| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||
Correspondence: Address correspondence to Mary L. James, MA, Institute of Gerontology, University of Michigan, 300 North Ingalls, Ann Arbor, MI 48109-2007. E-mail: mljames{at}umich.edu
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
Key Words: Long-term care • Nursing facility • Transition • Community-based care • Screening • Targeting
A significant literature is available describing the program structure and operations among these first-generation NFT programs (Eiken, Holtz, & Steigman, 2005; Kasper, 2005; Kasper & O'Malley, 2006a, 2006b; Mollica, 2003; Mollica & Gillespie, 2004; O'Connor, Long, Quach, Burgess, & Shea-Delaney, 2006; Reinhard & Farnham, 2006; Reinhard & Gillespie, 2005; Reinhard & Hendrickson, 2006; Reinhard, Hendrickson, & Bemis, 2005; Reinhard & Scala, 2001; Siebenaler, O'Keeffe, Brown, & O'Keeffe, 2005; Summer, 2005). A major theme in this literature illustrates the strategies used to identify individuals likely to participate or those individuals deemed most appropriate for transition services. Such targeting strategies have varied widely. For instance, Michigan sought to identify anyone residing in a nursing facility who expressed a wish to return home (Eiken, Burwell, & Asciutto, 2002). Several states, including Vermont, Pennsylvania, and Kansas, tried to use data on the functional capacities of institutionalized older adults from the nursing home Minimum Data Set (MDS) 2.0 to identify candidates for transition (Chapin, Wilkinson, Rachlin, & Levy, 1997; Eiken, Hatzmann, & Asciutto, 2003; Eiken & Heestand, 2003; Reinhard & Hendrickson, 2006). Advocates have supported the use of MDS Item Q1a ("Resident expresses/indicates preference to return to the community") to identify likely transition candidates.
Although targeting strategies appear to be of great interest, only a few researchers have systematically analyzed the characteristics of individuals served by transition programs (Bolda, Fralich, Keith, Leighton, & Richards, 1998). This evidence gap hinders efforts to evaluate the outcome of different targeting strategies, to define what constitutes a successful transition, and to estimate future program costs. The following case study describes one analytic approach that begins to flesh out these key issues. It is important to note that the method uses a data set all states have in common: the MDS 2.0.
In 2001, Arkansas initiated the Passages program through the Division of Aging and Adult Services to transition eligible individuals from nursing facilities to the community. The state's eight Area Agencies on Aging and four Centers for Independent Living provided services. To be eligible, individuals had to meet five criteria: the individuals' medical needs could be met in the community, a safe and healthy physical environment could be provided, transition was cost effective, 24-hr skilled care was not required, and an adequate support system was available that included both formal and informal supports. Passages participants received assistance in locating housing and appropriate services in the community, in addition to assistance while actually moving from the nursing facility to the community. Participants were eligible for continued transition services for up to 3 months post-transition (Schaefer & Eiken, 2003).
The Passages program employed several strategies to identify potential participants. It sent information about the program to advocacy groups, nursing home directors, and social workers. In some cases, program staff made onsite presentations. Local newspapers and the aging network newsletter published information about the program. A Web site and toll-free number were also available for information and referral. Finally, ombudsmen assisted in the identification of potential participants (Schaefer & Eiken, 2003). Many of the Passages strategies resembled those used by other first-generation NFT programs.
|
Methods |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Using MDS 2.0 data, we had an extremely broad range of characteristics available by which to classify individuals who transitioned. These included close to 400 items in a full MDS assessment and a variety of validated scales (available in both full and quarterly assessments) that pulled together multiple MDS items into a single concept (such as cognitive performance). Given that there were only 111 individuals in the transition sample, there was concern that the analysis could be overspecified with too many classification variables. Thus, the approach to select the classification variables included several steps. First we selected major scales known to be useful in profiling nursing facility populations. These included the Cognitive Performance Scale, the Activities of Daily Living Hierarchy, the Depression Rating Scale, the Communication Scale, and the Psychosocial Well-Being Scale. (Details of the construction of these scales and citations to the research describing their development are all available from the corresponding author. Along with most MDS items, all scales listed assign 0 to the absence of a problematic condition, with higher numbers associated with higher dependency or dysfunction.) We measured the intensity of care provided to an individual using the nursing case-mix index of the Resource Utilization Group (RUG-III) system. The project group reviewed the remaining items as being potentially important in identifying the population, representing different dimensions than the selected scales and the RUG-III case-mix index, and having nontrivial prevalence (i.e., more than 5%) in the 111-person NFT group.
The second step was to identify whether items of interest were available in the data sets drawn for the research. Some variables of interest (e.g., change in cognitive status), although standard MDS items, are not on the MDS quarterly assessment and, therefore, were missing for many cases. Other variables of interest, such as living arrangement at admission, education, and race, are collected only on the MDS background face sheet filled out at the initial assessment. The data sets did not include face sheets for individuals admitted outside our research time frame. Hence, such items were also missing for part of the sample, and we did not know if this would provide a bias to any results using them. Thus, the results describe the distribution of these items in the subsample where they are not missing, but we excluded them all from the remainder of the present analysis. Finally, we computed two variables by comparing dates in the MDS: the day of stay (i.e., current day of stay at the time of the assessment) and the current age of the individual (in years, at the time of the assessment). In calculating day of stay in the derivation sample, we discovered missing data for 6 Passages participants and 159 nonparticipants. For the present analysis, we assumed these individuals to have resided in a nursing facility for more than 730 days, under the assumption that shorter lengths of stay were likely known (dropping these cases would not have significantly changed the results). The variables identified in this process are those listed in Table 1. We tested all differences with the chi-square statistic and deemed them significant if p <.05.
|
The AID package used here was part of the Enterprise Miner programs in SAS (SAS Institute, 2003). This software allows for an interactive mode in which the user decides which split makes both statistical and "clinical" sense (e.g., the AID-recommended split on age was at a threshold slightly less than 65 years, but testing showed that splitting at 65 would have had almost no effect on the results; we took the same approach for day of stay). AID also permits the researcher either to omit missing data from the sample or to put them into a specified group; we discuss this issue later with respect to day of stay, which was missing on a large number of observations. Although it is preferable to include in any such work a final validation step using a sample not employed in the derivation, the number of NFT individuals was too small to consider splitting them out into an independent validation sample. Thus, we performed a partial validation by fitting the AID model to the larger data set so that we could compute the prevalence of Passages individuals in each group and compare it to the results from the random sample of 1,000 individuals.
|
Results |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Out of those for whom data were available, similar percentages—71% (32 out of 45) of Passages participants versus 61% (467 out of 769 non-Passages participants)—indicated a preference to return to the community (MDS Item Q1a), and the difference was not significant.
Several of these bivariate results remained in the multivariate analysis. Application of the AID technique provided a system of 13 terminal groups broken out variously by six criteria: age; current day of stay; RUG-III case-mix index; behavior problems; cognition; and the presence of hemiplegia, paraplegia, or quadriplegia. Figure 1 shows the terminal groups and their ability to predict NFT or non-Passages individuals.
|
|
|
Discussion |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
The results of the AID analysis confirm this observation. The analysis utilized age as a first split to identify a group of individuals aged 65 and older and a group younger than age 65; this represents statistically the most substantial characteristic differentiating the NFT individuals. Significantly, we found only one of the end groups that was most like the Passages sample in the older-than-65 branch; this group, identified as E, had no cognitive impairment and was very light care (case-mix index less than 0.58). Cognition also plays a central role in the AID analysis and was a third-level split in the two main branches; there were no combinations of characteristics identifying Passages individuals that included those with a CPS score above mildly impaired. Resource use intensity, as measured by case-mix scores, was a second-level split only within the older-than-65 branch; however, various kinds of paralysis were used to differentiate Passages individuals in the younger-than-65, very long-stay branch. Day of stay proved another predictive characteristic, being employed both as a second- and a fourth-level split.
These results also demonstrate the utility of AID techniques for identifying unique groups among the total sample in an efficient manner. The five groups that represent 62% of the Passages individuals, as well as the sixth group representing the preponderance of nursing facility residents who had not transitioned, were created using at most five characteristics. (Severe behavior was useful in differentiating groups in the middle of the spectrum of sensitivity and specificity, but it was not needed to identify either of the two versions of target groups that were very likely to contain high percentages of Passages residents.) Although we do not put this particular algorithm forward here as the way to target likely transition candidates, the approach does suggest a straightforward set of criteria with which to design an initial screening algorithm. Such an algorithm could be built into a state's MDS reporting system to flag potential participants, or it could be used by staff at local programs as a quick screener.
Depending on the desired goal in this case, two possible approaches could be employed. The first would identify individuals who "look like" Passages participants (boxes A through E in Figure 1), only 1.5% of individuals in nursing facilities. Over the course of a year, this approach would identify for consideration approximately 336 of the more than 16,700 Medicaid-eligible individuals who utilize the Arkansas nursing facilities each year (Fries, James, & Park, 2005; Nursing Home Data Compendium, 2005). If one bases the computations on the larger sample of the estimated population from which the Passages individuals were identified, the sensitivity of this approach is 62%; in other words, this strategy would correctly place into the target group two thirds of the individuals resembling Passages participants. The specificity of this approach is 98.5% (98.4% in the validation data); thus, such a strategy would incorrectly identify individuals as resembling non-Passages individuals only 1.5% of the time.
The second, broader approach would identify all nursing facility residents except those meeting the criteria for group X in Figure 1 (i.e., this approach would target all residents who either are younger than age 65 or have a case mix index of less than 0.58). This approach would have a sensitivity of 92% and a specificity of 83% (78% in the validation data). However, over the course of a year it would identify for evaluation more than 22%, or 3,685, of all 17,600 Medicaid-eligible individuals. Given that the goal of most NFT programs is to cast a wide net to find every person who might be a likely candidate, this latter approach seems preferable. Although it would require substantial effort to contact more than 3,500 individuals to find what might be as small as an expected 102 cases (2.55%, the positive predictive value), it would also substantially reduce the search of all residents by three quarters. In another state—or even in Arkansas as more experience is obtained—the characteristics of those identified for NFT will be different and researchers can undertake a similar analysis to rederive an identification algorithm.
Examining information describing nursing facility residents in Arkansas who were discharged to the community has enabled us to pinpoint an efficient system to help identify similar future targets for NFT. The approach uses data that are readily available and can be modified over time as additional experience is gained. Nevertheless, such targeting should not replace the intent of NFT initiatives (i.e., to allow any person to question whether institutionalization is necessary and to be considered for discharge to the community).
Although the approach put forth here was successful in describing Arkansas's NFT program, it would be a mistake to assume that the results will generalize to other state NFT programs. These other programs may target (and successfully transition) completely different individuals, and our findings are not meant to suggest what is either the only or the best target for NFT. Thus, the primary purpose of displaying this research is to illustrate the potential of using these approaches and techniques to inform stakeholders' understanding of who was actually served in any NFT effort. Whether the Arkansas results can generalize to other states is the subject of future analyses.
Even within the Arkansas context, the research is limited in several respects. The sample size was small, and some key information gathered only on the MDS face sheet was missing; this limited the choice of candidate predictor items. Also, longitudinal data on participant outcomes after transition were not available, thus it was not possible to examine here whether particular characteristics or additional analyses might have identified not only those individuals who left nursing facilities, but those who were able to remain in the community on a long-term basis. From a technical point of view, there were insufficient numbers of Passages residents to split the sample and validate the AID classification model; the application of the derived algorithm to the larger database but with the same 111 Passages residents is certainly not a strong test of validity. AID analysis applied to the full validation data set provided relatively similar splits to those found with the derivation data, but, as is often the case with AID, alternative models could be derived, and the larger number of observations made it possible to provide alternative, sometimes more predictive, groups with the same statistical significance. As well, without detailed information on which specific residents were considered for NFT under the Passages program, the comparison sample in the validation database is, at best, an approximation. As the number of individuals served by the Arkansas Passages program increases, as the state's experience with identifying candidates and accomplishing such transitions grows, and, perhaps most important, as information accumulates about the outcomes of transitioned participants, researchers can extend these analyses.
It is tempting to speculate about the degree to which the organizations that carried out Passages on the local level influenced the participant profile. It is possible that the staff of the Centers for Independent Living, already serving younger individuals with disabilities, were equipped with the skills and knowledge necessary to establish rapport with and arrange services (particularly housing) for the younger disabled individuals encountered in the nursing facilities. The staff from the Area Agencies on Aging who took part in Passages may have faced a steeper learning curve that resulted in proportionately fewer older adults being served within the time frame of the study.
We hope that these findings will stimulate similar research, particularly as states move forward with the Money Follows the Person Demonstration authorized by the federal Deficit Reduction Act of 2005 (DRA). Certain mandates in this legislation are likely to pose new and significant challenges for NFT programs, particularly the requirement that restricts enhanced state Medicaid matching funds for home- and community-based services used after transitions to individuals who have resided in a nursing facility or other institution for at least 6 months (Crowley, 2006). Curiously, the optimum split points selected by the AID techniques, close to 60 and 730 days, do not parallel the target of 180 days to 2 years as specified in the DRA; this suggests that the DRA target may be overly restrictive. However, in light of the DRA mandate, it behooves states wishing to earn additional matching funds to compare how closely their first-generation NFT population resembles the new Money Follows the Person target group. For instance, the rarity of the Passages population suggests that future program costs in Arkansas might be different if another target population were served.
To benefit from this kind of analysis, states overseeing NFT initiatives need to adopt data collection procedures that track individuals longitudinally across home- and community-based services and institutional settings; this would also enable the development of a more robust definition of NFT success. Future efforts to pool NFT data across states also seem desirable, as this would enrich the emerging literature on NFT program structures with a comparative analysis of the subpopulations and outcomes of individuals served in each state.
|
Footnotes |
|---|
1 Institute of Gerontology, University of Michigan, Ann Arbor.
2 School of Public Health, University of Michigan, Ann Arbor.
3 Geriatric Research, Education, and Clinical Center, Ann Arbor VA Medical Center, Ann Arbor, MI.
Decision Editor: Linda S. Noelker, PhD
Received for publication November 7, 2006. Accepted for publication June 6, 2007.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
| ||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
