INTRODUCTION

During recent years, there has been an increasing focus, by administrators, payers, and health services researchers, on length of stay (LOS) as an important indicator of quality and efficiency of inpatient psychiatric hospitalization. This growing attention is related to the increasing pressure for containment of costs of psychiatric care and the ongoing focus on treatment in the least restrictive setting of care. Although longer inpatient hospitalizations are associated with greater direct and indirect costs of treatment, the costs of inadequate stabilization during hospitalizations that are too short also must be considered. As interest in LOS continues, the identification of the predictors of LOS is crucial.

Prior research has provided some potential determinants of psychiatric LOS, and three important factors that have emerged from this literature were of particular interest in this investigation. These include: (1) commonly used measures of clinical and psychosocial functioning, including the Global Assessment of Functioning (GAF) scale, (2) the use of seclusion and/or restraints during hospitalization, and (3) the presence of a comorbid substance use disorder. Several studies have found the simple GAF rating or other measures of illness severity to be predictive of LOS, with greater illness severity associated with longer LOS (1, 2). Concerning seclusion and/or restraints as a determinant of LOS, LeGris and coworkers reviewed the charts of 85 patients with psychosis on admission to an urban community general hospital, and found a significantly longer mean LOS in secluded patients compared to non-secluded patients (3). Incidents of physical aggression during hospitalization (which are presumably closely associated with the use of seclusion and/or restraints) also have been shown to be associated with longer LOS (4). A number of studies have demonstrated that the presence of a substance use disorder is related to shorter psychiatric LOS (2,5–8). Interestingly, while psychotic disorders and severe affective disorders are associated with longer LOS, the presence of a comorbid substance use disorder decreases LOS (6, 7). Although other factors have been found to be associated with LOS in various settings, the results are quite inconsistent, and are likely heavily dependent on institutional and local practices. While some have noted that the field has become skeptical that any particular variables can reliably predict LOS across psychiatric facilities, prior studies show that patient-related correlates of LOS in a single hospital can be stable over time and can be readily discovered using relatively simple multivariable statistical procedures (5).

This analysis was conducted using data from a study that examined predictors of nonadherence with the first outpatient community mental health follow-up appointment after psychiatric hospitalization (9). The aim of this analysis was to investigate the predictors of LOS on two inpatient psychiatric units in a large, urban, university-affiliated, county hospital. It was hypothesized that three variables would be significantly independently associated with shorter LOS, based on prior research and clinical experience: (1) higher GAF scale scores, (2) not requiring the use of seclusion or restraints during hospitalization, and (3) the presence of a comorbid substance use disorder diagnosis. In addition to these specific hypothesized predictors, this study sought to determine other sociodemographic and clinical predictors of LOS on these two inpatient units and to assess the amount of variance in LOS that can be explained using these variables in this particular setting. The determinants of LOS were studied for each unit separately because of the very different treatment philosophies and average lengths of hospitalization on the two units.

METHOD

Setting and Sample

This study took place at large, urban, university-affiliated, county hospital that serves a predominantly African American, socially disadvantaged population. Of the two locked inpatient psychiatric units in the hospital, one is a crisis stabilization unit (CSU) with a capacity of 8 patients and a target LOS of 3–7 days. Due to its focus on crisis intervention and rapid return to the community, short LOS traditionally has been one of the defining features of this unit. The other unit is a longer-stay “milieu” unit (LSMU), which can accommodate 22 patients with a target length of stay of about 10–14 days. This unit provides more intensive evaluation and is better equipped to care for psychiatrically, medically, and psychosocially complex cases. On this unit, LOS traditionally has been minimized as a focus of attention, with more emphasis placed on thorough evaluation by an interdisciplinary team, in the milieu model of inpatient care. However, reducing LOS increasingly has become important for the LSMU over recent years. The majority of the patients on both of these units are treated for acute exacerbations of major psychotic and affective disorders and there is a high incidence of comorbid substance use disorders among these patients.

Consecutive discharges from each of these two units were studied as part of a project to examine predictors of outpatient nonadherence (9). Due to the inclusion criteria of the parent study, charts of patients discharged with follow-up appointments to clinics other than the designated clinics in the metropolitan county in which the hospital is located (e.g., to private practices or to community mental health centers in other counties) were excluded. Only the first hospitalization per patient during the study period was included, to maintain independence of observations. Thus, the charts of patients who were readmitted during the study period were excluded.

Procedures

The study design was a medical record review of patients’ charts during their hospitalization. Patients were not assessed directly. Basic demographic, socioeconomic, psychosocial, and clinical data were extracted from the medical charts and were obtained by consulting with the social worker, the psychiatry resident or psychology intern, and the attending psychiatrist caring for the patient during hospitalization. Variables were recorded in a systematic manner using a structured data collection instrument. Data were gathered on a total of 234 patients, between December 2003 and July 2004. The research was approved by the university's institutional review board and by the hospital's research oversight committee.

Analyses

Patients on the CSU and the LSMU were compared to assess for any statistically significant differences between the two groups with respect to sociodemographic characteristics. Distributions of primary discharge diagnoses and other psychiatric diagnoses were also compared. Bivariate analyses (using the independent samples Student's t-test, Pearson's product-moment correlation, and analysis of variance) were performed on the potential sociodemographic and clinical predictor variables for each respective unit in relation to the outcome variable of interest (LOS). The bivariate analyses were conducted to exclude from the linear regression models those clinical and sociodemographic variables that were not associated with LOS.

Separate multiple linear regression models were built for the CSU and LSMU using those variables that were associated (p < .10) with LOS in the bivariate tests. The stepwise entry method was used to model the relationship between LOS and the predictor variables for each inpatient unit. In this selection method, which is a combination of forward selection and backward elimination, after a variable is entered into the model, any variables already in the model that are no longer significant predictors are removed (10). The criterion for entering the variable was set at <.10, which allowed for exploration of all variables that were associated with LOS in bivariate tests at the p < .10 level. Although the p < .10 level was used for model entry, interpretation of model results focused on independently statistically significant predictors (p < .05). Regression diagnostic procedures were performed to verify that all of the assumptions for multiple linear regression modeling were met. Normality of LOS on both the CSU and LSMU was assessed prior to performing linear regression modeling, and transformation was used as needed to reduce skewness of distributions for linear regression. The SPSS 12.0 software was used for all statistical tests.

RESULTS

Sociodemographic characteristics of the two groups of patients are presented in Table 1. Among these 10 variables (age, gender, race, years of education, marital status, employment status, living with children ≤18 years of age, homelessness, receiving disability payments, and largest source of payment for medical care), three were significantly different between the two units. CSU patients were slightly older (40.8±11.8 years) than those admitted to the LSMU (37.4±12.6 years; t=2.00; df=230; p=.05). CSU patients were more likely to be receiving disability payments—Supplemental Security Income (SSI) or Social Security Disability Insurance (SSDI)—compared to LSMU patients (60.2% compared to 46.6%; χ2=4.10; df=1; p=.04). Similarly, CSU patients were more likely to have Medicare (19.8% compared to 7.6%; χ2=7.84; df=2; p=.02).

TABLE 1 Characteristics of Patients \(\boldmath{(n=234)}\) Admitted to Two Psychiatric Inpatient Units in a Large, Urban, University-Affiliated, County Hospital
Full size table

Distributions of primary discharge diagnoses for the two groups are shown in Table 2. The two groups did not differ in terms of the proportions diagnosed with schizophrenia and other psychotic disorders (60.9% of CSU patients compared to 67.8% of LSMU patients) or depressive, bipolar, and anxiety disorders (27.6% of CSU patients compared to 30.2% of LSMU patients). In addition to primary discharge diagnoses, other psychiatric diagnoses at the time of discharge did not differ between the two groups. Specifically, 45.5% of CSU patients and 44.1% of LSMU patients had discharge diagnoses that included a substance use disorder (abuse or dependence). Some 15.9% of CSU patients and 14.4% of LSMU patients had comorbid personality disorder diagnoses at discharge.

TABLE 2 Distributions of Primary Discharge Diagnoses of Hospitalized Patients \(\boldmath{(n=234)}\)
Full size table
FIGURE 1.
figure 1

Length of stay distributions.

Full size image

The distribution of the LOS was assessed for both units (Figure 1). Among CSU patients, the range of the LOS was 1–22 days. The mean LOS was 6.8±3.0 days, with a median and mode of 7 and 8 days, respectively. To normalize the distribution for the multiple linear regression model, one outlier observation (LOS of 22 days) was removed. Excluding the outlier resulted in a LOS distribution with a mean, median, and mode of 6.7±2.6, 7.0, and 8 days, respectively. The skewness and kurtosis statistics for this distribution were 0.61 and 1.55, respectively. On the LSMU, the range of the LOS was 5–48 days, with a mean and median of 14.4±8.4 and 12.0 days, respectively; and modal values of 7 and 8. Because of obvious skewing to the right, this distribution was normalized by log transformation. (The transformed variable is referred to as log10LOS.) This transformation resulted in a log10LOS distribution with a mean, median, and mode of 1.09±0.2, 1.08, and .85, respectively. The skewness and kurtosis statistics for this distribution were 0.40 and –0.40, respectively.

Bivariate tests were conducted to determine whether or not any of the 10 sociodemographic variables were significantly associated with LOS in either of the two hospital units. Among the 88 patients on the CSU, there were no significant associations between LOS and any of the sociodemographic variables. Among the 146 patients on the LSMU, two sociodemographic variables were significantly associated with LOS—gender and marital status. Specifically, the mean LOS for females on the LSMU was 12.8±6.3, compared to 16.8±10.2 for males (t=2.94; df=144; p < .01), and the mean LOS for separated/divorced/widowed patients on the LSMU was 11.8±6.8, compared to 15.6±9.0 among those who were single/never married (F=3.62(2,143); p=.03). Additionally, the association between LOS and two other sociodemographic variables approached statistical significance for the LSMU patients. That is, there was a slight inverse correlation between LOS and age (r=–0.16; p=.05), and the mean LOS of employed patients was lower (9.4±2.3) than that of unemployed patients (14.8±8.5; t=1.88; df=143; p=.06).

Bivariate tests were also carried out to determine significant associations between LOS and 13 clinical variables of interest: (1) presence of psychotic symptoms on admission, (2) comorbid personality disorder discharge diagnosis, (3) comorbid substance use disorder discharge diagnosis, (4) prior psychiatric hospitalization, (5) medical problems requiring medications during hospitalization, (6) the use of seclusion or restraints, (7) whether or not a family meeting was held, (8) treatment adherence on the unit during the week prior to discharge, (9) whether or not the patient was established with an outpatient clinician, (10) number of psychiatric medications on discharge, (11) legal status on discharge, (12) admission GAF, and (13) discharge GAF. In these bivariate tests, using a p < .10 criterion for significance for later entry into regression models, seven of these clinical variables were associated with LOS on the CSU (presence of psychotic symptoms on admission, personality disorder discharge diagnosis, substance use disorder discharge diagnosis, medical problems requiring medications during hospitalization, use of seclusion or restraints, legal status on discharge, and discharge GAF). Six variables were associated with LOS on the LSMU (presence of psychotic symptoms on admission, substance use disorder discharge diagnosis, use of seclusion or restraints, whether or not a family meeting was held, admission GAF, and legal status on discharge).

Sociodemographic and clinical variables associated with LOS were entered into multiple linear regression models for the respective inpatient units. Nine variables were entered into the model for LOS for CSU patients (the seven clinical variables found in bivariate tests, as well as age and gender). The final regression equation for the CSU (Table 3) was significant (F (5,79)=8.21; p < .01), with an R value of .59. The adjusted R 2 value was .30, indicating that the five remaining predictors explained 30% of the variance in LOS on the CSU. Using a more stringent criterion for significance (p < .05), two variables in this model were independently predictive: personality disorder diagnosis and substance use diagnosis. The presence of either of these types of diagnoses predicted a shorter length of stay on the CSU. Specifically, holding other variables constant, the presence of a personality disorder decreased the LOS on average 1.83 days (95% CI: .43, 3.23 days) and the presence of a substance use disorder comorbidity decreased the LOS on average 1.07 days (95% CI: .08, 2.06 days).

TABLE 3 Determinants of LOS Among Patients Admitted to the CSU \(\boldmath{(n=87)^*}\)
Full size table

Table 4 shows the results of the multiple linear regression model predicting log10LOS on the LSMU, which initially included ten variables (the six clinical variables found in bivariate tests, as well as age, gender, marital status, and employment status). The same entry method and parameters were used in this model. The final regression equation was significant (F (6,136)=15.39; p < .01), with an R value of .64. The adjusted R 2 value was .38, indicating that the six remaining predictors explained 38% of the variance in log10LOS. The five independently significant (p < .05) predictors in this model were: legal status on discharge, use of seclusion or restraint, GAF on admission, gender, and substance use discharge diagnosis. Specifically, involuntary legal status on discharge, not requiring the use of seclusion or restraints, higher admission GAF, female gender, and the presence of a comorbid substance use discharge diagnosis were independently associated with a shorter LOS on the LSMU. The finding of five independently significant predictors in the LSMU model, compared to only two in the CSU model may be related to the fact that there was greater variability in LOS on the LSMU. In terms of actual LOS in days (rather than log10LOS), holding other variables constant, involuntary legal status on discharge and not requiring the use of seclusion or restraints decreased the LOS on average 1.54 days (95% CI: 1.33, 1.78 days) and 1.45 days (95% CI: 1.21, 1.73 days), respectively. Female gender and the presence of a substance use disorder decreased the LOS on average 1.20 days (95% CI: 1.05, 1.37 days) and 1.16 days (95% CI: 1.01, 1.32 days), respectively. For each 10-point increase in the admission GAF, the LOS decreased on average 1.17 days (95% CI: 1.07, 1.29 days).

TABLE 4 Determinants of log10LOS Among Patients Admitted to the LSMU \(\boldmath{(n=146)^*}\)
Full size table

For both of the final models, regression diagnostics were examined to assure that the assumptions for multiple linear regression were met (10). The assumptions of constant variance and independence of LOS were examined through residual plots. Multicollinearity was assessed using variance inflation factors, and a global test was performed on the final model to assure goodness of fit. Both final models were found to have no significant outliers or influential observations.

DISCUSSION

Patients on the two units were largely comparable with respect to the measured socioeconomic and diagnostic characteristics—both groups consisted of predominantly African American, unmarried, unemployed patients with severe and persistent mental illnesses, who had either public-sector health insurance (Medicare and Medicaid) or no insurance of any sort. The CSU and the LSMU had median LOS values of one week and nearly two weeks, respectively (7 days and 12 days). This substantial LOS difference between the two units was expected, because of the differing treatment philosophies on the two units. The CSU specializes in rapid evaluation, stabilization, and return to community care. Patients also are selected for this unit based on a presenting history for which faster stabilization can be expected (e.g., several days of medication nonadherence causing symptom exacerbation, minor suicidal behavior). The LSMU had a longer median LOS likely due to: (1) admission bias toward complex cases with medical and substance use comorbidities, severe psychosocial problems including homelessness, or first-episode cases in need of thorough evaluation; (2) exclusion of cases expected to rapidly stabilize (they are preferentially admitted to the CSU); and (3) a treatment philosophy based on an interdisciplinary team approach in a “milieu” setting, comprehensive discharge planning, and the idea that longer hospitalizations are sometimes required for stabilization and prevention of relapse upon discharge. According to national data from adult specialty mental health care programs in 1997, the median lengths of inpatient care for patients in state/county mental hospitals and for patients in non-federal general hospitals were: 6.8 and 7.7 days for patients with schizophrenia, and 9.8 and 5.6 days for patients with affective disorders (11). These estimates indicate that the LOS on the CSU may be more representative of national averages.

While differences in LOS were expected, differences in the predictors of LOS were not. In terms of the original hypotheses, (that higher GAF scores, not requiring the use of seclusion or restraints during hospitalization, and the presence of a comorbid substance use disorder would predict shorter LOS), the analyses revealed that: (1) GAF scores were associated with LOS in the predicted direction on both units, (2) the use of seclusion or restraint was predictive of LOS on the LSMU but not the CSU, and (3) the presence of a substance use disorder diagnosis was predictive of LOS on both units. These findings are consistent with the results of some previous studies (1–8). In addition to these three hypothesized predictors, several other characteristics (personality disorder diagnosis, legal status on discharge, and gender) were found to be predictive of LOS.

The linear regression models examined the relative influence of a set of factors that were associated with LOS in bivariate tests. Two diagnostic variables (personality disorder diagnosis and substance use disorder diagnosis) were predictive of shorter length of stay on the CSU, a unit that is designed specifically for rapid stabilization and return to community care. This finding is in keeping with some prior research. For example, Fisher and colleagues found that discharge diagnoses of substance abuse and adjustment disorder were associated with significantly lower LOS in their sample of 368 adults admitted to general and private psychiatric specialty hospitals under Massachusetts commitment law (12). Other research also has demonstrated that the presence of a personality disorder diagnosis may be associated with a shorter LOS (11,13). Also consistent with previous research (14), the absence of psychosis was predictive of shorter LOS, though the effect of this variable failed to reach statistical significance when adjusting for the other variables in the models for both units.

Several other studies have used a similar analytic method to determine independently significant predictors of LOS. Variables including severity of psychiatric illness, level of psychosocial disability, number of previous hospitalizations, and substance abuse (either as a primary or a secondary diagnosis) have been found repeatedly as predictors, with the latter being a negative predictor of LOS (2,5). In a retrospective chart review from a locked, acute medical/psychiatric 21-bed unit at the Massachusetts General Hospital, Blais and colleagues found a model that accounted for 62% of the variance in LOS (with a significance level set at p≤.10) (2). Legal status and admission GAF were among their 10 significant predictor variables. The basic sociodemographic, diagnostic, and clinical variables assessed in the present review of medical records were able to predict a substantial proportion of the variability in LOS on the CSU and LSMU (30% and 38%, respectively). This is similar to the R 2 values obtained by other research groups, which have ranged from 7% to 57%, though most studies report an R 2 value of 25% or less (3,12,15,16). While the models in this study explained a substantial proportion of the variance, most of the variance in this and other studies is unexplained. This suggests a need to better understand the factors driving these treatment variations.

Concerning the finding that the presence of a substance use disorder was predictive of shorter LOS on both units, this may reflect the fact that those presenting with comorbid substance abuse have less serious conditions and thus require less time for stabilization. However, prior research has demonstrated that substance abuse comorbidity in the context of severe and persistent mental illnesses is associated with more severe illness and disability. Stevens and coworkers suggested that the reasons for this shorter LOS associated with dual-diagnosis may include an interaction effect of some diagnoses, a factor present among the dually-diagnosed that is not being assessed by the research, or discharge after either disorder has remitted, rather than continuing hospitalization until both disorders are remitted (6). Farris and colleagues noted that the finding may be due to: (1) an increased likelihood of individuals with dual-diagnosis leaving the hospital against medical advice, (2) an increased speed of symptom resolution when the effects of acute intoxication have resolved, and/or (3) motivation by dually-diagnosed patients to show speedy improvement so as to obtain release and return to an environment that provides access to their substances of choice (7). Another potential driving force behind this association is the possibility that countertransference issues among clinicians influence discharge decisions for those with substance use disorders (or personality disorders). The treatment course for patients with severe and persistent mental illnesses with comorbid substance use disorders may be negatively affected by shorter lengths of inpatient care, poorer outpatient treatment adherence, and more frequent inpatient admissions (7). Further research is needed on this important issue.

Findings from this investigation should be interpreted in light of several methodological limitations. First, diagnostic and other variables were based on a structured chart review, without the use of diagnostic interviews or rating scales administered directly to patients. However, the purpose of this analysis was to study basic demographic, psychosocial, diagnostic, and clinical determinants of LOS that do not require patient assessment beyond that done as part of routine evaluation and treatment. The implementation of formal rating scales requires a period of special instruction and periodic retraining, which often precludes their use in daily practice (13). The investigative team was particularly interested in readily available variables that could be accessed from the patient's chart and treatment team. Some variables that could impact LOS (e.g., symptom severity, patient insight and motivation) could not be assessed by this method. Second, the generalizability of the specific determinants of LOS may be limited given the characteristics of the samples and, more importantly, given the fact that this research was conducted within a single hospital. Differences in LOS across psychiatric units may be strongly influenced by local economic pressures, regional and institutional variation, case-mix differences, and aspects of the physical environment, including the climate in which the medical center is located and the season during the sampling (17). In a study examining hospital, physician, and patient components of LOS, Harman and colleagues found that hospitals and physicians within hospitals account for nearly 41% of the inter-hospital variation in LOS for schizophrenia (8). Aside from these important limitations, the findings herein demonstrate the ability to predict a substantial portion of the variance in LOS using basic, readily-available characteristics from the medical record during hospitalization. Furthermore, an important strength of this research is the focus on predominantly African American, socially disadvantaged patients with serious mental illnesses treated in a large, urban, county hospital.

LOS has become a widely used indicator of resource utilization. Because of the importance of the balance between further reductions in LOS and maintaining quality of inpatient care, sources of variance in LOS represent an important area for research. Prior analyses with a sample of 221 patients in this dataset indicated that only 36.2% kept their first outpatient community mental health appointment (9). This very low rate of outpatient follow-up adherence, in conjunction with increasing demands to reduce the inpatient LOS, suggests that greater attention must be given to effective discharge planning and bridging strategies (including early planning and attention to disposition, and more inpatient focus on collaboration with families, outpatient clinicians, and social services) designed to avert relapse and rehospitalization (18,19).

Strategies implemented to further reduce inpatient LOS and enhance continuity of care must recognize the possibility that there may be a bedrock of illness which will always need inpatient care despite increased community resources (20). One approach to reducing costs of psychiatric care is to reduce the LOS. Some research suggests, however, that reduction in LOS may increase readmission rates (20,21). Some have argued that the impetus to reduce LOS has gone too far, losing sight of the crucial role of the inpatient setting (22). The findings herein, in addition to other research, may be useful for informing staffing decisions, hospital practice planning, and reimbursement negotiations (19); assessing productivity and comparing different hospital settings (23); as well as providing clinicians with information on potential predictors of LOS in acute psychiatric units.