The use of clinical decision support for hospital discharge disposition was associated with a reduction in spending and readmissions without negatively affecting emergency department use.
Objectives: To assess the association of a clinical decision support (CDS) algorithm for hospital discharge disposition with spending, readmissions, and postdischarge emergency department (ED) use.
Study Design: A retrospective study in a cohort of fee-for-service Medicare patients 65 years or older linked to a database of patients receiving CDS.
Methods: We evaluated (1) patients whose discharge disposition was concordant with the CDS recommendation versus those whose disposition was not and (2) patients receiving CDS for discharge disposition versus those not receiving CDS, regardless of concordance. Outcomes were spending over a 90-day episode, 90-day readmissions, and postdischarge ED utilization not associated with a readmission.
Results: Analysis of concordant versus discordant cases showed decreased spending for concordant cases ($860 savings; 95% CI, $162-$1558; P = .016), a decrease in readmissions (adjusted odds ratio [OR], 0.920; 95% CI, 0.850-0.995; P = .038), and no change in rate of postdischarge ED use (adjusted OR, 0.990; 95% CI, 0.882-1.110; P = .858). Analysis of patients receiving CDS versus not receiving CDS showed no significant difference in spending ($221 savings; 95% CI, —$115 to $557; P = .198), ED use (adjusted OR, 0.959; 95% CI, 0.908-1.012; P = .128), or readmission rate (adjusted OR, 1.004; 95% CI, 0.966-1.043; P = .840).
Conclusions: Following the recommendation of a CDS algorithm for hospital discharge disposition was associated with lower spending, fewer readmissions, and no change in ED use over a 90-day episode of care.
Am J Manag Care. 2019;25(6):288-294Takeaway Points
Following the recommendation of a clinical decision support (CDS) algorithm for hospital discharge disposition was associated with lower spending and reduced readmissions with no change in emergency department (ED) use.
Spending on postacute care accounts for a substantial portion of overall healthcare costs and is growing faster than other spending categories.1,2 For conditions like pneumonia, chronic obstructive pulmonary disease, heart failure, and joint replacement, Medicare spends nearly as much in the 30 days after discharge as it does during hospitalization.3 Moreover, postacute costs are associated with large geographic variation across the United States, with three-fourths of all regional variation in Medicare spending attributable to postacute care spending.4
Achieving the judicious, appropriate use of resources following hospitalization can be a critical success factor for organizations participating in full risk capitation and in payment models such as accountable care organizations and bundled payments. Additionally, Medicare spending per beneficiary—a measure of spending encompassing the 30 days subsequent to hospitalization—is publicly reported and tied to financial incentives for all US Inpatient Prospective Payment System hospitals as part of Hospital Value-Based Purchasing and is a component of the physician payment formula under the Merit-based Incentive Payment System.5
Despite growing pressure for hospitals to develop a systematic approach to using postacute care for all patients, most published studies examining such approaches address a single diagnosis, such as stroke or joint replacement; describe a method that may be too complex and time-consuming for widespread use; or rely on information that is unavailable early in the hospitalization, when factors influencing discharge destination may be modified.6-12
We sought to build a clinical decision support (CDS) algorithm to assist hospital discharge planning teams in identifying the most appropriate discharge care level while avoiding untoward effects such as increases in readmissions, emergency department (ED) use, and overall spending. To assess the algorithm, as a convener in CMS’ Bundled Payments for Care Improvement initiative (BPCI),13 we accessed Medicare claims data for acute hospitalizations and the subsequent 90-day period.
We evaluated the effect of the algorithm on spending, 90-day readmissions, and postdischarge 90-day ED use in cases in which discharge disposition was concordant or discordant with the algorithm’s recommendation and in which the algorithm was or was not used (regardless of concordance).
The setting of this study was acute hospitalization and the 90-day period following discharge, encompassing patient care in the home, home with a home health agency (hereafter written simply as home health agency), and postacute facility (including skilled nursing facilities, inpatient rehabilitation facilities, and long-term acute care hospitals).
We developed a proprietary CDS tool incorporating an algorithm to help teams determine an appropriate level of care following hospital discharge. Inputs to the algorithm were ambulatory status; ability to perform activities of daily living; cognitive status; availability of a capable caregiver; postacute physical, occupational, and speech therapy needs; and postacute skilled nursing needs. In developing the CDS tool, we reviewed the literature to identify patient-level factors to serve as the basis of decision support.14-19 Caregiver support was also recognized as critical to effective discharge planning.20-24 We convened experts in home health, postacute facility care, and hospital care in a series of working sessions to finalize factors driving a decision to discharge patients to 1 of 3 options. Other potentially informative factors, including comorbidities, polypharmacy, environmental factors (eg, stairs, home modifications), and social determinants, were intentionally omitted to increase the tool’s ease of use. After assessing the tool for user requirements and reproducibility of results across users, we created a scoring system based on the identified inputs that yielded a recommendation for 1 of 3 postdischarge care intensity levels: home, home health agency, or postacute facility.
A pilot was then conducted, using an analysis of 1537 BPCI patients to whom the CDS tool was retrospectively applied. This pilot yielded a proposal that made recommendations as follows (results shown as pilot vs controls): More patients go home or to a home health agency (home, 50.6% [95% CI, 47.7%-53.6%] vs 36.1% [95% CI, 32.8%-39.5%]; home health agency, 32.5% [95% CI, 29.1%-35.1%] vs 26.0% [95% CI, 16.8%-24.4%]), and fewer go to a postacute facility (16.9% [95% CI, 13.1%-20.8%] vs 43.3% [95% CI, 40.2%-46.5%]). The tool’s performance was then evaluated using risk-adjusted regression models created to predict the rate of 90-day readmissions for patients discharged to home, a home health agency, and a postacute facility. These rates were compared with observed 90-day readmissions and found to be not statistically different across the 3 discharge dispositions: home (32.4% [95% CI, 30.4%-34.4%] vs 32.8% [95% CI, 29.5%-36.0%]), home health agency (33.1% [95% CI, 31.1%-35.1%] vs 32.0% [95% CI, 27.5%-36.4%]), and postacute facility (33.3% [95% CI, 31.4%-35.3%] vs 31.7% [95% CI, 28.7%-34.7%]).25
Subsequently, the algorithm was embedded in the convener’s proprietary software platform used by bundled payment program operating personnel to identify and manage patients during the bundle episode. These users—personnel at the convener’s BPCI episode—initiating hospitals—were trained in the use of the tool, which involved a webinar, a user’s manual, and targeted face-to-face and remote individualized education and support. Training addressed how and when to populate the components of the algorithm and also the importance of creating a process to discuss the tool’s inputs and recommendation with the interdisciplinary discharge team, including the patient/surrogate decision maker. Although all episode-initiating hospitals were encouraged to use the tool, only a subset did so, which was at the discretion of the organizations’ leadership.
All Part A and Part B claims for the convener’s population of fee-for-service Medicare BPCI patients encompassing acute hospitalization and the subsequent 90-day period occurring between January 1, 2016, and March 31, 2017, were accessed. After deleting claims with incomplete information, there were 148,385 episodes available for analysis. During the same time period, a subset of 15,887 patients in this population were tested using the CDS tool. Our analytic data set consists of Medicare claims data on 132,498 episodes that did not receive the CDS tool and Medicare claims data plus CDS testing results for the 15,887 episodes that did receive the CDS tool.
For each episode, we examined the following outcomes: allowed payment amounts (spending), discharge disposition (home, home health agency, or postacute facility), readmission within 90 days, and postdischarge ED visits not associated with a readmission within 90 days of hospital discharge.
We used propensity modeling with inverse probability weighting (IPW) for each analysis26,27; propensity to follow the CDS recommendation was used in the concordant versus discordant analysis, and propensity to use the test was used in the intention-to-treat (ITT) analysis (tested vs untested). We used logistic regression for both propensity models. The following variables were included as independent variables: length of stay, race, sex, dual-enrolled status (binary), prior system utilization (the number of days spent as a hospital inpatient in the year prior to the index hospitalization), entered through the ED, clinical episode name, Medicare Severity Diagnosis Related Group tier (no comorbid complication, comorbid complication, major comorbid complication), and primary diagnosis code. In all analyses, we used IPW to compute average treatment effect among the treated (ATT) estimates. Although these variables were all selected a priori without first examining their relationship to the probability of testing, most of them are significantly associated with it.
We performed 2 analyses of the effect of the CDS tool’s use on patient outcomes: concordant versus discordant and tested versus untested.
Concordant Versus Discordant
Concordant was defined as discharge disposition agreeing with the CDS tool’s recommendation, whereas discordant was defined as discharge disposition disagreeing with the recommendation. We used IPW regression (spending) and logistic regression (ED use and readmissions) to test for the effect of agreeing with the CDS tool recommendation on outcomes.
Within this analysis, we also examined the impact of more intense and less intense levels of care. In an attempt to control for hospital-level differences, we performed a post hoc analysis of CDS-concordant versus CDS-discordant cases, showing results for when the hospital is controlled for as a main effect.
A more intense level of care is defined as either (1) the CDS tool recommends home and the actual disposition was either home health agency or postacute facility or (2) the CDS tool recommends home health agency and the actual disposition was postacute facility. A less intense level of care is defined as either (1) the CDS tool recommends postacute facility and the actual disposition was either home health agency or home or (2) the CDS recommends home health agency and the actual disposition was home. In these models we included the impact of more or less intense levels of care on the outcomes.
Tested Versus Untested
The second analysis consisted of tested versus untested cases. Tested cases received the CDS tool, whereas untested cases did not receive the CDS tool. This was an ITT analysis, examining outcomes independent of whether the CDS tool recommendation was followed. This analysis reflects the expectation that the decision of discharge disposition is based on a merging of clinical expertise and the CDS tool recommendation, with responsibility for the final decision resting with the discharge planning team and patient/caregiver.
We acknowledge that in actual use, providers may not apply the CDS tool to all patients, nor do we expect that everyone in the population will receive the test. Therefore, we used a propensity model to estimate the ATT, or the effect of the test on the outcome of those who received it.
Federal common rule28 provides an exemption from institutional review board requirements when the purpose is to study, evaluate, or otherwise examine a public benefit or a service program—in this case, CMS’ BPCI program. The contractor signed a data use agreement stating that all data were securely and solely used for the purposes of this study.
In general, compared with patients found to be discordant, those who were concordant were about 1 year younger, had shorter lengths of stay, were less likely to be women, were less likely to be dual enrolled, and had lower rates of major complications. Concordant patients were also less likely to be sent to home health agencies and postacute facilities and were more likely to go home (Table 1). Adjustment for differences in concordant versus discordant patients using IPW resulted in the 2 groups being similar in all categories (eAppendix Table 1A [eAppendix available at ajmc.com]).
There were 148,385 patients in the sample in total. Of the 15,887 who received CDS, 10,218 were CDS concordant and 5669 were CDS discordant. Of the latter group, 2066 were discharged to a less intense level of care than what the CDS proposed and 3603 were discharged to a more intense level of care (Figure 1).
Episode spending was $860 less (95% CI, $162-$1558; P = .016), 90-day readmissions were lower (adjusted odds ratio [OR], 0.920; 95% CI, 0.850-0.995; P = .038), and ED use was unchanged (adjusted OR, 0.990; 95% CI, 0.882-1.110; P = .858) for concordant compared with discordant cases (Figure 2). A post hoc analysis, controlling for the hospital as a main effect, showed episode spending to be $934 less (95% CI, $247-$1621; P = .008), 90-day readmissions lower (adjusted OR, 0.916; 95% CI, 0.846-0.992; P = .031), and ED use unchanged (adjusted OR, 0.997; 95% CI, 0.887-1.120; P = .957) for concordant versus discordant cases.
More Intense and Less Intense Levels of Care
When concordant cases were compared with discordant cases discharged to more intense levels of care, concordance was associated with decreased spending ($4802; 95% CI, $3896-$5709; P <.001), decreased readmission rates (adjusted OR, 0.834; 95% CI, 0.757-0.920; P <.001), and unchanged ED use (adjusted OR, 0.956; 95% CI, 0.832-1.096; P = .517). Results of the adjusted analysis of concordant versus discordant cases discharged to less intense care suggest that concordance was more expensive ($6417; 95% CI, $5551-$7283; P <.001), with no changes in readmission rates (adjusted OR, 1.086; 95% CI, 0.951-1.240; P = .222) or ED use (adjusted OR, 1.086; 95% CI, 0.893-1.319; P = .410) (Figure 3).
Disposition to Home, Home Health Agency, and Postacute Facility
Recommended rates for disposition to home, home health agency, and postacute facility were 41.5%, 29.6%, and 28.9%, respectively, whereas actual disposition rates among the tested population were 40.8%, 20.6%, and 38.7% (Table 2). Compared with actual, the CDS tool recommended fewer patients be sent to a postacute facility and more patients be sent to a home health agency than was observed. Approximately the same proportion of patients were sent home (40.8%) as were recommended (41.5%). Conversely, more patients received a recommendation to go to a home health agency (29.6%) than was observed (20.6%), and fewer received a postacute facility recommendation (28.9%) than was observed (38.7%). The overall rate of concordance with the CDS recommendation was 64%. Table 2 shows the numbers of patients for each combination of disposition and recommendation.
Tested Versus Untested Populations and Outcomes
Overall, there was no difference in age between the tested and untested populations. However, the tested were less likely to be white, were more likely to be female, had longer lengths of stay, and had higher rates of dual enrollment. See eAppendix Table 1B for the rate of testing split across a selection of demographic variables.
The IPW-adjusted ITT analysis showed decreased spending among those tested, but it was not statistically significant ($221 savings; 95% CI, —$115 to $557; P = .198). There was also no difference among those tested in ED use (adjusted OR, 0.959; 95% CI, 0.908-1.012; P = .128) or readmission rates (adjusted OR, 1.004; 95% CI, 0.966-1.043; P = .840) (see the eAppendix Figure).
We tested a CDS tool for post—hospital discharge destination and found that following its recommendation was associated with reduced spending and readmissions, with no change in ED use, over the course of an episode encompassing a hospitalization and the ensuing 90-day postdischarge period for patients participating in the Medicare bundled payment program. For cases that were discharged to a more intense care level than the recommendation, associated spending and readmissions were greater, whereas ED use was unchanged. For cases discharged to less intense care than recommended, spending was reduced, whereas readmissions and ED use were unchanged.
Because of the observational nature of the study, it is difficult to definitively state the reasons for lower spending or readmission reductions when following the tool’s recommendation. It is possible that the recommendation-concordant group—who went home more often and to postacute facilities less often (Table 1)—recovered more successfully because they were in the home environment. The hazards of postacute facilities (eg, falls, delirium, infection, poor nutrition, decreased mobility, deconditioning29,30) may play a role in increased readmissions and ED use, although the evidence on the impact of home care versus alternative locations on health outcomes is inconclusive.31 Alternatively, in select patients, such as those undergoing elective total joint replacement, a home discharge has been associated with lower readmissions.32
Similarly, the reasons for findings associated with less intense and more intense discharge decisions are speculative. It is conceivable that for many patients, receiving less intense posthospital care than the tool recommends is in actuality the appropriate care level, thus explaining why spending is lower and readmissions and ED use were no different from those of the recommendation-concordant group. For patients receiving more intense care, which may in part be driven by patient/caregiver preferences, one may argue that higher spending and readmissions are explained by patient factors; however, the propensity model was designed to adjust for such factors.
For patients discharged to a less intense level of care than recommended, because readmissions and ED use were unchanged in that group, it is likely that if the discharge team’s judgment supports the decision to discharge to a less intense level, such a decision is safe and appropriate. Conversely, if a more intense level of care is felt to be required, the team should consider this study’s findings of higher readmissions and ED use and carefully consider the purported benefits of the decision.
The results of the study also showed that in the comparison of cases receiving CDS tool testing—regardless of whether the recommendation was followed—versus no testing, there was no statistically significant change in spending and no change in readmission rate or ED use. It is likely that providers were selective about who received the tool, namely that those tested had longer hospital lengths of stay and higher rates of dual enrollment, suggesting they may be sicker and more likely in need of additional care after discharge. This is relevant because even with the propensity model, it is difficult to adjust for all the differences between the tested and untested, yet the potentially sicker tested group did not have worse outcomes.
It should be stated that the intention of the CDS tool is to inform the discharge planning team’s evaluation regarding the factors influencing discharge destination, with a final decision arrived at with input from the patient/caregiver and the judgment of the team. Discussion and evaluation by the team of the tool’s data elements—including measures of independence, availability of a capable caregiver, and postacute needs—along with other details of each particular case can form the basis of a structured process yielding a final decision. The adoption of such a process for evaluating patients’ postdischarge destination may help hospitals looking to improve the precision with which various postacute services and settings are recommended.
The study was limited to the use of observational data. Thus, if there are unmeasured confounders associated with the decision to follow the CDS tool’s recommendation, or to use the tool at all, there may be uncorrected bias in the results. Because providers may exercise discretion as to who receives the CDS tool, selection bias may be a significant factor in differences between the tested and untested groups. Also, the study used only Medicare Part A and Part B claims data in its outcomes analysis. It is possible that clinical data would have improved the propensity model and increased the relevance of the outcomes. It is also possible that there were differences in spending not reflected in Part A and Part B claims, such as out-of-pocket spending or that associated with supplemental insurance, that were not measured.
The population in the study was limited to patients 65 years or older. We cannot rule out the possibility that the impact of the test will differ among younger patients. Moreover, the analysis of recommendation concordance is confounded by the fact that providers’ discharge decisions are potentially affected by the tool’s recommendations.
This study demonstrated an association between concordance with a CDS algorithm and decreased 90-day episode spending and readmissions, with no adverse effect on postdischarge ED visits. The study is an example of an innovative approach to care redesign under a bundled payment model. Because bundled payments create an incentive to critically evaluate decisions affecting discharge destination, the development and implementation of the CDS tool can be viewed as a result of a new payment incentive.
The authors acknowledge the valuable contributions of Georgine Schmidt, RN; Susana Hall, RN, MBA; Benjamin Record, BA; Luann Tammany, PT, MBA; and Steve Wiggins, MBA, to the manuscript. All are affiliated with Remedy Partners.Author Affiliations: Remedy Partners (WFW, JLC, PH), Norwalk, CT; Village Care of New York (RT), New York, NY; Vital Statistics, LLC (JEL), Chapel Hill, NC.
Source of Funding: None.
Author Disclosures: Drs Whitcomb and Hayward report current employment with and stock options in Remedy Partners. Ms Chiu reports previous employment with Remedy Partners. Ms Tornheim reports previous employment with and stock options in Remedy Partners. Dr Lucas reports receiving payment from Remedy Partners for involvement in the preparation of this manuscript.
Authorship Information: Concept and design (WFW, RT, PH); acquisition of data (WFW, JLC, PH); analysis and interpretation of data (WFW, RT, JEL, JLC, PH); drafting of the manuscript (WFW, JEL); critical revision of the manuscript for important intellectual content (WFW, JEL, PH); statistical analysis (JEL, PH); administrative, technical, or logistic support (WFW, RT, JLC, PH); and supervision (PH).
Address Correspondence to: Winthrop F. Whitcomb, MD, Remedy Partners, 800 Connecticut Ave, 3rd Floor, Norwalk, CT 06854. Email: email@example.com.REFERENCES
1. Huckfeldt PJ, Mehrotra A, Hussey PS. The relative importance of post-acute care and readmissions for post-discharge spending. Health Serv Res. 2016;51(5):1919-1938. doi: 10.1111/1475-6773.12448.
2. Chandra A, Dalton MA, Holmes J. Large increases in spending on postacute care in Medicare point to the potential for cost savings in these settings. Health Aff (Millwood). 2013;32(5):864-872. doi: 10.1377/hlthaff.2012.1262.
3. Mechanic R. Post-acute care—the next frontier for controlling Medicare spending. N Engl J Med. 2014;370(8):692-694. doi: 10.1056/NEJMp1315607.
4. Newhouse JP, Garber AM. Geographic variation in Medicare services. N Engl J Med. 2013;368(16):1465-1468. doi: 10.1056/NEJMp1302981.
5. Cost requirements: PY2019. CMS Quality Payment Program website. qpp.cms.gov/mips/cost. Accessed May 3, 2018.
6. Unsworth CA. Selection for rehabilitation: acute care discharge patterns for stroke and orthopedic patients. Int J Rehabil Res. 2001;24(2):103-114.
7. Lockery SA, Dunkle RE, Kart CS, Coulton CJ. Factors contributing to the early rehospitalization of elderly people. Health Soc Work. 1994;19(3):182-191.
8. Bohannon RW, Lee N, Maljanian R. Postadmission function best predicts acute hospital outcomes after stroke. Am J Phys Med Rehabil. 2002;81(10):726-730. doi: 10.1097/01.PHM.0000027422.46697.A7.
9. Ekstrand E, Ringsberg KA, Pessah-Rasmussen H. The Physiotherapy Clinical Outcome Variables Scale predicts length of hospital stay, discharge destination and future home facility in the acute comprehensive stroke unit. J Rehabil Med. 2008;40(7):524-528. doi: 10.2340/16501977-0210.
10. van der Zwaluw CS, Valentijn SAM, Nieuwenhuis-Mark R, Rasquin SM, van Heugten CM. Cognitive functioning in the acute phase poststroke: a predictor of discharge destination? J Stroke Cerebrovasc Dis. 2011;20(6):549-555. doi: 10.1016/j.jstrokecerebrovasdis.2010.03.009.
11. Gambier N, Simoneau G, Bihry N, et al. Efficacy of early clinical evaluation in predicting direct home discharge of elderly patients after hospitalization in internal medicine. South Med J. 2012;105(2):63-67. doi: 10.1097/SMJ.0b013e318242d74d.
12. Jette DU, Stilphen M, Ranganathan VK, Passek SD, Frost FS, Jette AM. AM-PAC “6-Clicks” functional assessment scores predict acute care hospital discharge destination. Phys Ther. 2014;94(9):1252-1261. doi: 10.2522/ptj.20130359.
13. Bundled Payments for Care Improvement (BPCI) initiative: general information. CMS website. innovation.cms.gov/initiatives/bundled-payments. Updated April 17, 2019. Accessed April 29, 2018.
14. Graf C. Functional decline in hospitalized older adults. Am J Nurs. 2006;106(1):58-67.
15. Hartigan I. A comparative review of the Katz ADL and the Barthel Index in assessing the activities of daily living of older people. Int J Older People Nurs. 2007;2(3):204-212. doi: 10.1111/j.1748-3743.2007.00074.x.
16. Katz S. Assessing self-maintenance: activities of daily living, mobility, and instrumental activities of daily living. J Am Geriatr Soc. 1983;31(12):721-727. doi: 10.1111/j.1532-5415.1983.tb03391.x.
17. Katz S, Down TD, Cash HR, Grotz RC. Progress in the development of the index of ADL. Gerontologist. 1970;10(1):20-30.
18. Katz S, Ford AB, Moskowitz RW, Jackson BA, Jaffe MW. Studies of illness in the aged. The index of ADL: a standardized measure of biological and psychosocial function. JAMA. 1963;185(12):914-919. doi: 10.1001/jama.1963.03060120024016.
19. Kresevic DM. Assessment of physical function. In: Boltz M, Capezuti E, Fulmer T, Zwicker D, eds. Evidence-Based Geriatric Nursing Protocols for Best Practice. 4th ed. New York, NY: Springer Publishing Company, LLC; 2012:89-103.
20. Kane RL. Finding the right level of posthospital care: “we didn’t realize there was any other option for him.” JAMA. 2011;305(3):284-293. doi: 10.1001/jama.2010.2015.
21. D’Souza MF, Davagnino J, Hastings S, Sloane R, Kamholz B, Twersky J. Preliminary data from the Caring for Older Adults and Caregivers at Home (COACH) program: a care coordination program for home-based dementia care and caregiver support in a Veterans Affairs Medical Center. J Am Geriatr Soc. 2015;63(6):1203-1208. doi: 10.1111/jgs.13448.
22. Black BS, Johnston D, Rabins PV, Morrison A, Lyketsos C, Samus QM. Unmet needs of community-residing persons with dementia and their informal caregivers: findings from the maximizing independence at home study. J Am Geriatr Soc. 2013;61(12):2087-2095.
23. Discharge planning proposed rule focuses on patient preferences [news release]. Baltimore, MD: CMS; October 29, 2015. cms.gov/newsroom/press-releases/discharge-planning-proposed-rule-focuses-patient-preferences. Accessed May 3, 2019.
24. Medicare benefit policy manual: chapter 7—home health services [40: covered services under a qualifying home health plan of care]. CMS website. cms.gov/Regulations-and-Guidance/Guidance/Manuals/downloads/bp102c07.pdf. Updated March 22, 2019. Accessed May 3, 2019.
25. Care at the Right Location (CARL). Remedy Partners website. remedypartners.com/our-platforms/episode-connect/care-at-the-right-location-carl. Accessed April 29, 2018.
26. Curtis LH, Hammill BG, Eisenstein EL, Kramer JM, Anstrom KJ. Using inverse probability-weighted estimators in comparative effectiveness analyses with observational databases. Med Care. 2007;45(10 suppl 2):S103-S107. doi: 10.1097/MLR.0b013e31806518ac.
27. Austin PC, Stuart EA. Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observation studies. Stat Med. 2015;34(28):3661-3679. doi: 10.1002/sim.6607.
28. Protection of human subjects, subpart A: basic HHS policy for protection of human research subjects: to what does this policy apply? 46 CFR §46.101(b)(5) (2009). govinfo.gov/content/pkg/CFR-2016-title45-vol1/pdf/CFR-2016-title45-vol1-part46.pdf. Accessed May 3, 2019.
29. Hakkarainen TW, Arbabi S, Willis MM, Davidson GH, Flum DR. Outcomes of patients discharged to skilled nursing facilities after acute care hospitalizations. Ann Surg. 2016;263(2):280-285. doi: 10.1097/SLA.0000000000001367.
30. Wysocki A, Butler M, Kane RL, Kane RA, Shippee T, Sainfort F. Long-term services for older adults: a review of home and community-based services versus institutional care. J Aging Soc Policy. 2015;27(3):255-279. doi: 10.1080/08959420.2015.1024545.
31. Boland L, Légaré F, Perez MM, et al. Impact of home care versus alternative locations of care on elder health outcomes: an overview of systematic reviews. BMC Geriatr. 2017;17(1):20. doi: 10.1186/s12877-016-0395-y.
32. Dummit LA, Kahvecioglu D, Marrufo G, et al. Association between hospital participation in a Medicare bundled payment initiative and payments and quality outcomes for lower extremity joint replacement episodes. JAMA. 2016;316(12):1267-1278. doi: 10.1001/jama.2016.12717.