American Journal of Emergency Medicine (2013) 31, 1-7

Original Contribution

Effect of advanced age and vital signs on admission from an ED observation unit^?,??

Jeffrey M. Caterino MD, MPH ^a,?, Emily M. Hoover ^b, Mark G. Moseley MD ^a

^aDepartment of Emergency Medicine, The Ohio State University, Columbus, OH 43210, USA

^bCollege of Medicine, The Ohio State University, Columbus, OH 43210, USA

Received 8 December 2011; revised 5 January 2012; accepted 5 January 2012

Abstract

Objectives: The primary objective was to determine the relationship between advanced age and need for admission from an emergency department (ED) observation unit. The secondary objective was to determine the relationship between initial ED vital signs and admission.

Methods: We conducted a prospective, observational cohort study of ED patients placed in an ED- based observation unit. Multivariable penalized maximum likelihood logistic regression was used to identify independent predictors of need for hospital admission. Age was examined continuously and at a cutoff of 65 years or more. Vital signs were examined continuously and at commonly accepted cutoffs. We additionally controlled for demographics, comorbid conditions, laboratory values, and observation protocol.

Results: Three hundred patients were enrolled, 12% (n = 35) were 65 years or older, and 11% (n = 33) required admission. Admission rates were 2.9% (95% confidence interval [CI], 0.07%-14.9%) in older adults and 12.1% (95% CI, 8.4%-16.6%) in younger adults. In multivariable analysis, age was not associated with admission (odds ratio [OR], 0.30; 95% CI, 0.05-1.67). Predictors of admission included systolic pressure 180 mm Hg or greater (OR, 4.19; 95% CI, 1.08-16.30), log Charlson comorbidity score (OR, 2.93; 95% CI, 1.57-5.46), and white blood cell count 14 000/mm³ or greater (OR, 11.35; 95% CI, 3.42-37.72).

Conclusions: Among patients placed in an ED observation unit, age 65 years or more is not associated with need for admission. Older adults can successfully be discharged from these units. Systolic pressure 180 mm Hg or greater was the only predictive vital sign. In determining appropriateness of patients selected for an ED observation unit, advanced age should not be an automatic disqualifying criterion.

(C) 2013

Introduction

^? Sources of support: This project was supported in part by a Medical Student Research Scholarship from the Samuel J. Roessler Fund of the Ohio State University College of Medicine.

^?? Prior presentations: NA.

* Corresponding author. Tel.: +1 614 293 8905; fax: +1 614 293 3124.

E-mail address: [email protected] (J.M. Caterino).

Short-stay Observation Units affiliated with hospital emergency departments (EDs) are becoming increasingly common. Their goal is to provide a brief period for diagnostic and therapeutic interventions to aid physicians in making the decision whether to admit or discharge the patient. Disposition of ED patients to these units has been proven to be appropriate for a number of conditions [1-7].

0735-6757/$ - see front matter (C) 2013 http://dx.doi.org/10.1016/j.ajem.2012.01.002

Although a percentage of patients are expected to fail ob- servation status and require hospital admission, patients who ultimately require hospital admission would ideally be identified in the ED rather than after an observation unit stay [8]. Several authors have identified factors, such as comorbid conditions and ED laboratory parameters, which may be predictive of need for admission among observation unit patients [5,9,10]. These studies have generally examined specific observation protocols but are not generalizable to the broad population of patients seen in the observation unit.

One important potential predictor of admission is advanced age. There are greater than 19 million yearly ED visits by adults 65 years or older in the United States, and these older adults are admitted to the hospital at a higher rate than younger patients [11,12]. As a result, there is a need to identify the patterns of care for these patients in an observation unit. One study has found that elders were more likely to be admitted than younger patients from an observation unit (26% vs 18%) [8]. However, this study did not control for comorbidities, protocol type, or other potential predictive factors. Others have examined age as it pertains to specific observation protocols, but not the entire observation unit population [13]. It is therefore unknown if age is independently predictive of admission from an obser- vation unit.

Abnormal vital signs may also represent a differentiat- ing factor to predict the need for admission from an observation unit. Vital signs are part of the Emergency Severity Index triage classification, which has been shown to predict admission in older ED patients [14,15]. However, they have not been well studied in the specific observation unit population.

The goal of this study was to better investigate the relationship between both age and vital signs and the patterns of disposition from an ED observation unit. Our primary objective was to determine the relationship between ad- vanced age and odds of admission from an observation unit. Our secondary objective was to determine if initial ED vital signs are predictive of the need for admission.

Methods

Study design and setting

This was a prospective, observational cohort study conducted in a 20-bed observation unit affiliated with and adjacent to the ED of a tertiary-care level 1 trauma center that sees approximately 72 000 patients yearly. The study took place from July 2010 through March 2011 and was approved by the hospital’s institutional review board. The observation unit is staffed by 3 attending physicians (1 emergency/ internal medicine, 1 emergency medicine, and 1 internal medicine trained) for 8 to 10 hours per day and has around- the-clock nurse practitioner coverage. The unit treats an

average of 475 patients per month on more than 30 clinical protocols. Length of stay ranges from 6 to 24 hours (average of 16 hours), with occasional patients remaining for 24 to 48 hours. Transfer to the observation unit is at the discretion of the ED attending physician with the acceptance of the observation unit provider staff.

Study population and protocol

All patients 18 years or older on an observation unit protocol were potentially eligible to participate. Pregnant women and trauma patients were excluded. A convenience sample of patients was enrolled on shifts when a research assistant was available between 8:00 AM and midnight, 7 days per week. Not every eligible patient was approached during the enrollment window, and not every shift was covered with a research assistant. Upon enrollment, study personnel administered a patient survey gathering patient data such as demographics and medical history. Additional study data were gathered by review of the ED electronic medical record including ED vital signs, ED laboratory values, observation protocol, and ultimate disposition.

The patient survey was administered either by one author (EH) or by undergraduate research assistants. All personnel underwent 1 hour of training on the survey instrument. Research assistants were blinded to study hypotheses. The chart review was performed by one author (EH) who was not blinded to study hypotheses. There were both a standardized survey instrument and standardized abstraction form with predefined variables.

Measurements

The primary outcome variable was disposition, which was defined as either admission to the hospital or discharge from the observation unit. Age was both examined as a continuous variable and divided between adult and older adult (age

>=65 years). The 33 possible observation unit protocols were

initially grouped into 5 Diagnostic categories: cardiac (rule out myocardial infarction, syncope, congestive heart failure, hypertensive urgency, and supraventricular tachycardia), infectious disease (cellulitis, pneumonia, and genitourinary tract infections), pain related (abdominal pain, nephro- lithiasis, and back pain), neurologic (headache, pseudotumor cerebri, transient ischemic attack, and vertigo), and other (dehydration, general observation, pulmonary protocols, endocrine protocols, and other assorted protocols). During data analysis, these 5 categories were combined into 3 categories as described in the “Results” section. Race was divided into white and nonwhite. medical comorbidities were recorded individually, and the Charlson comorbidity (CCM) score was calculated [16]. Immunosuppression was defined as the presence of human immunodeficiency virus/Acquired immunodeficiency syndrome, multiple myeloma, systemic steroid use (past 30 days), malignancy, organ transplant, or current use of immunosuppressive medication.

Initial ED vital signs were examined as continuous variables. They were also dichotomized at cut points chosen from common clinical use. A fever was defined as tem- perature of 38.0?C or greater, tachycardia as a pulse of 100 beats per minute or more, tachypnea as a respiratory rate of 24 breaths per minute or more, and hypotension as a systolic blood pressure of less than 90 mm Hg. Additional potentially useful cutoffs were identified through construc- tion of smoothed plots of each vital sign vs the primary outcome variable.

Recorded laboratory values included white blood cell count, hemoglobin level, platelet count, and creatinine level. Because of the Acuity level of the patients and the variety of clinical conditions present, one or more of these studies were not obtained in many patients, making interpretation as continuous variables problematic. There- fore, laboratory values were dichotomized at common clini- cal cut points. Abnormal values included WBC count of 14 000/mm³ or greater, hemoglobin level of less than 10 g/dL, platelets of less than 150 000/mm³, and serum creatinine level of 2 mg/dL or greater. For purposes of the multivariable analysis and to allow use of the complete data set, patients who did not have particular laboratory tests ordered were assumed to have normal values. This was felt appropriate because laboratory testing in these cases was not felt necessary by the treating clinician and did not influence Disposition decision. The use of multiple impu- tation analysis for laboratory variables was not considered appropriate because of the obviously nonrandom nature of the missing values.

Data analysis

All data analysis was completed using Stata version 11 (STATACorp, College Station, TX). Variables were reported as proportions for dichotomous variables and median with interquartile range for continuous variables. Comparisons were considered significant at P b .05. There was only one outcome event (hospital admission) in the Age 65 years or older group, which was the primary independent variable of interest. The presence of such quasi-complete separa- tion in logistic regression models can result in bias of odds ratio (OR) estimates away from 1. To avoid biased estimates, we analyzed the data using penalized maximum likelihood logistic regression with the Stata command firthlogit [17-19], which reduces bias by penalizing the calculated log likelihood. Significance was tested using the penalized likelihood ratio test [18]. Continuous variables were tested for linearity in the logit using Lowess smoothed plots and fractional polynomial analysis (in unadjusted, nonpena- lized models).

Independent variables were first entered into a series of unadjusted models with disposition as the dependent vari- able. Variables with unadjusted P values less than .20 were then entered into a multivariable model. Age 65 years or older was retained in all models. Variables were examined

for collinearity, and rational choices were made regarding inclusion in the model. An initial full model was created with all candidate independent variables. To address possible sample size concerns, another model was then created by removing selected nonsignificant variables at the P = .05 level through a manual, rational process.

Sensitivity analyses were performed by repeating the multivariable analysis in standard logistic regression and in exact logistic regression. Goodness of fit and discrimination are not available for firthlogit but were calculated in the standard logistic regression models using the Hosmer- Lemeshow test and area under the receiver operating charac- teristic curve (AUROCC), respectively. In the exact model, we substituted the dichotomous variable abnormal CCM score for the continuous CCM score, as required by this technique. We also tested age as a continuous variable in the original models. Finally, we substituted variables initially found to be collinear into the original models.

The initial study sample size of 300 patients was based on an estimated admission rate of 20% with 10% of patients aged 65 years and older. Given that the admission rate was lower than expected, we report Power calculations for the comparison between proportion admitted in the 2 age groups assuming a 10% admission rate and a change in admission proportion from 10% to 20%, which was a priori considered clinically significant. Power was calculated using a 2-sided test with ? = .05.

Results

A total of 396 screened subjects met enrollment criteria and were approached for consent. Ninety-six declined to participate, leaving 300 subjects enrolled. The characteristics of the study subjects are demonstrated in Table 1. Twelve percent (n = 35) of the study population were elders 65 years or older. Eleven percent (n = 33) of the population were admitted, including 32 (12.1%) of 265 patients younger than 65 years and 1 (2.9%) of 35 elders. Older adults had a greater Comorbidity burden and were more likely to be on a cardiac protocol than younger adults. The only missing data were the lack of any recorded temperatures in 3 patients. They were assigned the study average temperature.

Because of the lower-than-expected admission rate, the Power of the study to detect a difference in proportion admitted in each age group was lower than originally planned. We used the number of subjects and the proportions admitted in each age group in our study sample for the power calculations. The power to detect a significant difference between age groups in the admission rates we obtained in the study (2.9% and 12.1%) was only 0.15. However, the study had adequate power to detect clinically significant increases in admission among older adults. Power was 0.90 to detect a difference in proportion admitted among older adults from a hypothesized 20% compared with the 3% we found. In addition, power was 0.81 to detect an increase

Table 1 Characteristics of 300 subjects in an ED observation unit

Variable Entire population (n = 300) Age b65 y (n = 265) Age >=65 y (n = 35)

Frequency (n)

Percentage or median (IQ range)

Frequency (n)

Percentage or median (IQ range)

Frequency (n)

Percentage or median (IQ range)

Demographics
Age (y) (continuous)	-	44 (32-54)	-	42 (31-51)	-	68 (66-75)
Age >=65 y	35	12%	NA	NA	NA	NA
Female sex	185	62%	161	61%	24	69%
Nonwhite race	78	26%	71	27%	7	20%
Hispanic ethnicity	3	1%	3	1.1%	0	0%
Nursing home resident	0	0%	0	0%	0	0%
Comorbid conditions
CCM score (continuous)	-	1 (0-2)
CCM score >=4	44	15%	33	12%	11	31%
Diabetes mellitus	65	22%	55	21%	10	29%
Immunosuppression	80	27%	68	26%	12	34%
Initial vital signs
Temperature (?C), median (IQ range)	-	36.7 (36.6-36.9)	-	36.7 (36.6-36.9)	-	36.7 (36.5-37.0)
Pulse, median (IQ range)	-	85 (75-96)	-	86 (76-98)	-	78 (70-88)
Respiratory rate, median (IQ range)	-	16 (16-18)	-	16 (16-18)	-	16 (16-18)
Systolic blood pressure, median (IQ range)	-	138 (123-154)	-	136 (122-154)	-	147 (129-158)
Temperature >=38.0?C	12	4%	9	3%	0	0%
Tachycardia (>=100 beats per minute)	62	21%	60	23%	2	6%
Tachypnea (>=24 breaths per minute)	6	2%	6	2.3%	0	0%
Systolic blood pressure >=180 mm Hg	15	5%	14	5%	1	3%
Hypotension (b90 mm Hg)	2	1%	2	1%	0	0%
Laboratory values
WBC count >=14 000/mm3	18	6%	17	6%	1	3%
Hemoglobin level b10 g/dL	9	3%	8	3%	1	3%
Platelets b150 000/mm3	12	4%	10	4%	2	6%
Creatinine level >=2 mg/dL	4	1%	3	1%	1	3%
Observation protocol
Cardiac	118	39%	99	37%	19	54%
Infectious disease	47	16%	43	16%	4	11%
Neurologic	28	9%	23	9%	5	14%
Pain	48	16%	47	18%	1	3%
Other	59	20%	53	20%	6	17%
Disposition
Admission	33	11%	32	12%	1	3%
IQ indicates interquartile; NA, not applicable.

in proportion admitted from 10% in younger adults to 30% in older adults.

Results of the unadjusted analyses are shown in Table 2. After initial examination of the ORs for protocol type, we combined cardiac (OR, 1.00; 95% confidence interval [CI], 0.35-2.81) (P = 1.000) and neurologic (OR, 1.06; 95% CI,

0.24-4.59) (P = .938) protocols with the referent “other” protocol group. Examination of the smoothed plot of sys- tolic blood pressure revealed Increased admission rates both at the predefined cutoff of less than 90 mm Hg and at an additional cutoff of 180 mm Hg or greater.

Variables with univariate P values less than .200 included age 65 years or older, protocol category, CCM score, diabetes mellitus, immunosuppression, initial temperature, presence of fever, systolic blood pressure of 180 mm Hg or greater,

and elevated WBC count. In the multivariable models, the variables measuring medical history-CCM score, diabetes mellitus, and immunosuppression-were collinear. To achieve a complete overview of medical comorbidity, we retained CCM score for our primary multivariable models. However, in the fractional polynomial analysis, CCM score was not linear in the logit; and it was therefore log transformed. Collinearity was also found between the continuous temperature and the dichotomous fever variable, so fever was not tested in the primary model.

An initial full multivariable model was created from these identified predictors (Table 3). Age 65 years or older trended toward a decreased odds of admission (OR, 0.30) in this model but was not significant. Because the inclusion of 7 variables for 34 outcome events in the full model is at the

Variable	OR	P a	Wald 95% CI b
Demographics Age >=65 y Female Nonwhite Comorbid conditions CCM score Log of CCM score Diabetes Cancer Immunosuppression Initial vital signs Temperature Pulse Respiratory rate Systolic blood pressure Temperature >=38.0?C Tachycardia (>=100 beats per minute) Tachypnea (>=24 breaths per minute) Systolic blood pressure b90 Systolic blood pressure >=180 Laboratory values WBC count >=14,000/mm3 Hemoglobin level b10 g/dL Creatinine level >=2 mg/dL Platelets b150 000/mm3 Observation protocol Cardiac/neurologic/other Infectious disease Pain	0.31	.109	0.06-1.67
	1.08	.834	0.51-2.27
	1.30	.513	0.60-2.83
	1.21	.028	1.03-1.42
	2.09	.006	1.23-3.55
	2.70	.012	1.28-5.72
	1.31	.602	0.49-3.48
	1.46	.330	0.68-3.12
	1.52	.029	1.09-2.13
	1.00	.648	0.98-1.02
	1.05	.439	0.92-1.21
	1.01	.351	0.99-1.02
	4.92	.036	1.27-19.09
	1.30	.536	0.57-3.00
	2.20	.435	0.35-13.90
	1.58	.779	0.07-33.72
	3.40	.056	1.07-10.79
	8.17	b.001	3.03-22.03
	2.76	.216	0.63-12.09
	3.49	.253	0.50-24.44
	0.30	.326	0.018-5.27
	Referent	.033
	2.40		1.06-5.44
	0.46		0.12-1.78
a P values are based on the penalized likelihood ratio test. b Confidence intervals are based on Wald intervals, which may be inaccurate for variables with small numbers of outcome events.

limit of recommended events per variable in multivariable analysis [20], a reduced model was created (Table 3). Age 65 years or older continued to show a trend toward reduced odds of admission (OR, 0.28), but remained nonsignificant. Results for the remaining variables in this reduced model were similar to the full model. There were no interactions in either model.

Table 2 Univariate predictors of admission among ED observation unit patients using penalized log likelihood logistic regression

In the sensitivity analyses, the same variables were sig- nificant in both standard and exact logistic regression as in the primary analysis. Age 65 years or older remained non- significant in both standard (OR, 0.19; 95% CI, 0.02-1.51)

(P = .117) and exact (OR, 0.21; 95% CI, 0.00-1.43) (P =

.1672) multivariable logistic regression. There was no evidence of lack of fit of either the full (P = .1470) or reduced (P = .9295) models in standard logistic regression. Discrimination was good for full (AUROCC, 0.79) and reduced (AUROCC, 0.77) models. In the initial penalized

models, when substituting age as a continuous variable, it remained nonsignificant (OR, 0.98; 95% CI, 0.95-1.01) (P =

.200). Among the collinear variables, when substituting for the logCCM score, diabetes mellitus was significant (OR, 3.06; 95% CI, 1.36-6.87) (P = .008); but immunosuppression was not (OR, 1.89; 95% CI, 0.84-4.28) (P = .129).

Discussion

We found that only 2.9% (95% CI, 0.07%-14.9%) of older adults were admitted from an ED observation unit as compared with 12.1% (95% CI, 8.4%-16.6%) of adults. This finding supports the contention that the ED observa- tion unit is an appropriate environment for placing selected older adults.

Results of prior studies in this area have been mixed. In a large cohort of observation unit patients, Ross et al [8] found in 2003 that elders actually had higher rates of admission than younger adults (26.1% vs 18.5%). They did, however, conclude that elders were appropriately placed in the unit because they had an admission rate lower than a predefined 30% threshold. Although they considered protocol type, they did not adjust for other demographics, medical comorbid- ities, vital signs, or laboratory abnormalities. We therefore felt it appropriate to expand the analysis controlling for these variables. Our overall admission rate was lower than that in Ross et al, which may point to differences in patient selection policies, maturation of the observation unit concept, or other unmeasured differences in practice between the 2 sites.

Additional prior studies have supported our conclusions. Chan et al [21] studied all patients placed in an ED obser- vation unit in Australia and found that there was no rela- tionship between age and admission after controlling for multiple confounders. However, Chan et al examined age only as a continuous variable; and the distribution of ages in the study population was not reported. In studies limited to specific observation unit protocols, advanced age has not been associated with increased odds of admission [6,9,13].

Sample size and power to detect a difference are a concern in our study. The initial power calculation assumed an overall admission rate of 20%. Given the study data, our concern is to avoid type II error, that is, failing to detect a difference where one exists. Most clinically relevant would be failure to detect an increased rate of admission among older adults. We think that an increased rate of admission among older adults is unlikely for several reasons. First, the point estimate and CIs for admission of older adults were 2.9% (95% CI, 0.07%-14.9%) compared with 12.1% (95%

CI, 8.4%-16.6%) in adults. We therefore are 95% confident that the unadjusted admission rate in our population of older adults is less than 15%. Second, although the power to detect a difference in these point estimates was low, the power was adequate to detect a clinically significant difference from a 10% admission rate in younger adults to a 30% admission rate in older adults. Finally, the OR in the adjusted analysis,

OR P a 95% CI b					OR P a 95% CI b
Age >=65 y	0.30	.111	0.05-1.67	0.28		.091	0.05-1.57
Log of CCM score	2.93	b.001	1.57-5.46	2.91		b.001	1.59-5.30
Initial temperature	1.17	.503	0.64-2.11	-		-	-
Systolic blood pressure >=180 mm Hg	4.19	.058	1.08-16.30	5.20		.016	1.52-17.79
WBC count >=14 000/mm3	11.35	b.001	3.42-37.72	14.00		b.001	4.48-43.44
Observation protocol		.254				-
Infectious disease protocol	1.71		0.64-4.55	-		-	-
pain protocol	0.50		0.12-2.14	-		-	-
a P values are based on the penalized likelihood ratio test. b Confidence intervals are based on Wald intervals, which may be inaccurate for variables with small numbers of outcome events.

although nonsignificant and with CIs encompassing 1, trended toward decreased rather than increased odds of admission for older adults.

Table 3 Multivariable analysis of predictors of admission among ED observation unit patients using penalized log likelihood logistic regression

Variable Full model Reduced model

Because selection of patients for the observation unit in our study was at the discretion of the ED and observation unit staff, it could be argued that only less ill or uncomplicated elders were placed in the unit. Assessment of patient severity is complicated by the large variety of observation unit protocols; but possible measures of severity include Vital sign abnormalities, comorbidity burden, and laboratory abnor- malities. In our cohort, abnormal vital signs and laboratory values were similar between the 2 age groups. The excep- tion was tachycardia, which was more common in younger adults. However, elders had a much greater overall comor- bidity burden than younger adults. Overall, these facts sug- gest that elders in our observation unit had similar severity of illness to younger adults.

An additional difference between elders and younger adults was in distribution of protocol type. These differences may be due to different proportions of Presenting complaints in different age groups, differences in risk stratification for these presenting complaints (eg, younger adults with chest pain are less likely to undergo a cardiac workup), or differ- ences in admission patterns based on age (eg, older adults with infection may be more likely to be admitted). We are unable to determine which of these factors, if any, affected initial decision to place in the observation unit. However, we did adjust for protocol type in the multivariable analyses to control for potential confounding.

Abnormal vital signs might also predict admission from the observation unit but have not previously been well studied. We found that initial temperature, heart rate, and respiratory rate were not predictive of admission. The lack of a relationship is not surprising because patients with severely abnormal vital signs would likely be admitted rather than placed in an observation unit.

Systolic blood pressure was also not predictive of ad- mission when examined continuously. However, graphical analysis identified a cut point at blood pressures of 180 mm Hg or greater where admission rates increased. Although

constituting only 5% of the population, 26.7% of the 15 patients with systolic blood pressures of 180 mm Hg or greater were admitted, more than double the overall cohort rate. As a result, severely elevated systolic blood pressure could be considered as a relative contraindication to place- ment into an observation unit. Others have not identified this cut point at 180 mm Hg or greater, likely because of confining their analysis of blood pressure to a continuous variable [9].

We also examined protocol type, comorbidities, and laboratory studies as potential predictors of admission. Protocol type was not associated with admission, a finding consistent with the findings of Chan et al [21]. Comor- bidity burden and elevated WBC count were associated with admission. A larger sample would be required to determine what specific comorbidities are most contributory.

The study was limited by the fact that we only enrolled patients after initial transfer to the observation unit. This study therefore cannot be used as a guide for initial dis- position decision of patients during their ED stay. Rather, it demonstrates characteristics of patients who are likely to fail observation care after being placed in Observation status. We have a large, well-established observation unit; and the patterns seen in this single-center study may not apply to other settings with different observation referral patterns. It would be appropriate to study the ED population directly to account for potential differentiation of patients selected for the observation unit based on age.

The admission rate in the study cohort was 11%, but our overall observation unit admission rate is closer to 20%. There are several potential explanations for this finding. First, a proportion of patients are placed in observation status; but upon initial evaluation by observation unit staff, they are felt to be more appropriate for admission. Although such patients are refused for observation level of care and immediately admitted, because of the mechanics of our electronic medical record, they appear among the observation unit population, increasing its admission rate. It may also be that patients were less willing to participate if they were more ill. Finally, the shift distribution of study personnel may have resulted in

missing patients admitted when the observation unit physician first came on his or her shift at 7:00 AM. It is unclear what effect including these patients may have had on study outcomes. We were able to confirm that we did not preferentially enroll patients on specific protocols because our distribution of protocols mirrored that of the observa- tion unit overall for the study period.

One consequence of the lower admission rate was a less- than-expected number of outcome events. We have already discussed concerns over study power above. Because of the quasi-complete separation for the age variable, we used penalized logistic regression, which prevents bias, but may be overly conservative [18,19]. The Wald CIs provided in Tables 2 and 3 are the only ones available in Stata using the firthlogit command. These may be inaccurate for variables with small numbers of outcome events (eg, age >=65 years) because their likelihood profile may be asymmetric. The CIs for age should therefore be interpreted with caution. However, the penalized likelihood ratio tests’ P values are valid measures; and conclusions regarding significance of age 65 years or older are considered valid. Although there are less than the traditional 10 events per independent vari- able in the models, recent work indicates that 5 to 9 events per variable provides similar error rates and bias in logistic regression as 10 events per variable [20]. To confirm our findings, we also constructed the reduced model. Finally, the assumption that results of laboratory studies, which were not ordered, were normal, although reasonable to make, may have attenuated the effect of abnormal values on admission. As a result, we would exercise caution in concluding that laboratory parameters were not associated with admission.

We conclude that, among patients placed in an ED ob-

servation unit, older adults are admitted 2.9% (95% CI, 0.07%-14.9%) of the time as compared with an admission rate of 12.1% (95% CI, 8.4%-16.6%) of adults. In both adjusted and unadjusted analysis, advanced age was not significantly associated with admission and demonstrated a trend toward decreased odds of admission. The study was underpowered to detect a difference in these analyses, however. Older adults can successfully be cared for in these units. Initial temperature, respiratory rate, and pulse were not predictive of admission; but extremely elevated blood pres- sure was predictive.

References

1. Stead LG, Bellolio MF, Suravaram S, et al. Evaluation of transient ischemic attack in an emergency department observation unit. Neurocrit Care 2009;10:204-8.
2. Decker WW, Smars PA, Vaidyanathan L, et al. A prospective, ran- domized trial of an emergency department observation unit for acute onset atrial fibrillation. Ann Emerg Med 2008;52:322-8.
3. Ross MA, Compton S, Medado P, et al. An emergency department Diagnostic protocol for patients with transient ischemic attack: a randomized controlled trial. Ann Emerg Med 2007;50:109-19.
4. Bledsoe J, Hamilton D, Bess E, et al. Treatment of low-risk pulmonary embolism patients in a chest pain unit. Crit Pathw Cardiol 2010;9: 212-5.
5. Sabbaj A, Jensen B, Browning MA, et al. soft tissue infections and emergency department disposition: predicting the need for inpatient admission. Acad Emerg Med 2009;16:1290-7.
6. Schrock JW, Laskey S, Cydulka RK. Predicting observation unit treatment failures in patients with skin and soft tissue infections. Int J Emerg Med 2008;1:85-90.
7. Schrock JW, Reznikova S, Weller S. The effect of an observation unit on the rate of ED admission and discharge for pyelonephritis. Am J Emerg Med 2010;28:682-8.
8. Ross MA, Compton S, Richardson D, et al. The use and effectiveness of an emergency department observation unit for elderly patients. Ann Emerg Med 2003;41:668-77.
9. Burkhardt J, Peacock WF, Emerman CL. Predictors of emergency department observation unit outcomes. Acad Emerg Med 2005;12: 869-74.
10. Madsen T, Bossart P, Bledsoe J, et al. Patients with coronary disease fail observation status at higher rates than patients without coronary disease. Am J Emerg Med 2010;28:19-22.
11. LaMantia MA, Platts-Mills TF, Biese K, et al. Predicting hospital admission and returns to the emergency department for elderly patients. Acad Emerg Med 2010;17:252-9.
12. Burstein JL, Hollander JE, Henry MC, et al. Association of out-of- hospital criteria with need for hospital admission. Acad Emerg Med 1995;2:863-6.
13. Madsen TE, Bledsoe J, Bossart P. Appropriately screened geriatric chest pain patients in an observation unit are not admitted at a higher rate than nonGeriatric patients. Crit Pathw Cardiol 2008;7:245-7.
14. Baumann MR, Strout TD. Triage of geriatric patients in the emergency department: validity and survival with the Emergency Severity Index. Ann Emerg Med 2007;49:234-40.
15. Platts-Mills TF, Travers D, Biese K, et al. Accuracy of the Emergency Severity Index Triage Instrument for identifying elder emergency department patients receiving an immediate life-saving intervention. Acad Emerg Med 2010;17:238-43.
16. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: develop- ment and validation. J Chronic Dis 1987;40:373-83.
17. Firth D. Bias reduction of maximum likelihood estimates. Biometrika 1993;80:27-38 [Ref Type: Generic].
18. Heinze G, Schemper M. A solution to the problem of separation in logistic regression. Stat Med 2002;21:2409-19.
19. Heinze G. A comparative investigation of methods for logistic regres- sion with separated or nearly separated data. Stat Med 2006;25: 4216-26.
20. Vittinghoff E, McCulloch CE. Relaxing the rule of ten events per variable in logistic and Cox regression. Am J Epidemiol 2007;165: 710-8.
21. Chan T, Arendts G, Stevens M. Variables that predict admission to hospital from an emergency department observation unit. Emerg Med Australas 2008;20:216-20.