The effect of pneumonia on mortality among patients hospitalized for acute stroke

I.L. Katzan, MD, MS; R.D. Cebul, MD; S.H. Husak, BA; N.V. Dawson, MD; and D.W. Baker, MD, MPH

AbstractObjective: To determine the effect of pneumonia on 30-day mortality in patients hospitalized for acute stroke. Methods: Subjects in the initial cohort were 14,293 Medicare patients admitted for stroke to 29 greater Cleveland hospitals between 1991 and 1997. The relative risk (RR) of pneumonia for 30-day mortality was determined in a final cohort (n 11,286) that excluded patients dying or having a do not resuscitate order within 3 days of admission. Clinical data were obtained from chart abstraction and were merged with Medicare Provider Analysis and Review files to obtain deaths within 30 days. A predicted-mortality model (c-statistic 0.78) and propensity score for pneumonia (c-statistic 0.83) were used for risk adjustment in logistic regression analyses. Results: Pneumonia was identified in 6.9% (n 985) of all patients and in 5.6% (n 635) of the final cohort. The rates of pneumonia were higher in patients with greater stroke severity and features indicating general frailty. Unadjusted 30-day mortality rates were six times higher for patients with pneumonia than for those without (26.9% vs 4.4%, p 0.001). After adjusting for admission severity and propensity for pneumonia, RR of pneumonia for 30-day death was 2.99 (95% CI 2.44 to 3.66), and population attributable risk was 10.0%. Conclusion: In this large community-wide study of stroke outcomes, pneumonia conferred a threefold increased risk of 30-day death, adding impetus to efforts to identify and reduce the risk of pneumonia in patients with stroke.
NEUROLOGY 2003;60:620 625

In addition to recurrent stroke, myocardial infarction, and death, stroke places individuals at increased risk for nonvascular medical sequelae or complications. Over half of all readmissions following stroke are for nonvascular diagnoses.1 Many of these are related to the neurologic deficits and increased frailty associated with stroke. Pneumonia is among the most common2 and has a reported incidence of 1.5 to 13.0%3 after acute stroke. However, many of the studies on the incidence of pneumonia after stroke were based on cohorts derived from inpatient rehabilitation units or were from other countries, making it difficult to extrapolate to the typical US hospital setting.4-7 Pneumonia is thought to occur most often as a result of oropharyngeal dysphagia with secondary aspiration. It is estimated that 42 to 76% of patients with acute stroke develop dysphagia. Of these, about half will experience aspiration; however, not all develop pneumonia.3 In addition, up to half of patients who develop pneumonia do not aspirate.8 Little information is available on which specific factors increase
Additional material related to this article can be found on the Neurology Web site. Go to www.neurology.org and scroll down the Table of Contents for the February 25 issue to find the title link for this article.

the risk for pneumonia, although it appears to be more common among older patients and among those with more severe strokes.6,7 Even less information is available regarding the outcomes of pneumonia. Current data and estimates are severely restricted, derived principally from general neurologic cohorts not specific to the stroke population9 or from very limited cohorts of patients with stroke.3,10 Additional information on predictors and associated outcomes will create opportunities to identify those at greatest risk for pneumonia and to institute preventive therapies. To address this problem, we used data from the Cleveland Health Quality Choice Program (CHQC), a region-wide quality measurement initiative including all 29 nonfederal hospitals in northeast Ohio. This project included all Medicare patients admitted for acute stroke, including detailed clinical information at the time of admission, as well as hospital outcomes.11 Our goal was to determine the incidence of pneumonia as well as the independent effect of pneumonia on 30-day mortality.
Methods. Hospitals and procedures in the CHQC. The hospital sample includes all 29 nonfederal hospitals in the four-county greater Cleveland area that participated in the CHQC from January 1991 through December 1997. Ten of the hospitals had accred-

From the Center for Health Care Research & Policy (Drs. Katzan, Cebul, Dawson, and Baker, and S.H. Husak) and Departments of Medicine and Epidemiology and Biostatistics (Drs. Cebul, Dawson, and Baker), Case Western Reserve University at MetroHealth Medical Center, Cleveland; and Department of Neurology (Dr. Katzan), Cleveland Clinic Foundation, OH. Supported by Agency for Healthcare Research and Quality RO1-HS09969 (D.W.B.) and T32-HS00059 (I.L.K.). Received February 25, 2002. Accepted in final form October 25, 2002. Address correspondence and reprint requests to Dr. Irene L. Katzan, Center for Health Care Research and Policy, MetroHealth Medical Center, 2500 MetroHealth Drive, 237A, Cleveland, OH 44109-1998; e-mail: ikatzan@metrohealth.org 620 Copyright 2003 by AAN Enterprises, Inc.

ited residency programs (eight medical, two osteopathic), and five were members of the Council of Teaching Hospitals (COTH) of the Association of American Medical Colleges. Mean hospital bed size was 307 (median, 272; range, 97 to 1,014), and nine hospitals had 300 or more beds. Every 6 months, each hospital provided CHQC with a list of all patients who were discharged from their facility with one of six medical conditions as a principal diagnosis, including stroke. Data were abstracted from patients hospital records by CHQC-trained medical records technicians using standard data forms. Several procedures were established to ensure data reliability,11,12 including explicit definitions for each variable, double-keystroke data entry, and independent evaluation of reliability of data abstraction at each hospital by CHQC chart abstractors. The standard database included patient demographics; admission source (e.g., home, nursing home); presence and type of insurance; comorbid conditions; admission vital signs and neurologic findings; results of laboratory, radiologic (including, for stroke, brain imaging), and electrocardiographic testing; dates of do not resuscitate (DNR) orders; and discharge status and disposition. All International Classification of Diseases, 9th revision, clinical modification (ICD-9-CM) diagnosis and procedure codes were collected on each patient at the time of discharge. No specific informed consent statement was signed and no patient identifying data were produced. Patients and cohort definitions. The study was approved by the Institutional Review Board at MetroHealth Medical Center. Study subjects included all patients with Medicare as the primary insurer who were admitted for an ischemic (ICD-9-CM 434, 436) or intraparenchymal hemorrhagic stroke (ICD-9-CM 431, 997.02). Patients with ICD-9-CM 433 (occlusion/stenosis of precerebral artery) were excluded from this analysis because of the low specificity associated with this code.13,14 Only the first hospital admission with principal discharge diagnosis of stroke during the time period of 1991 to 1997 was used for this analysis. We defined two cohorts for the purposes of this investigation. The initial cohort included all patients as described above. In the final cohort, we excluded patients who died within 3 days of admission and those who had DNR orders written in the medical record within their first 3 days of hospitalization. This was done to limit our analyses to those patients for whom intensive efforts at pneumonia prevention would be clearly desirable and for whom aggressive treatment would be clearly appropriate for patient management. Deaths occurring within 3 days of admission are more likely to be due to the direct effect of the neurologic event rather than pneumonia.15 Similarly, documented DNR orders within 3 days of admission may have been an indication that patients or their loved ones may have preferred not to have aggressive interventions undertaken, including those designed to prevent pneumonia. Data and measurements. Merging CHQC and Medicare files. The CHQC focused its quality measurement activities on inhospital mortality and length of stay and did not collect patientspecific identifiers. Therefore, in order to identify patient-level 30-day mortality, we merged data from CHQC with the Medicare Provider Analysis and Review (MEDPAR) files. Six fields common to both CHQC and Center for Medicare and Medicaid Services (formerly Health Care Financing Administration) files were used to link CHQC stroke admissions to their corresponding admission in MEDPAR: hospital identifier, discharge date, principal diagnosis code, sex, birth month, and birth year. Of all CHQC admissions for stroke, 88% were for patients with Medicare as the insurer. Matching between MEDPAR and CHQC was obtained for 89% of the CHQC Medicare patients. Nine patients were excluded owing to missing data on critical demographic features (age, sex, or race). Missing data on most other variables were rare (less than 1%). Based on a priori clinical assumptions and exploratory analyses, we either assigned normal results to these missing values or imputed the results using the S-Plus version 4 Transcan program (MathSoft; Seattle, WA). Details on procedures for missing data are reported elsewhere16 and described further in Appendix E-1 on the Neurology Web site (access www.neurology.org). Main independent variable: pneumonia. Pneumonias occurring during the index hospital admission were identified using secondary ICD-9-CM codes recorded in the CHQC database (507.0 507.8, 481.xx 482.xx, 485.xx, 486.xx, 997.3). Covariate: admission severity of illness. To adjust for baseline differences in risk of death between patients with stroke who

developed or did not develop pneumonia, we developed a covariate to reflect admission severity of illness. We adapted the predicted hospital mortality model for stroke developed and validated by CHQC17 using vital status (dead or alive) at 30 days as the dependent variable in a logistic regression analysis. Variables in the predicted mortality model were age, blood urea nitrogen, highest glucose within 48 hours of admission, pH, initial platelet count, arterial pCO2 level, initial systolic blood pressure, arrival within 24 hours of stroke onset, the presence of atrial fibrillation, highest temperature within the first 48 hours, coma, lethargy or stupor, metastatic cancer, white blood cell count, presence of cardiac arrest before arrival, initial hematocrit, presence of paralysis, admission to the hospital from a nursing home, and nonreactive pupils. The predicted probability obtained from this model was entered as a covariate in the main analyses. We then examined its performance in the final cohort that excluded stroke patients who died or had DNR orders written within the first 3 days of admission. For each patient in the final cohort, the model calculated the risk of dying (0.0 to 1.0) by 30 days after admission. The c-statistic of the severity of illness model was 0.78. The c-statistic can be interpreted as the area under the receiver operating characteristic (ROC) curve, where a value of 0.50 indicates no discrimination and a value of 1.0 indicates perfect discrimination. Covariate: propensity for pneumonia. In order to better isolate the effect of pneumonia from the effect of features of patients who are more likely to develop pneumonia, we developed a covariate that represented patient propensity for pneumonia. The approach of propensity analysis facilitates the balancing of factors that are likely to differ among patients who either develop pneumonia or do notfactors such as being admitted from a nursing home, or having poor nutritional status, that could confound the relationship of pneumonia to mortality. In propensity analysis, a collection of predictors is collapsed into a single summarizing predictor, the propensity score, using logistic regression.18 For each patient in the final cohort, the regression model calculates the probability (0.0 to 1.0) of developing pneumonia during the hospital stay, using data available around the time of admission. To develop a propensity score for pneumonia, we assembled a focus group (three vascular neurologists and an internist) to identify factors posited to be associated with the development of pneumonia. Bivariate analyses of these factors with pneumonia were initially performed. Continuous variables were assessed graphically and were dichotomized at various cutoffs to try to elucidate nonlinear associations. All 38 factors identified by the focus group were then placed into a logistic regression model with pneumonia as the dependent variable. Adequacy of the propensity score in adjusting for the effect of covariates on the development of pneumonia was evaluated by matching patients with and without pneumonia on propensity score. To create the match, a patient with stroke was randomly chosen from the patients with pneumonia in the final cohort. That patient was then matched with a patient without pneumonia having the closest propensity score within 0.02.19 The baseline characteristics between the matched pairs were then assessed for how well propensity scoring equalized the distribution of covariates. Overall discrimination of the model also was assessed by the c-statistic. Statistical methods. The incidence of pneumonia occurring during the hospital admission was calculated for both the initial and final cohorts. Bivariate analyses examined the association between pneumonia and admission characteristics of patients in the final cohort, and relative risks (RR) were calculated along with 95% confidence intervals. We examined the effect of pneumonia on 30-day mortality, both without adjustment and using logistic regression models that successively added covariates accounting for patient demographics (age, sex, race) and COTH status of the patients hospital, severity of illness, and propensity for pneumonia. We then estimated the effect of pneumonia in various subgroups in the final cohort, including stroke subtype, sex, race, older age, admission from a nursing home, prior stroke/TIA, and chronic obstructive pulmonary disease/asthma. This was done with stratification, using the same regression model for all subgroups.19 Odds ratios (OR) and 95% CI were calculated and converted to RR using published formula.20 To further evaluate the impact of pneumonia on mortality, population attributable risk (PAR) was calculated using the forprevalence*(RR 1)/[prevalence*(RR 1) 1]. mula21,22 PAR
February (2 of 2) 2003 NEUROLOGY 60 621

Table 1 Incidence of pneumonia and unadjusted 30-day mortality Cohort Initial cohort Early death or do not resuscitate, days 13 Final cohort* N 14,293 2,998 11,286 Pneumonia, n (%) 986 (6.9) 351 (11.7) 635 (5.6) 30-Day mortality, n (%) 2,085 (14.6) 1,448 (48.3) 644 (5.7)

* Excluding patients with early death or do not resuscitate orders and nine additional patients who had missing data on key covariates. This is equivalent to the percentage of deaths in patients hospitalized for stroke that is attributable to pneumonia. Prevalence refers to the proportion of patients with pneumonia (0.056). The number of cases of pneumonia avoided to prevent one death was also calculated using the following formula: number needed to treat 1/[(RR 1)*prevalence] (http://www.library.utoronto.ca/medicine/ ebm/practice/ca). Prevalence in this latter instance refers to the proportion of patients without pneumonia who died (0.044). Finally, we determined the RR of 90-day death in patients who developed pneumonia during their hospitalization after adjusting for the same covariates used in the main analysis. Results. There were 14,293 patients admitted to Cleveland area hospitals with the primary discharge diagnosis of stroke between 1991 and 1997. Pneumonia was identified in 6.9% of patients and Table 2 Characteristics of patients with and without pneumonia Crude relative risk (95% CI) of developing pneumonia if characteristic present p Value 0.007 0.55 (0.470.64) 0.87 (0.721.04) 2.0 (1.642.57) 0.001 0.12 0.001 30-day mortality was 14.6% (table 1). Of the initial cohort, 21% died or were made DNR within the first 3 hospital days. The incidence of pneumonia in this group was 11.7% (see table 1). This group was excluded from the final cohort in order to focus on patients in whom the prevention and treatment of pneumonia were most relevant. An additional nine patients were excluded because critical demographic data were missing, leaving 11,286 in the final cohort. The majority of identified cases were coded as aspiration pneumonia (40.7%, ICD-9-CM 507.0 507.8) or pneumonia, organism unspecified (37%, ICD-9-CM 486.xx). At baseline, there were significant differences between patients with and without pneumonia in almost all measured characteristics (table 2). Men, patients admitted from nursing homes, and those with more comorbid illnesses had higher rates of pneumonia. In addition, patients with more severe neurologic impair-

Characteristics Age, mean, y Female White Admission from nursing home Comorbidities History of COPD or asthma History of stroke or transient ischemic attack Congestive heart failure Neurologic findings Intracranial hemorrhage Lethargy or stupor Coma Paralysis Nonreactive pupils Vital signs/laboratory values, mean Initial SBP, mm Hg Creatinine mg/dL mol/L Albumin, g/L Highest glucose mg/dL mmol/L

Pneumonia patients, n 635* 77.6 [71.8, 77.7, 83.6] 266 (41.9) 508 (80) 79 (12.4)

No pneumonia, n 10,651* 76.7 [71.5, 76.8, 82.3] 6,139 (57.6) 8,777 (82.4) 652 (6.1)

119 (18.7) 212 (33.4) 183 (28.8)

1,272 (11.9) 3,134 (29.4) 1,202 (11.3)

1.6 (1.31.98) 1.19 (1.011.40) 2.9 (2.463.40)

0.001 0.034 0.001

124 (19.5) 166 (26.1) 65 (10.2) 128 (20.2) 36 (5.7)

807 (7.6) 1,311 (12.3) 154 (1.4) 841 (7.9) 171 (1.6)

2.7 (2.243.24) 2.3 (1.982.78) 5.8 (4.637.18) 2.69 (2.243.23) 3.22 (2.374.38)

0.001 0.001 0.001 0.001 0.001

160.7 [138, 158, 182]

163.3 [142, 160, 182]

0.06

1.4 [1, 1.2, 1.5] 127.3 [88.4, 106.1, 132.6] 36.4 [33, 38, 40]

1.3 [0.9, 1.1, 1.4] 114.9 [79.6, 97.2, 123.8] 37.8 [35, 39, 40]

0.001

0.0001

185.8 [119, 150, 21] 10.3 [6.6, 8.3, 1.2]

156.6 [107, 129, 176] 8.7 [5.9, 7.2, 9.8]

0.0001

* Values in parenthesis represent percentages; straight brackets represent quartiles [25th, 50th, 75th]. Identified using secondary International Classification of Diseases, 9th revision, clinical modification (ICD-9-CM) codes. COPD
622

chronic obstructive pulmonary disease; SBP

February (2 of 2) 2003

systolic blood pressure.

NEUROLOGY 60

Table 3 Effect of risk adjustment on estimated impact of pneumonia on mortality Relative risk (95% CI) of 30-day death 5.9 (5.16.8) 4.4 (3.72.5) 3.0 (2.43.7) Population attributable risk, % 21.5 16.1 10.0 No. prevented pneumonia cases needed to avoid one death 4.6 6.6 11.4

Model* A B C

Adjustors in model Basic demographics Demographics severity

C-statistic 0.671 0.769 0.800

Demographics severity propensity for pneumonia

* Dependent variable 30-day mortality; independent variable pneumonia. Relative risk for 30-day death in patients with pneumonia. Demographics include age, sex, race, hospital teaching status (member of Council of Teaching Hospitals). Severity of illness predicted probability of 30-day mortality (see text for details).

ments at the time of admission, who had sustained a cardiac arrest before admission, and who had abnormal laboratory values were more likely to develop pneumonia. Pneumonia was identified in 5.6% (635/11,286) of patients in the final cohort. Thirty-day mortality in patients with pneumonia was 26.9% (171/635) as compared to 4.4% (473/10,651, p 0.001) in those without pneumonia (RR 5.9 [95% CI 5.1 to 6.8]). Propensity model. The nonparsimonious model of propensity to develop pneumonia included as covariates 38 factors identified by focus group (see Appendix E-2 on the Neurology Web site; www.neurology.org). The c-statistic for the propensity model was 0.83, indicating very good ability to discriminate patients who developed pneumonia from those who did not. Pairs matched on propensity scores. Of the 635 patients with pneumonia, we matched 614 (96.7%) with patients who did not develop pneumonia, matching on propensity scores within 0.02. Distribution of characteristics among the 614 matched pairs was very similar (see table E-1 on the Neurology Web site; www. neurology.org), suggesting excellent adjustment with the use of propensity scores. Among all 11,286 patients, analyses by quintile of propensity score revealed similar balance among covariates within each quintile (data not shown). Multivariable analyses. Multivariable logistic regression analysis was the main method used to determine the effect of pneumonia on 30-day mortality. Table 3 summarizes the effect of sequentially more complete adjustments on RR, PAR, and the number of pneumonias needing to be prevented in order to avoid one death by 30 days. When adjusting only for age, race, sex, and COTH status, pneumonia was estimated to confer an almost sixfold increased risk of 30-day mortality. After additionally adjusting for admission severity of illness, RR declined to 4.43. Finally, after also adjusting for patients propensity to develop pneumonia, RR declined to 2.99 (95% CI 2.44 to 3.66). Analogous decreases were observed in PAR, along with increases in the number of pneumonias needing to be prevented in order to avoid one death by 30 days. As measured by the c-statistic, the discrimination improved (from 0.67 to 0.80) with addition to the baseline model admission severity and propensity to develop pneumonia. We sought to determine whether the effect of pneumonia on 30-day mortality varied among predefined subgroups after adjusting for the same covariates used in the main analysis (figure). Pneumonia conferred a higher RR of mortality in hemorrhagic stroke (RR 3.46 [95% CI 2.78 to 4.27]) as compared with ischemic stroke (RR 1.81 [95% CI 1.05 to 3.37]). There also was a significant difference among racial groups: the RR of pneumonia for 30-day mortality in white patients was 3.43 (95% CI 2.75 to 4.24) as compared with 1.56 (95% CI 0.90 to 2.65) in black patients. To evaluate possible explanations for the racial disparity in the effect of pneumonia, we compared race-related differences in the frequencies of various baseline characteristics associated with 30-day death. Although the white patients were older than the black patients who developed pneumonia (78.3 years vs 74.7 years, p 0.0001), there were no significant race-related differences among patients with or without pneumonia by stroke type (hemorrhagic vs ischemic), DNR status after 3 days, the presence of paralysis, lethargy or stupor, or overall severity of illness on admission (data not shown). Although the rates of coma were low in both groups, black patients had a higher incidence of coma than white patients

in the nonpneumonia group (2.1% vs 1.3%, p 0.006, unadjusted for multiple comparisons). The adjusted RR of 90-day death for patients who developed pneumonia during their initial hospitalization was 2.15 (95% CI 1.83 to 2.51), slightly lower than the RR for 30-day death.

Discussion. Among 14,293 Medicare patients who were hospitalized for acute stroke in the Cleveland area between 1991 and 1997, 6.9% developed pneumonia, including 5.6% of those who survived their early hospitalizations and did not have DNR orders written within their first 3 days of admission. In this final cohort for whom efforts to prevent pneumonia are most relevant, pneumonia conferred a threefold increased risk of 30-day mortality after rigorous adjustment for severity of illness and other baseline features associated with the development of pneumonia. We estimate that 10% of deaths within 30 days of admission among hospitalized patients with stroke are attributable to pneumonia, and that one death could be averted for every 11 patients in whom stroke-related pneumonia is prevented. Several factors on admission were associated with increased risk of pneumonia. Relatively common baseline characteristics found to be associated with

Figure. Relative risks of 30-day death with pneumonia among subgroups. NH admit admission from nursing home. Black box yes; gray x no.
February (2 of 2) 2003 NEUROLOGY 60 623

pneumonia included admission from a nursing home, diminished nutritional status as defined by admission albumin level, male sex, selected comorbid illnesses (including obstructive airways disease and cardiac conditions), and hemorrhagic stroke. Stroke patients with paralysis and those with impaired level of consciousness also had had higher rates of pneumonia. These findings have been noted previously.9 Patients with altered level of consciousness not only have more severe neurologic impairments but also are less able to protect their airway effectively. Our finding of a 26.9% mortality rate among the 635 patients with stroke who developed pneumonia is remarkably similar to the 24% rate reported in a prior study,10 although that study included only 21 patients with pneumonia. The unadjusted sixfold increased risk of death among patients who developed pneumonia reflected not only the effects of their pneumonia, but also both their greater stroke severity on admission and other factors that predisposed them to developing pneumonia, factors not rigorously examined in previous work. After adjusting for admission severity using a validated model of predicted mortality, the RR of pneumonia declined to 4.4, and declined further to 3.0 after adjusting for patients propensity to develop pneumonia. Propensity analysis has been used increasingly in observational studies to balance factors associated with exposures that are not assigned at random, including both tests19 and treatments.23,24 Factors on admission that were imbalanced included features that are not typically considered in adjusting for severity of illness. By using an expert panel to specifically address features that might predispose toward pneumonia among patients with stroke, including markers for general frailty such as nutritional status and nursing home residence before admission, we were able to more accurately estimate the independent effect of pneumonia on all-cause mortality. We believe our estimate that 10% of deaths occurring in patients with stroke were attributable to pneumonia is a conservative one. We excluded patients who died or who had DNR orders written within 3 days of admission because we wanted to examine patients for whom efforts at pneumonia prevention, early detection, and rapid treatment were clearly pertinent. Pneumonia occurring so early after stroke may have been less preventable and the medical care of early DNR patients may have focused on comfort measures, with pneumonia developing as an expected part of terminal events. Our cohort also included only first-time admissions for stroke between 1991 and 1997. Patients with documented prior stroke admissions during this period are likely to have had increased neurologic impairments and may have been at an even greater risk of developing pneumonia. Finally, because we did not record process-of-care measures designed to treat identified pneumonias in this observational study, the estimated effect of pneumonia does not account for these beneficial interventions. By extrapolating the cur624 NEUROLOGY 60 February (2 of 2) 2003

rent findings to the 700,000 strokes that occur annually in the United States (http://www.stroke.org/ brain_stat.cfm), including similar rates of early DNR and deaths within 3 days, we estimate that 3,155 deaths in the remaining 553,000 patients with stroke can be attributed to pneumonia annually. Based on our number needed to treat analyses (see table 3), we estimate that one stroke-related death at 30 days could be avoided for each 11 cases of pneumonia that are prevented. A recent Evidence Report of the Agency for Health Care Research and Quality3 highlighted several means by which patients with stroke who are at increased risk for pneumonia can be identified and targeted for preventive interventions. Although the report called for more research on risk stratification, clinical interventions to reduce the incidence of pneumonia, and the effectiveness of early treatments, it concluded that clinical assessments can identify the presence of dysphagia and aspiration central posited causes of pneumonia using swallowing evaluations performed either at the bedside or using videofluoroscopy.3 Pneumonia conferred a lower RR of mortality at 90 days than 30 days, although there was still a twofold increased risk. Deaths occurring farther from the hospital admission where the pneumonia occurred were likely less related to pneumonia and more related to other factors such as cardiovascular events or intervening therapies. In subgroup analyses, white patients who developed pneumonia had a significantly higher relative risk of death than black patients. There were no differences between white and black patients with pneumonia in variables generally associated with higher mortality rates. Possible explanations for our findings include a survival bias among black patients who live long enough to be age-eligible for Medicare,25 the occurrence of pneumonias that were less severe among black patients with stroke, and more aggressive preventive or treatment efforts among black than white patients. Our data do not permit us to examine these alternative explanations and suffer from three general limitations. First, we used ICD-9-CM codes to identify the occurrence of pneumonia and we were unable to identify the severity of pneumonia. The sensitivity of identifying complications using ICD-9-CM codes is low, ranging from 25 to 57%.26-29 Nonetheless, coding of complications on medical services appears to fare better than on surgical services,26 and infections have been reported to have the highest overall accuracy, with an ROC curve area of 0.71 and a high specificity of 95%.29 Thus, false-positive rates of identified pneumonia are likely to be low. Further, we restricted the spectrum of pneumonia-related codes to those most likely to reflect complications, with over three quarters of identified cases coded either as aspiration pneumonia (ICD-9-CM 507.0 507.8) or pneumonia, organism unspecified (ICD-9-CM 486.xx). Finally, the overall incidence of pneumonia in the current investigation matches well with other

studies, including a prospective investigation undertaken as part of a clinical trial.30 Second, although our data permitted rigorous adjustment for patient features at the time of admission, certain desirable admission data were not recorded in our records reviews, and process-related data during the course of hospitalization were not obtained. Many important neurologic findings on admission were assiduously recorded as part of the CHQC, although the severity of neurologic impairment as defined by a stroke impairment scale such as the NIH Stroke Scale was not determined. Our predicted 30-day mortality model, however, had good performance as assessed by the c-statistic. In addition, the presence of dysphagia or related signs of risk for aspiration and subsequent pneumonia were not recorded. Because these signs are likely to be highly collinear with other captured elements, howeversuch as the presence of lethargy or stupor, or the presence of blood or mass effect on brain imaging studies our propensity model was able to accurately discriminate patients who eventually developed pneumonia from those who did not. The absence of process-of-care measures in our data has other important implications for our findings and identifies the need for future research. First, as mentioned earlier, we did not record information about treatment of pneumonias that were identified clinically, and we therefore cannot isolate the effects of pneumonia from the effects of treatment. Because patients who were effectively treated may have been more likely to survive, we believe that our findings are likely to have underestimated the independent mortality risk of pneumonia. Future research should examine the effectiveness of alternative treatment strategies in patients with stroke who develop pneumonia. Finally, because the presence of measures designed to prevent aspiration and pneumonia was not recorded, we are unable to precisely isolate the risk of developing stroke-related pneumonia or to determine the effectiveness of preventive measures. As a result, in addition to the insensitivity of using ICD-9-CM codes to identify complications, as mentioned above, we expect that our observed incidence of pneumonia is likely to underestimate the true incidence (i.e., the incidence of pneumonia in the absence of preventive interventions). These results underscore the call by others1 for further research to identify patients with stroke at increased risk for pneumonia and for studies of the effectiveness of alternative approaches for prevention.
Acknowledgment
The authors thank Dr. Lawrence M. Brass for his thoughtful review of this manuscript. They also acknowledge the assistance of Charles Thomas, who provided SAS macros for matching and helped with statistical applications in SAS.

References
1. Andersen HE, Schultz-Larsen K, Kreiner S, Forchhammer BH, Eriksen K, Brown A. Can readmission after stroke be prevented? Results of a randomized clinical study: a postdischarge follow-up service for stroke survivors. Stroke 2000;31:1038 1045.

2. Roth EJ, Lovell L, Harvey RL, Heinemann AW, Semik P, Diaz S. Incidence of and risk factors for medical complications during stroke rehabilitation. Stroke 2001;32:523529. 3. ECRI. Diagnosis and treatment of swallowing disorders (dysphagia) in acute-care stroke patients. 99-E024. Evidence report/technology assessment, number 8. Rockville, MD: Agency for Health Care Policy and Research, 1999. 4. Langhorne P, Stott DJ, Robertson L, et al. Medical complications after stroke: a multicenter study. Stroke 2000;31:12231229. 5. Haerer AF, Smith RR. Medical and surgical experiences in patients of a large southern stroke center. South Med J 1974;67:667 671. 6. Davenport RJ, Dennis MS, Wellwood I, Warlow CP. Complications after acute stroke. Stroke 1996;27:415 420. 7. Pinto AN, Melo TP, Lourenco ME, et al. Can a clinical classification of stroke predict complications and treatments during hospitalization? Cerebrovasc Dis 1998;8:204 209. 8. Riquelme R, Torres A, el Ebiary M, et al. Community-acquired pneumonia in the elderly: a multivariate analysis of risk and prognostic factors. Am J Respir Crit Care Med 1996;154:1450 1455. 9. Thompson RS, Hall NK, Szpiech M, Reisenberg LA. Treatments and outcomes of nursing-home-acquired pneumonia. J Am Board Fam Pract 1997;10:82 87. 10. Reynolds P, Gilbert L, Good D, et al. Pneumonia in dysphagic stroke patients: effect on outcomes and identification of high risk patients. J Neuro Rehab 1998;12:1521. 11. Rosenthal GE, Harper DL. Cleveland health quality choice: a model for collaborative community-based outcomes assessment. Jt Comm J Qual Improv 1994;20:425 442. 12. Rosenthal GE, Harper DL, Quinn LM, Cooper GS. Severity-adjusted mortality and length of stay in teaching and nonteaching hospitals. Results of a regional study [see comments]. JAMA 1997;278:485 490. 13. Benesch C, Witter DM, Jr., Wilder AL, Duncan PW, Samsa GP, Matchar DB. Inaccuracy of the International Classification of Diseases (ICD-9-CM) in identifying the diagnosis of ischemic cerebrovascular disease [published erratum appears in Neurology 1998;50:306] [see comments]. Neurology 1997;49:660 664. 14. Goldstein LB. Accuracy of ICD-9-CM coding for the identification of patients with acute ischemic stroke: effect of modifier codes. Stroke 1998;29:16021604. 15. Bamford J, Dennis M, Sandercock P, Burn J, Warlow C. The frequency, causes and timing of death within 30 days of a first stroke: the Oxfordshire Community Stroke Project. J Neurol Neurosurg Psychiatry 1990; 53:824 829. 16. Baker D, Einstadter D, Thomas C, et al. Mortality trends during a program that publicly reported hospital performance. Med Care 2002; 40:879 890. 17. Shepardson LB, Youngner SJ, Speroff T, OBrien RG, Smyth KA, Rosenthal GE. Variation in the use of do-not-resuscitate orders in patients with stroke. Arch Intern Med 1997;157:18411847. 18. Rubin DB. Estimating causal effects from large data sets using propensity scores. Ann Intern Med 1997;127(8 Pt 2):757763. 19. Connors AF, Jr., Speroff T, Dawson NV, et al. The effectiveness of right heart catheterization in the initial care of critically ill patients. SUPPORT Investigators [see comments]. JAMA 1996;276:889 897. 20. Zhang J, Yu KF. Whats the relative risk? A method of correcting the odds ratio in cohort studies of common outcomes. JAMA 1998;280:1690 1691. 21. Goldstein LB, Adams R, Becker K, et al. Primary prevention of ischemic stroke: a statement for healthcare professionals from the Stroke Council of the American Heart Association. Circulation 2001;103:163182. 22. Whisnant JP. Modeling of risk factors for ischemic stroke. The Willis Lecture. Stroke 1997;28:1840 1844. 23. Gum PA, Thamilarasan M, Watanabe J, Blackstone EH, Lauer MS. Aspirin use and all-cause mortality among patients being evaluated for known or suspected coronary artery disease: a propensity analysis. JAMA 2001;286:11871194. 24. Radford MJ, Foody JM. How do observational studies expand the evidence base for therapy? JAMA 2001;286:1228 1230. 25. Howard G, Russell GB, Anderson R, et al. Role of social class in excess black stroke mortality. Stroke 1995;26:1759 1763. 26. Massanari RM, Wilkerson K, Streed SA, Hierholzer WJ, Jr. Reliability of reporting nosocomial infections in the discharge abstract and implications for receipt of revenues under prospective reimbursement. Am J Public Health 1987;77:561564. 27. Faciszewski T, Johnson L, Noren C, Smith MD. Administrative databases complication coding in anterior spinal fusion procedures. What does it mean? Spine 1995;20:17831788. 28. Hartz AJ, Kuhn EM. Comparing hospitals that perform coronary artery bypass surgery: the effect of outcome measures and data sources. Am J Public Health 1994;84:1609 1614. 29. Geraci JM, Ashton CM, Kuykendall DH, Johnson ML, Wu L. International Classification of Diseases, 9th revision, clinical modification codes in discharge abstracts are poor measures of complication occurrence in medical inpatients. Med Care 1997;35:589 602. 30. Johnston KC, Li JY, Lyden PD, et al. Medical and neurological complications of ischemic stroke: experience from the RANTTAS trial. RANTTAS Investigators. Stroke 1998;29:447 453. February (2 of 2) 2003 NEUROLOGY 60 625