Abstract
Conducting studies with sufficient statistical power has always been seen as an impediment for quantitative research in the field of AN [1]. Although AN is the third most common chronic illness in adolescent girls [2], within the whole spectrum of medical and psychiatric conditions, it is a low prevalence disorder. Most inpatient centres have few beds, and this combined with lengthy admissions, means that it would take years to systematically collect sufficient data in any one location. Multisite studies are now encouraged as a means of resolving such issues [3]. Sometimes also referred to as a practice research network (PRN) [4], such databases provide a mechanism for networks of clinicians to study patients, treatments and outcomes in typical clinical settings. Both the clinician/researcher and patient, are ‘real world’ [4: p.1199] participants. In this way, it is argued that the research findings are more easily transferred into clinical practice. This form of research design also potentially minimizes the problem of variables emerging as significant because of site idiosyncrasies. Research networks such as this have been established in psychiatry investigating a range of diagnostic groups and aspects of their routine care [5–7].
Recently developed at the University of Sydney, the AN inpatient database is a multisite collaborative project collecting data on admission and discharge on all consenting participants admitted to selected inpatient settings. The purposes of this database are multiple, and include the following:
To determine the demography of people with AN who access specialist inpatient treatment; To provide ongoing evaluation that such treatment programs are admitting patients in line with existing and relevant admission guidelines To document response to inpatient treatment and any complications that arise during its course To document clinical condition at discharge and follow-up; To develop clinical predictors of service utilization, distribution and provision.
Length of hospitalization
Although there is ongoing debate about which treatment modalities work best in anorexia nervosa (AN), there is wider consensus that hospitalization is necessary under certain circumstances. Namely, this step should be taken to manage specific symptoms such as severe bingeing and purging, to interrupt and reverse weight loss, and to treat serious physical or psychiatric complications and risks. Indeed, not to do so under such circumstances has been described as perilous [8]. Reflecting this consensus of opinion, clinicians are now guided to hospitalize before such patients present as medically unstable [9].
Inpatient treatment for psychiatric disorders is widely considered the most expensive and time-consuming episode of care, and AN is no exception. Striegel-Moore et al. [10] recently reported an average admission cost for women at $17 384 but with significant variation around this (SD = $24 394). At an average of 26 days, such patient stays were significantly shorter than that considered necessary for adequate nutritional rehabilitation, or that reported by other specialist treatment centres in Australia and New Zealand [11–14]. Such short stays have been described as cost-ineffective and potentially counterproductive [15, 16], yet shortening admissions remains an obvious and ongoing ambition for funders who describe length of stay (LOS) as both a proxy for cost [17] and a measure of success [18].
Examined in this way, predicting LOS emerges as an important focus for a variety of stakeholders and especially important for clinicians and consumers of such services. Not only is it a pragmatic consideration in the development of inpatient programs, but informing consumers of the likely course of a treatment also constitutes an important part of the explanation regarding informed treatment choices and procedures.
The prediction of LOS in general psychiatric settings has been the subject of multiple studies investigating a multitude of factors [19]. These generally cluster around three types of variables: those relating to patients; those relating to systems; and those relating to treatment [17]. While including something from each cluster generally has been found to increase predictability [20], findings regarding individual variables within each cluster have not always been consistent. For example, older age and being female have been associated both with longer stays [21–23] and shorter stays [24, 25].
Although many studies describe length of AN hospitalization, we are aware of only one study investigating factors predictive of this [26]. These authors found that poor social adaptability (defined as nonattendance at either school or work prior to onset), habitual stimulant abuse, low body weight after AN onset and at admission, younger age at onset and at admission, and previous treatment for an eating disorder, were all associated with a longer LOS. However, this study excluded patients who left treatment early (prior to attaining 80% of minimum healthy body weight) and hence it is limited in its ability to predict LOS for any one patient approaching an inpatient facility.
Inevitably, LOS in AN will depend on a number of factors, and only some of these may be intuitively known in clinical practice. For example, those admitted in an extreme state of emaciation can be expected to take longer to reach their goal weight range or nutritional stability. Often the more immediate problem is one of encouraging patients to stay long enough to meaningfully achieve the goals of hospitalization; many have strong resistance to treatment and, despite the efforts of family and clinicians, prematurely discontinue therapy [27]. Indeed, LOS may be as much a result of person factors as of treatment protocol, treatment setting, or clinical phenomena. Whatever the case, the current eating disorders literature leaves clinicians and researchers with few clues about what factors may be pertinent in determining hospital LOS in AN, and this constitutes part of the wider paucity of information about hospitalization in this population [9].
Utilizing this database, we report on the cohort of participants in the first 20 months of the data collection. Specifically, the study investigates the degree to which demographic, clinical, and patient factors collected at admission through this database may predict LOS in hospital.
Method
Establishment and description of the Multisite Database Project
In Sydney Australia there are seven hospitals and clinics providing treatment for AN (Concord Hospital, New Children's Hospital, Northside Clinic, Rivendell Adolescent Unit, Royal Prince Alfred Hospital, Wesley Private Hospital and Westmead Hospital). In 1997 a collaboration agreement was made between all sites to begin development of a common dataset for the assessment and evaluation of inpatient treatment for AN. At a later stage two international sites (Eating Disorders Service, Leeds, UK, Christchurch Eating Disorders Service, New Zealand) joined the project. At the time of writing, complete data had been obtained from five of the eight sites: Concord Hospital; Northside Clinic; Royal Prince Alfred Hospital; Wesley Hospital (all Australia); and the Christchurch Eating Disorders Service (New Zealand).
Variables included in this common dataset were derived collaboratively by a small team of experts from the Royal Prince Alfred Hospital, Westmead Hospital, Rivendell Adolescent Unit, and the Department of Obstetrics and Gynaecology at the University of Sydney, led by one author (PB). The data items for initial consideration were chosen in accordance with the Australian National Minimum Dataset (NMDS). Only items applicable to AN were included, and some additional items were designed to measure specific aspects of AN and its treatment. This resulted in a dataset of variables justified by one or more of the following reasons:
It is pertinent to the development and symptomatology of AN; It is thought to be a predictor of illness outcome; It is an indicator of epidemiological trends; It is an indicator of service utilization, distribution, and provision; It assesses and relates to the treatment being provided on site.
Thereafter a data dictionary defining each item of the dataset, the rationale for its inclusion, and the precise system of measurement, was written and disseminated to sites (the format of the complete dataset, and detailed methods of its administration, is available on request from the corresponding author). Data collection began at a number of sites in late 1998, with active involvement being determined by the amount of available resources at the site. Other sites delayed data collection until adequate resources could be found to fund the project.
Systematic data checks by the project coordinator (SM) ensured that each site conformed to the data dictionary in a uniform fashion, and that data integrity was maintained from collection through to data entry and analysis. The participating treatment sites all provide 24 h inpatient treatment, and although each contends with its unique service provision issues and policies, clinicians report a high degree of liaison and collaboration with each other through joint research and professional meetings. Each treatment site participated in the preparation of common clinical practice guidelines for treatment of anorexia nervosa in Australia and New Zealand, and the health professionals across these two countries are subject to reciprocal registration standards. Weekly communication was maintained between all sites and the project coordinator (SM) who also maintained the master database held at the University of Sydney.
Participants and recruitment procedure
This paper includes data from the five active sites that began data collection in late 1998. At the time of analysis in late 2000, this sample constituted 218 cases (admissions) with a primary diagnosis of AN. Diagnosis according to DSM-IV criteria [28] was confirmed at each site by the clinical team at the first ward round following admission. Unless clinical status over the admission week rendered any approach inappropriate (e.g. ongoing medical crisis, patient unable to give informed consent) all such patients were approached by the site database co-ordinator during the week of their admission, and invited to participate in the database. After obtaining written consent, all data were collected within 7 days of the admission date, and again within 7 days of hospital discharge. Sources of data included the clinical file, clinician ratings from the primary responsible clinician, patient checklists, and the following standardized self-report inventories: Eating Disorder Inventory-2 (EDI-2) [29]; Eating Attitudes Test (EAT) [30]; Beck Depression Inventory (BDI) [31–34]; and Rosenberg Self-Esteem Scale (RSES) [35, 36].
Statistical analysis and procedures
Possible predictors of outcome in AN, as reported in other studies, were evaluated for their relationship to LOS inpatient treatment. These included: gender; occupation; education; marital status; age at onset of eating disorder; age at admission; minimum body weight after onset of eating disorder; body weight at admission; previous treatment for an eating disorder; a binge/purge subtype; abuse of dieting prosthetics; and a comorbid diagnosis. In addition we examined the relationship between LOS and the following factors: drug use prior to admission; menstrual status on admission; and presence of medical abnormalities (electrolytes, liver/renal enzymes, ECG.).
The relationship between LOS and patient and clinical variables was investigated in two stages. First, by means of one-way ANOVA for categorical variables, and Pearson correlation for continuous variables. Second, the utility of significant demographic or clinical variables (independent variables) for the prediction of LOS (dependent variable) was investigated using multiple linear regression creating dummy variables for the categorical independent variables. An alpha level of 0.05 was used on all tests.
Results
Characteristics of the sample
The demographic and clinical characteristics of the sample are described in Table 1. As expected from the known epidemiology of AN, of the 218 cases (n = 154 participants) included in the database, almost all (98%) were young women experiencing many years of AN. Most (70.6%) cases had a single admission recorded on the database, 17.6% had two admissions, while the remaining 11.5% had three or more admissions.
Admission demographic and clinical characteristics of the cases
The vast majority of cases (60.8%) reported ‘study’ as their main occupation, while 17.5% were employed, 12.4% unemployed, 5.1% performing home duties and 1.8% unable to work due to illness. Sixtysix percent reported some prior inpatient treatment for an eating disorder. Just under a half of cases (49.5%) met DSM-IV criteria for a comorbid diagnosis: 28.2% with an affective disorder, 6.4% with an anxiety disorder, and 3.2% with a personality disorder. With a mean BMI (body mass index) of 14.5, most participants presented with severe AN, and as such, were representative of the cases expected to enter inpatient facilities [9].
Relationship between admission variables and LOS
A significant relationship was found between LOS and nine admission variables (Table 2). Length of stay was significantly associated with BMI (r = −0.28, p = 0.000), and lowest weight ever (r = −0.17, p = 0.014), and number of years of education (r = 0.17, p = 0.015). ANOVA indicated a significant association between LOS and site (F-statistic [4, 206] = 4.64, p = 0.001) (Table 2). Pairwise comparisons indicated that LOS at the Christchurch site was significantly longer than all other sites. Area of residence was also significantly associated with LOS (F-statistic [2, 207] = 5.64, p = 0.004), with pairwise comparisons indicating that participants residing in areas outside the state/ region housing the treatment facility had a significantly longer LOS than participants living locally or within the state/region. The category of number of previous inpatient admissions was significantly associated with LOS (F-statistic [2, 205] = 3.91, p = 0.021), with pairwise comparisons indicating that participants with 2–3 previous admissions had significantly longer admissions than participants with 0–1 or more than three previous admissions. Admission menstrual status category was significantly associated with LOS (F-statistic [5, 198] = 2.97, p = 0.013), with those reporting amenorrhea unrelated to an eating disorder (e.g. menopausal, breast-feeding, etc.) experiencing longer LOS than all other groups except those reporting normal menstrual cycles. Finally, LOS was significantly associated with the presence of an abnormal white blood cell (WBC) count (F-statistic [1, 188] = 4.88, p = 0.26), compliance with psychometric questionnaires on admission (F-statistic [1, 209] = 9.49, p = 0.002).
Length of stay for categorical and dichotomous variables where ANOVA or t-test indicated significance
To assess the utility of each significant variable in independently contributing to LOS, a multiple linear regression was used. All variables found to have a significant univariate association, including site, were entered into the model (Table 3). As a result, two variables were found to contribute significantly to the model, namely BMI (F-statistic = 18.92, p = 0.000), and the category of previous inpatient admissions to hospital (F-statistic = 8.36, p = 0.000). The overall adjusted R 2 for this model was 0.198.
Multiple regression of significant admission variables on length of stay
Discussion
We describe the development and implementation of an Australasian international database focusing on the most expensive and intensive aspects of AN treatment, namely inpatient care in specialized centres. We then investigate the extent to which this database can predict, at point of admission, LOS in the participating treatment centres. Although multisite database studies such as this overcome many commonly reported problems in AN research, capture data on real patients receiving real treatments, and can collect data from a variety of perspectives (in this case, clinicians, patients, files, and site), there are a number of limitations. The study design limits the ability to make causal inferences, and data collection was carried out by clinicians with varying levels of specialized research training who all had other ongoing clinical responsibilities [4]. A comprehensive data dictionary precisely defining each variable was provided to each site, and it is argued that this maximized the rigor of data collection methods. Despite the disadvantages of practice-based networks such as this, once established, the costs of ongoing data collection and collaboration with other treatment centres is low. They also provide a rapid method by which specialist and/ isolated services can, at any point, compare local standards and practices with similar services. Given the continually evolving and sometimes polarized debates about the utility of inpatient treatment for AN, the provision of systematic information regarding the clinical practice and outcome of this, by means of this database, makes an important contribution.
Length of stay is an important [15, 17] yet seemingly difficult aspect of treatment to predict. In this study, of the multitude of variables routinely assessed on admission, and typically of interest to clinicians in planning inpatient treatment, very few predicted LOS and the best-fit model accounted for just under one-fifth of the total variance. At this time when LOS, and its prediction, is a prominent issue for all stakeholders, existing research provides imprecise models, and future research is clearly needed.
The finding that a lower BMI was predictive of a longer LOS is expected. The primary focus of inpatient treatment at all sites involved in this study, and indeed most inpatient units for AN, is weight gain. Patients who present with a low BMI are likely to be encouraged to stay longer in order to regain weight, and this is a finding reported in a related study [26]. Leaving admissions until a lower weight clearly has fiscal, not just therapeutic implications.
The finding that the category of 2–3 previous inpatient admissions was associated with a longer LOS has not been reported elsewhere in the literature, and therefore is an important finding. It does however, fit with clinical observations that the first inpatient admission into a psychiatric setting can be very distressing, which in turn may lead to earlier discharge. At the other end of the continuum, those with chronic disorders having experienced multiple admissions may have different goals (e.g. crisis-focused symptom restabilization), which in turn may also lead to shorter admissions. It may well be that it is not until the second or third hospitalization that the more lengthy weight gain-orientated admissions can be tolerated by, or more successfully negotiated with, such patients. If so, this has important implications for how clinicians may develop a frame of reference around what may otherwise be construed as patients not completing treatment. Future research is needed to ascertain whether, in the overall treatment career of someone with AN, the time around the second or third inpatient admission presents a better window of opportunity for achieving certain goals of admission rather than others.
While treatment site no longer remained significant once other factors were taken into account, the considerably longer LOS at the Christchurch site invites comment. As the only New Zealand public sector specialist eating disorders service with a dedicated inpatient facility, potentially this site is more likely to see the out-ofregion patients who, this study has shown, are more likely to receive lengthy hospitalizations. Indeed, 18% of Christchurch admissions were from out-of-region (in this case, the North Island) whereas at 5% the Northside Clinic was the next site with most frequent admissions from out-of-region. Treating patients a long way from their domicile may mean that clinicians feel obliged to give as much treatment as possible, since access to discharge follow-up from the treating team is not available. Those admitted out-of-region also presented with a lower BMI than those within the region (p = 0.043), or within the metropolitan centre (p = 0.065) of the site. This could also explain in part the LOS in Christchurch. Whatever the case, this situation is likely to have significant clinical and funding implications for this particular site if it were to remain the sole New Zealand option in the public sector.
Likewise, while men appeared to have shorter mean LOS compared with women (28.6 days vs. 63.9 days), this was not significantly different (p = 0.15). The small sample of men (n = 5) almost certainly contributed to the failure to find a difference, and larger studies are needed in order to pursue this issue. For example, the bias in the treatment program towards issues of concern to women may curtail admissions for men, or it could be that the clinical issues prompting their admissions are more quickly resolved.
Externally imposed criteria (e.g. from third parties) will always play an prominent role in determining LOS, and such predetermined criteria can already be part of the admission planning and negotiating discussions held with patients. However, the question remains as to what are the more subtle issues at play within an admission. In this sense, the predictors of LOS await further investigation as, even when very comprehensive information is available at admission, at best such variables (and any models derived from them at present) have very modest predictive ability. At the very least, future research needs to investigate other variables and even other research methodologies as a means of revealing candidate variables for consideration. Qualitative methodologies have recently proved fertile in developing understandings of related treatment phenomena on both sides of the eating disorder patient–clinician equation [37–39], and perhaps this could be a promising means of revisiting this issue.