposures and the outcome. This model simultaneously allows us to estimate the effect of MTD on function outcomes. Then, we estimated the associations between each exposure (PDPM and COVID-19) and the mediator (MTD) using a linear mixed-effects model, including PDPM and COVID-19 time periods as the independent variables, and the same set of covariates and random intercepts as was used in the total effects models. We multiply the estimated effect of MTD on outcomes with the estimated effect of PDPM and COVID-19 on MTD to arrive at the indirect effect of our exposures mediated through MTD.^44,46,47 Standard errors for indirect effects were derived using the multivariate delta method, and 95% confidence intervals were constructed using normal approximation.⁴⁸ All analyses were conducted in RStudio Version 2024.12.1.

Results

Our analysis included 3,534,928 SNF stays. As seen in Figure 2, the average unadjusted therapy volume in Period 1 (before PDPM) was 122.2 MTD, which declined to 96.5 MTD in Period 2 (after PDPM implementation but before pandemic onset) and declined further to 87.7 MTD during COVID-19. In Period 1, 11.0% of SNF patients experienced ADL decline. In Period 2 this increased to 13.0% and to 14.6% in Period 3. In Period 1, 23.9% of SNF patients experienced high ADL improvement, and this increased to 25.5% in Period 2 and to 25.7% in Period 3. Refer to Supplement Table S2 for descriptive statistics for covariates.

Figure 2. Trends in functional outcomes and therapy volume.

Unadjusted monthly function outcomes for 3,534,928 post-acute skilled nursing facility (SNF) stays from January 2018 through September 2021, including the proportion of patients experiencing activities of daily living (ADL) decline (red line) and the proportion experiencing high ADL improvement (green line). Functional outcomes are reflected on the left Y-axis and the mediator of therapy volume reflected as average minutes of therapy per day (MTD) is reflected on the right Y-axis. Period 1 reflects the time prior to October 1, 2019 when the Patient-Driven Payment Model (PDPM) was implemented. Period 2 includes the time period after PDPM was implemented but prior to the onset of the COVID-19 pandemic in March 2020. Period 3 reflects the COVID-19 pandemic.

In adjusted models, the total effect of PDPM on the probability of ADL decline was an increase of 1.7 percentage points (pp) (95% CI 1.6, 1.7). The total effect of COVID-19 on the probability of ADL decline was an increase of 3.7pp (95% CI 3.6, 3.7) (Table 1). Our mediation analysis showed that the indirect effect of PDPM on ADL decline, mediated through MTD, was a 0.8pp increase (95% CI 0.8, 0.8) and after COVID-19 the indirect effect was a 1.0pp increase (95% CI 0.9, 1.0). Therefore, the decline in MTD associated with PDPM accounted for 47.9% of the increased probability of ADL decline after PDPM implementation and 26.5% of the increased probability of decline during COVID-19.

Table 1.

Total effects of Patient Driven Payment Model (PDPM) implementation and COVID-19 onset on functional outcomes with mediation analysis estimates of indirect effects of declining minutes of therapy per day (MTD).

	Total Effects (95% CI)	Indirect Effects (95% CI) of Therapy MTD
Percentage Point Change in Probability of ADL Decline *
After PDPM Implementation	1.6 (1.5, 1.7)	0.8 (0.8, 0.8)
After COVID-19 Onset	3.7 (3.6, 3.7)	1.0 (0.9, 1.0)
Percentage Point Change in Probability of High ADL Improvement†,
After PDPM Implementation	1.2 (1.0, 1.2)	−3.0 (−3.0, −2.9)
After COVID-19 Onset	0.8 (0.7, 0.8)	−3.6 (−3.7, −3.6)

Notes: Results are adjusted estimates for the total effects of the implementation of PDPM and the onset of the COVID-19 pandemic on functional outcomes. Indirect effects are results of the mediation analysis, estimating the indirect effects of changes in total minutes of therapy per day (MTD) on the total effects for 3,534,928 skilled nursing facility (SNF) stays from January 2018 through September 2021.

Abbreviations: ADL – Activities of Daily Living, pp – percentage point.

^*ADL decline occurred with any increase in disability on a 28-point scale of mobility and self-care function between SNF admission and discharge.

^†High ADL improvement occurred with a ≥4-point reduction in disability on a 28-point scale of mobility and self-care function between SNF admission and discharge.

The total effect of PDPM on the probability of high ADL improvement was an increase of 1.3pp (95% CI 1.1, 1.3) and the total effect of COVID-19 was a 1.0pp increase (95% CI 0.8, 1.0). Conversely, the mediation analysis indicated that the decline in MTD associated with PDPM implementation had a negative indirect effect of −2.9pp (95% CI −2.9, −2.9) on the probability of high ADL improvement. In other words, the decline in MTD induced by PDPM was associated with a 2.9pp decrease in the probability of high ADL improvement. The decline in MTD also had a negative indirect effect on the probability of high ADL improvement after COVID-19 (−3.5pp, 95% CI −3.5, −3.6).

Discussion

The results of our total effects models demonstrate that the adjusted proportion of patients who experienced ADL decline during post-acute SNF stays increased after PDPM implementation and increased more during the COVID-19 pandemic, for a total change of approximately 4pp. Conversely, the proportion of patients who achieved high ADL improvement during this time was approximately 2pp higher. Our mediation analysis detected a statistically significant indirect effect of therapy MTD on these changes in functional outcomes, suggesting a mediating effect. Specifically, declining therapy volumes contributed to worse function for SNF patients no matter how we specified functional change (i.e., lower therapy volumes were associated with higher rates of functional declines, in addition to lower rates of high improvement in function). This novel finding provides rigorous evidence to build on prior literature suggesting links between high-volume therapy and positive functional outcomes published before PDPM and the pandemic.³ The consistent positive association between higher therapy volumes and improved functional outcomes underscores the importance of adequate therapy volume for achieving goals of post-acute SNF care.

The diverging pattern of more patients declining in function over time, while more patients also got substantially better, is a novel finding that was robust to adjustments for seasonality and differences in local COVID-19 impacts, SNF characteristics (e.g., nurse staffing levels, profit status), length of stay, and patient clinical factors. Notably, increases in the rates of ADL decline were larger relative to the increases in high ADL improvement. This worsening began prior to the pandemic, so cannot be fully attributed to potential case-mix changes that occurred during COVID-19.

Diverging patterns of rising rates of functional decline and high improvement have many potential explanations, including changes in how therapy services were deployed as SNFs responded to PDPM incentives and grappled with pandemic-related staffing issues.^20,49,50 After PDPM implementation, SNFs reduced therapy volumes, but also delivered more group and concurrent therapy rather than individual sessions, which saved staffing costs because one therapist could treat multiple patients.^20,23,50 Research suggests that individual sessions may be more appropriate for patients with more severe functional and cognitive impairments, who may also require higher volumes of therapy to meet discharge goals.^3,51–54 It is thus possible that the shift toward group and concurrent sessions was more beneficial for patients who experienced high ADL improvement.⁵³ However, increasing proportions of patients experiencing functional decline is a troubling trend, which suggests that a substantial portion of the PDPM-associated reduction in therapy volume does not reflect a ‘right-sizing’ of therapy provision in SNFs.²² This is especially important considering SNFs will soon become financially accountable for functional outcomes when patient function scores are included in SNF value-based purchasing program metrics starting in 2027.²⁶ Because goals of post-acute SNF stays are to improve – or at least maintain – function, determining which patient and facility characteristics predict low therapy volumes will be essential for determining when therapy provision should be increased to avoid unnecessary functional declines.

Limitations

While we adjusted for many known confounders and used a mediator that temporally precedes the outcomes, unmeasured confounding could still lead to bias, and findings should not be interpreted causally. For example, we could not account for factors such as patients’ willingness to participate in therapy, delivery of specific interventions, or expertise of the therapy provider. Indeed, declines in therapy volume did not mediate 100% of the changes in function during our study, suggesting unmeasured differences in patient characteristics and/or therapy practice changes that were relevant to functional outcomes. It is also possible that while therapy volume declined, the efficacy of therapy sessions improved, which may be especially salient to the slight increase we observed in high ADL improvement. Therapy practice improvements could be related to recent evidence that emphasizes the effectiveness of high-intensity physiologically challenging treatments with progressive levels of difficulty.^55–61 Additionally, because we aimed to isolate the mediation pathway through MTD, we could not condition on other potential mediators, including changes in SNF lengths of stay, which may also contribute to the fact that declining MTD did not mediate 100% of the changes in function.

Other limitations include using a slightly different method of creating the MTD variable before and after PDPM implementation. For pre-PDPM stays, this would have impacted longer stays with fewer days of therapy captured during interim assessment lookback periods. However, the average length of stay in our pre-PDPM dataset was 24.8 days (Supplement Table S2), so we were able to capture most days of therapy within the lookback periods for the admission and discharge assessments as well as the 7-day lookback periods for the 14-day and 30-day assessments. Additionally, we lost a relatively large number of observations due to missing data on facility-level confounders from CMS public files (see Supplement Figure S1). Public data files often exclude smaller SNFs due to censoring of small sample sizes or differences in reporting requirements, so results may not be generalizable to all smaller SNFs. Finally, the 28-point ADL scale may not adequately capture improvement or decline across all tasks relevant for ensuring a patient’s successful discharge or quality of life. Future work with newer data incorporating MDS section GG measures that include more functional domains would be valuable in exploring specific aspects of function that may be more sensitive to changes in MTD. Newer data would also allow for examination of relationships between therapy provision and outcomes after the public health emergency and maturation of SNF operation under PDPM. Finally, future studies examining whether declining MTD had different mediating effects on functional outcomes for specific patient groups (e.g., patients with different diagnoses, goals of care, or functional levels at admission) would also be valuable for helping prioritize potential increases in therapy volumes within SNFs.

Conclusions

Our mediation analysis found that declining therapy volumes in SNFs from January 2018 through September 2021 contributed to a substantial proportion of worsening functional outcomes. Results suggest that fewer FFS patients would have experienced functional decline – and more patients may have achieved high improvement – if therapy volumes had not declined to such an extent after PDPM implementation and the COVID-19 pandemic. Results highlight the important role of therapy for post-acute SNF outcomes of stabilizing and improving ADLs and suggest that increasing therapy volumes could be an effective strategy for SNFs hoping to improve patient functional outcomes.

Abbreviations

SNF

skilled nursing facility

FFS

fee for service

ADL

activity of daily living

PDPM

Patient-Driven Payment Model

CMS

Centers for Medicare & Medicaid Services

MedPAR

Medicare Provider Analysis and Review

MBSF

Master Beneficiary Summary File

MDS

Minimum Data Set

MTD

minutes of therapy per day