Logistic regression analysis of biomarker data subject to pooling and dichotomization.

There is growing interest in pooling specimens across subjects in epidemiologic studies, especially those involving biomarkers. This paper is concerned with regression analysis of epidemiologic data where a binary exposure is subject to pooling and the pooled measurement is dichotomized to indicate either that no subjects in the pool are exposed or that some are exposed, without revealing further information about the exposed subjects in the latter case. The pooling process may be stratified on the disease status (a binary outcome) and possibly other variables but is otherwise assumed random. We propose methods for estimating parameters in a prospective logistic regression model and illustrate these with data from a population-based case-control study of colorectal cancer. Simulation results show that the proposed methods perform reasonably well in realistic settings and that pooling can lead to sizable gains in cost efficiency. We make recommendations with regard to the choice of design for pooled epidemiologic studies.

Differential Misclassification of Disease under Partial-Mouth Sampling.

AIM:: The effect of misclassification of a cluster-level dichotomous outcome (disease) due to partial-cluster sampling on its association with a dichotomous exposure is investigated. METHODS:: Disease (e.g., chronic periodontitis) is deemed to exist in a cluster (e.g., full mouth) when a condition of interest (e.g., pocket depth or clinical attachment loss exceeding an established threshold) is present in number and pattern across observations (e.g., tooth sites) in the cluster according to a specific criterion. When a subset of observations within each cluster is selected (i.e., partial-mouth sampling), specificity of disease is 100% (in the absence of site-level measurement error), whereas sensitivity is imperfect and generally unknown. Using conditional probability arguments, we investigate disease misclassification under partial-cluster sampling and its impact on the estimated disease-exposure association when the exposure is cluster level and measured without error. RESULTS:: When the probability of disease varies by exposure status, outcome misclassification at the cluster level is differential under partial-cluster sampling and depends on 1) the partial recording protocol, including the number of observations sampled and the particular sites selected in a cluster; 2) the joint probability structure of the condition within clusters; and 3) the criterion for disease. A numeric example demonstrates that disease-exposure odds ratios under partial-cluster random sampling can be biased in either direction (toward or away from the null) relative to gold-standard odds ratios under full-cluster sampling. CONCLUSIONS:: In general, misclassification of disease is differential under partial-cluster sampling. In particular, sensitivity and negative predictive values depend on exposure status, which leads to biased inference. KNOWLEDGE TRANSFER STATEMENT:: Partial-mouth sampling causes disease misclassification probabilities, including sensitivity, to vary by exposure groups when disease prevalence differs between groups. As a result, disease-exposure associations may be under- or overestimated by standard analysis procedures for periodontal data relative to full-mouth estimates. Procedures that address bias are needed for partial-recording protocols.

(F-18) fluorodeoxyglucose positron emission tomography as a predictor of pathologic grade and other prognostic variables in bone and soft tissue sarcoma.

Positron emission tomography (PET) can be used to measure tumor metabolism in sarcomas by measuring the standard uptake value (SUV) of (F-18) fluorodeoxyglucose (FDG). FDG-PET SUV has been shown to correlate with histological grade. We compared FDG-PET SUV in 89 bone and soft tissue sarcomas with histopathological features, including tumor grade, as well as with markers of cell proliferation and cell cycle regulatory gene expression that may be prognostically or therapeutically important. All patients had undergone PET before biopsy. Features evaluated included grade (National Cancer Institute for soft tissue or Mayo Clinic for bone), cellularity, and the number of mitoses per 10 400x fields. Deparaffinized, formalin-fixed sections were immunostained with antibodies to Ki-67 (MIB-1), p53 (DO7), p21WAF1 (EA10), and mdm-2 (1B10). For Ki-67, results were estimated as a percentage of positive cells. For p53 and mdm-2, only cases with >20% positive cells were considered to be overexpressing these proteins. For p21WAF1, only cases with <10% positive cells were considered to have lost normal p21WAF1 expression. Tumor S-phase percentage and ploidy were determined by flow cytometry. FDG-PET SUV was associated with histopathological grade, cellularity, mitotic activity, MIB labeling index, and p53 overexpression. No association was seen with p21WAF1, mdm-2, S-phase fraction, or ploidy. Tumor metabolism data acquired by FDG-PET may help ensure accurate grading and prognostication in sarcoma by guiding biopsy toward the most biologically significant regions of large masses. Further follow-up will be necessary to determine whether FDG-PET provides independent prognostic information.

The potentially poor response to outpatient diabetes care in urban African-Americans.

OBJECTIVE: HbA1c levels can be reduced in populations of diabetic patients, but some individuals may exhibit little improvement. To search for reasons underlying differences in HbA1c outcome, we analyzed patients managed in an outpatient diabetes clinic. RESEARCH DESIGN AND METHODS: African-Americans with type 2 diabetes were categorized as responders, intermediate responders or poor responders according to their HbA1c level after 1 year of care. Logistical regression was used to determine baseline characteristics that distinguished poor responders from responders. Therapeutic strategies were examined for each of the response categories. RESULTS: The 447 patients had a mean age and disease duration of 58 and 5 years, respectively, and BMI of 32 kg/m2. Overall, the mean HbA1c level fell from 9.6 to 8.1% after 12 months. Mean HbA1c levels improved from 8.8 to 6.2% in responders, and from 9.5 to 7.9% in intermediate responders. In poor responders, the average HbA1c level was 10.8% on presentation and 10.9% at 1 year. The odds of being a poor responder were significantly increased with longer disease duration, higher initial HbA1c level, and greater BMI. Although doses of oral agents and insulin were significantly higher among poor responders at most visits, the acceleration of insulin therapy did not occur until late in the follow-up period. CONCLUSIONS: Clinical diabetes programs need to devise methods to identify patients who are at risk for persistent hyperglycemia. Whereas patient characteristics explain some heterogeneity of HbA1c outcome (and may aid in earlier identification of patients who potentially may not respond to conventional treatment), insufficient intensification of therapy may also be a component underlying the failure to achieve glycemic goals.

Prognostic value of plasma HIV RNA in the natural history of Pneumocystis carinii pneumonia, cytomegalovirus and Mycobacterium avium complex. Multicenter AIDS Cohort Study.

OBJECTIVES: To use follow-up on untreated HIV-positive men to assess the prognostic information provided by baseline data on plasma HIV RNA, CD4 cell count, age, and HIV-related symptom status, separately for three specific AIDS-defining illnesses: Pneumocystis carinii pneumonia (PCP), cytomegalovirus (CMV), and Mycobacterium avium complex (MAC). METHODS: The study population were 734 HIV-positive homosexual men enrolled in the Multicenter AIDS Cohort Study, with follow-up (1984-1985 through mid-1988) restricted to the antiretroviral treatment-free and prophylaxis-free era. Baseline marker values were categorized and assessed as predictor variables in separate time-to-event analyses for each of the three specific outcomes. RESULTS: A total of 138 cases of PCP, 25 cases of CMV, and 25 cases of MAC were observed. For PCP and CMV, higher categories of HIV RNA and lower categories of CD4 cell count were associated with increased risk relative to the respective reference groups. For MAC, oral candidiasis or fever and elevated HIV RNA at baseline were the primary risk factors. Further analysis highlighted the importance of monitoring HIV RNA levels in addition to CD4 cell counts when evaluating patients' risk of developing PCP. CONCLUSIONS: In the absence of treatment, plasma HIV RNA levels provide prognostic information about the risk of these three specific AIDS-defining illnesses, independently of the CD4 cell count. These data provide a useful reference as researchers investigate changing patterns in the incidence and predictors of opportunistic infections in the era of increasingly active antiretroviral therapies.

Plasma viral load and CD4 lymphocytes predict HIV-associated dementia and sensory neuropathy.

OBJECTIVE: To determine the predictive value of plasma HIV RNA and CD4 lymphocytes for HIV-associated dementia and sensory neuropathy. METHODS: A total of 1,604 AIDS-free HIV seropositive men from the Multicenter AIDS Cohort Study were followed over a 10-year period (1985 to 1995). HIV-associated dementia and sensory neuropathy were diagnosed according to standard definitions. Baseline samples were used to measure plasma HIV RNA levels with a branched DNA assay and levels of beta2-microglobulin, CD4 lymphocyte counts, and hemoglobin levels. RESULTS: Seventy-seven patients with HIV-associated dementia and 213 patients with sensory neuropathy were identified. Baseline HIV RNA levels above 3,000 copies/mL and CD4 counts below 500 cells/mm3 were predictive of both neurologic outcomes, but neither hemoglobin, body mass index, nor beta2-microglobulin were independently predictive. After adjusting for age and level of education, individuals with baseline plasma HIV RNA >30,000 copies/mL had a relative hazard for dementia 8.5 times (p < 0.001) that of those with <3,000 copies/mL, and those with CD4 counts <200 cells/mm3 had a 3.5-fold (p = 0.003) greater hazard relative to those with CD4 counts >500 cells/mm3. Individuals with HIV RNA >10,000 copies/mL had a 2.3-fold (p = 0.008) greater hazard of sensory neuropathy than those with <500 copies/mL, and men with <750 CD4 cells/mm3 had a 1.4-fold (p = 0.03) greater hazard than those with >750 CD4 cells/mm3. CONCLUSIONS: High levels of systemic HIV replication may "drive" the initiation of neurologic disease; effective suppression of HIV may reduce the incidence of dementia and neuropathy. Levels of plasma HIV RNA and CD4 counts, determined before the initiation of antiretroviral therapy, were predictive of HIV-associated dementia and sensory neuropathy.

HIV-associated neurologic disease incidence changes:: Multicenter AIDS Cohort Study, 1990-1998.

This study examined the temporal trends in the incidence rates of HIV dementia, cryptococcal meningitis, toxoplasmosis, progressive multifocal leukoencephalopathy, and CNS lymphoma from January 1990 to December 1998 in the Multicenter AIDS Cohort Study. The incidence rates for HIV dementia, cryptococcal meningitis, and lymphoma decreased following the introduction of highly active antiretroviral therapy (HAART). The proportion of new cases of HIV dementia with a CD4 count in a higher range (i.e., 201 to 350) since 1996 may be increasing.

Combination antiretroviral therapy improves psychomotor speed performance in HIV-seropositive homosexual men. Multicenter AIDS Cohort Study (MACS).

BACKGROUND: Combination antiretroviral therapy including protease inhibitors (combo+PI) is effective in suppressing systemic viral load in HIV infection, but its impact on HIV-associated cognitive impairment is unclear. OBJECTIVE: To determine whether psychomotor speed, a sensitive measure of impairment in HIV dementia, improves with combo+PI compared with other antiretroviral treatments. METHODS: A total of 411 HIV-seropositive (HIV+) homosexual men (with longitudinal neuropsychological testing) in the Multicenter AIDS Cohort Study and, in a separate analysis, 282 HIV+ homosexual men with psychomotor slowing at baseline were classified by treatment into four groups: antiretroviral naive (no antiretroviral medication treatment), monotherapy, combination antiretroviral therapy without protease inhibitors (combo-noPI), and combo+PI. We compared longitudinal performance on three tests of psychomotor speed: the Grooved Pegboard (GP) (nondominant and dominant hands), Trail Making Test B, and the Symbol Digit Modalities Test (SDMT). RESULTS: Relative to antiretroviral-naïve and monotherapy participants, on the GP nondominant hand test, combo+PI participants with abnormal baseline neuropsychological testing showed improved performance (difference = +0.63 standard deviation [SD], p = 0.02). For the SDMT, both combo+PI participants (difference = +0.26 SD, p = 0.03) and combo-noPI participants (difference = +0.29 SD, p = 0.01) with abnormal baseline neuropsychological testing improved compared with antiretroviral-naïve and monotherapy groups. CONCLUSION: Combo+PI and combo-noPI are associated with improved psychomotor speed performance in HIV+ homosexual men with abnormal neuropsychological testing.

Impact of vitamin D supplementation on markers of inflammation in adults with cystic fibrosis hospitalized for a pulmonary exacerbation.

Patients with cystic fibrosis (CF) suffer from chronic lung infection and inflammation leading to respiratory failure. Vitamin D deficiency is common in patients with CF, and correction of vitamin D deficiency may improve innate immunity and reduce inflammation in patients with CF. We conducted a double-blinded, placebo-controlled, randomized clinical trial of high-dose vitamin D to assess the impact of vitamin D therapy on antimicrobial peptide concentrations and markers of inflammation. We randomized 30 adults with CF hospitalized with a pulmonary exacerbation to 250,000 IU of cholecalciferol or placebo, and evaluated changes in plasma concentrations of inflammatory markers and the antimicrobial peptide LL-37 at baseline and 12 weeks post intervention. In the vitamin D group, there was a 50.4% reduction in tumor necrosis factor-α (TNF-α) at 12 weeks (P<0.01), and there was a trend for a 64.5% reduction in interleukin-6 (IL-6) (P=0.09). There were no significant changes in IL-1ß, IL-8, IL-10, IL-18BP and NGAL (neutrophil gelatinase-associated lipocalin). We conclude that a large bolus dose of vitamin D is associated with reductions in two inflammatory cytokines, IL-6 and TNF-α. This study supports the concept that vitamin D may help regulate inflammation in CF, and that further research is needed to elucidate the potential mechanisms involved and the impact on clinical outcomes.

Effects of model misspecification in the estimation of variance components and intraclass correlation for paired data.

Paired data occur in many experimental situations. When one views the subjects as a random sample from some large population, it may seem reasonable to model the data according to the typical one-way random effects analysis of variance (ANOVA). It is then usually of interest to estimate variance components and intraclass correlation. These estimators can be biased if key assumptions are violated, leading to erroneous interpretations and conclusions. We focus upon assumptions about the equality or inequality of means and/or variances of the two measures on each subject. In the framework of the one-way random effects ANOVA model, and three generalizations of it, we document estimators obtained as solutions to the likelihood equations. We consider the potentially serious effects of mistaken assumptions. Our findings suggest that the most general model considered is most desirable if consistent and efficient estimation of the between-subject variance component and intraclass correlation is the main goal. We also briefly connect our exposition to the study of reliability or agreement.

On strategies for comparing occupational exposure data to limits.

Parametric statistical approaches to assessing workplace exposure levels have typically focused either on the probability that a single measurement exceeds a limit or on whether the mean exposure for a population of workers exceeds a limit. This article reviews and clarifies some methods that have been proposed for each of these two approaches, on the assumption that the exposure data represent a random sample from a lognormal distribution. For tests concerning the mean exposure level, the authors developed a potentially useful new procedure based on a bound for noncentral t critical values. Appropriate sample size calculations are emphasized, and computer simulation is used to compare competing methods for assessing mean exposure. The authors conclude that the new proposed method offers an appealing alternative to existing methods in many cases. The importance of employing an exposure assessment strategy that is in concert with underlying etiologic considerations is stressed.

A detailed evaluation of adjustment methods for multiplicative measurement error in linear regression with applications in occupational epidemiology.

It is often appropriately assumed, based on both theoretical and empirical considerations, that airborne exposures in the workplace are lognormally distributed, and that a worker's mean exposure over a reference time period is a key predictor of subsequent adverse health effects for that worker. Unfortunately, it is generally impossible to accurately measure a worker's true mean exposure. We begin by introducing a familiar model for exposure that views this true mean, as well as logical surrogates for it, as lognormal random variables. In a more general context, we then consider the linear regression of a continuous health outcome on a lognormal predictor measured with multiplicative error. We discuss several candidate methods of adjusting for the measurement error to obtain consistent estimators of the true regression parameters. These methods include a simple correction of the ordinary least squares estimator based on the surrogate regression, the regression of the outcome on the covariates and on the conditional expectation of the true predictor given the observed surrogate, and a quasi-likelihood approach. By means of a simulation study, we compare the various methods for practical sample sizes and discuss important issues relevant to both estimation and inference. Finally, we illustrate promising adjustment strategies using actual lung function and dust exposure data on workers in the Dutch animal feed industry.

Classifying individuals based on predictors of random effects. Multicenter AIDS Cohort Study.

Often one wishes to describe individuals according to whether their average exposure over a period of time is above or below some meaningful threshold. In this article, we treat predictors of random effects as diagnostic tools to aid in such classification, given that the true unobservable mean exposure for each of a set of individuals is defined according to a mixed linear model. Viewing candidate predictors in this light engenders the consideration of a unique set of performance criteria, and invites the use of nomenclature commonly used by epidemiologists and decision analysts to evaluate diagnostic techniques. We describe these criteria analytically and graphically under a random effects analysis of variance model, with the expressed goal of classifying subjects with regard to their true mean. Given knowledge of the model parameters, we compare typical predictors and illustrate the fact that completely new alternatives can arise depending on the particular set of criteria emphasized. We include a brief simulation study in which we also compare prediction methods according to various classification criteria, after incorporating estimates of the unknown model parameters. We provide two examples using data from participants in the Multicenter AIDS Cohort Study. In the first example, we seek to classify HIV seronegative individuals based on their mean diastolic blood pressure. In the second, via a natural extension to a randomized regression model, we classify HIV seropositive individuals according to their CD4+ slope over time.

Effects of covariate measurement error in the initial level and rate of change of an exposure variable.

When repeated measures of an exposure variable are obtained on individuals, it can be of epidemiologic interest to relate the slope of this variable over time to a subsequent response. Subject-specific estimates of this slope are measured with error, as are corresponding estimates of the level of exposure, i.e., the intercept of a linear regression over time. Because the intercept is often correlated with the slope and may also be associated with the outcome, each error-prone covariate (intercept and slope) is a potential confounder, thereby tending to accentuate potential biases due to measurement error. Under a familiar mixed linear model for the exposure measurements, we present closed-form estimators for the true parameters of interest in the case of a continuous outcome with complete and equally timed follow-up for all subjects. Generalizations to handle incomplete follow-up, other types of outcome variables, and additional fixed covariates are illustrated via maximum likelihood. We provide examples using data from the Multicenter AIDS Cohort Study. In these examples, substantial adjustments are made to uncorrected parameter estimates corresponding to the health-related effects of exposure variable slopes over time. We illustrate the potential impact of such adjustments on the interpretation of an epidemiologic analysis.

An exposure-assessments strategy accounting for within- and between-worker sources of variability.

A strategy is presented for comparing exposures to an occupational exposure limit (OEL) and for suggesting appropriate interventions when exposures are unacceptable. The major departure from previous approaches is the explicit recognition that exposures vary both within and between workers in a given occupational group. The primary goal is to determine whether the probability of overexposure is acceptably small (a value of 0.10 or less is recommended), with overexposure being defined as the likelihood that a randomly selected worker's true mean exposure exceeds the OEL. The exposure-assessment protocol contains five levels. It is suggested that at least two shift-long measurements be randomly collected from each of 10 workers for preliminary analysis. If the logged exposure data appear to be appropriate for testing (Level 1), the probability of overexposure is compared to the pre-determined value via a rigorous test of statistical significance (Level 2). Based upon published data, this test is likely to classify exposures as acceptable with 20 measurements when the group mean exposure is less than one-fifth of the OEL. However, if exposure is found to be unacceptable, re-sampling can be considered to increase the power of the test (Level 3). Otherwise, it is necessary to reduce exposures and then to re-apply the protocol. If it appears that all persons in the group have essentially the same predicted mean exposures (Level 4), then engineering or administrative controls are recommended. If, on the other hand, substantial differences appear to exist amongst these predicted mean values, regrouping and/or modifications of tasks and work practices should be considered (Level 5). Application of the protocol is illustrated with samples of data from four groups of workers exposed to inorganic nickel in the nickel-producing industry.

Correlation coefficient estimation involving a left censored laboratory assay variable.

When assessing a correlation between two exposure or biological marker variables, one sometimes encounters the problem of indeterminate values for one of the variables due to an assay detection limit. In this event, investigators often report correlation coefficients computed after removing the pairs involving non-detectable values, or after substituting some small constant for those values. These ad hoc practices can lead to bias in both point and confidence interval estimates of the true correlation coefficient. To address this issue, we consider two parametric techniques for estimating the correlation in the presence of left censoring for one of the variables. The first is a maximum likelihood approach, and the second is an adaptation of multiple imputation motivated primarily by potential benefits in confidence interval coverage. Both of the estimators studied reduce to the standard Pearson's correlation coefficient in the event of no censoring, and hence are valid in cases where this measure would be appropriate for the complete data. We assess these approaches empirically and contrast them with ad hoc methods for estimating the correlation between cervicovaginal human immunodeficiency virus (HIV) viral load measurements and CD4+ lymphocyte counts from HIV positive women enrolled in a clinical trial conducted in Bangkok, Thailand.

A lognormal distribution-based exposure assessment method for unbalanced data.

We present a generalization of existing statistical methodology for assessing occupational exposures while explicitly accounting for between- and within-worker sources of variability. The approach relies upon an intuitively reasonable model for shift-long exposures, and requires repeated exposure measurements on at least some members of a random sample of workers from a job group. We make the methodology more readily applicable by providing the necessary details for its use when the exposure data are unbalanced (that is, when there are varying numbers of measurements per worker). The hypothesis testing strategy focuses on the probability that an arbitrary worker in a job group experiences a long-term mean exposure above the occupational exposure limit (OEL). We also provide a statistical approach to aid in the determination of an appropriate intervention strategy in the event that exposure levels are deemed unacceptable for a group of workers. We discuss important practical considerations associated with the methodology, and we provide several examples using unbalanced sets of shift-long exposure data-taken on workers in various sectors of the nickel-producing industry. We conclude that the statistical methods discussed afford sizable practical advantages, while maintaining similar overall performance to that of existing methods appropriate for balanced data only.

Quantitative and qualitative differences in the serum antibody profiles of human immunodeficiency virus-infected persons with and without Cryptococcus neoformans meningitis.

The importance of humoral immunity for resistance to Cryptococcus neoformans is uncertain. A case-controlled study of the human antibody response to C. neoformans comparing the serum antibody profiles of human immunodeficiency virus (HIV)-infected persons who did (HIV+/CM+) or did not (HIV-infected controls) develop cryptococcal meningitis (CM) and HIV-uninfected persons with samples obtained from the Multicenter AIDS Cohort Study was performed. Total immunoglobulin concentrations were determined, and the specificity, isotype, and idiotype expression of antibodies to C. neoformans capsular glucuronoxylomannan were analyzed by ELISA. Compared with the HIV+/CM+ group, the HIV-infected control group had significantly lower levels of total IgM, IgA, and antibodies expressing a certain VH3 determinant. The HIV-infected control group manifested an increase in immunoglobulin levels with a decrease in CD4 lymphocytes. The findings suggest a possible association between reduced expression of certain immunoglobulin subsets and HIV-associated CM.

Correlating two viral load assays with known detection limits.

A timely objective common to many HIV studies involves assessing the correlation between two different measures of viral load obtained from each of a sample of patients. This correlation has scientific utility in a number of contexts, including those aimed at a comparison of competing assays for quantifying virus and those aimed at determining the level of association between viral loads in two different reservoirs using the same assay. A complication for the analyst seeking valid point and interval estimates of such a correlation is the fact that both variables may be subject to left censoring due to values below assay detection limits. We address this problem using a bivariate normal likelihood that accounts for left censoring of two variables that may have different detection limits. We provide simulation results to evaluate sampling properties of the resulting correlation estimator and compare it with ad hoc estimators in the presence of nondetects. In an effort to obtain improved confidence interval properties relative to the Wald approach, we evaluate and compare profile likelihood-based intervals. We apply the methods to HIV viral load data on women and infants from a trial in Bangkok, Thailand, and we discuss an extension of the original model to accommodate interval censoring arising due to the study design.

Extragastrointestinal (soft tissue) stromal tumors: an analysis of 48 cases with emphasis on histologic predictors of outcome.

The clinicopathologic features of 48 tumors that were histologically similar to gastrointestinal stromal tumors but occurred in the soft tissues of the abdomen were analyzed to determine their overall similarity to their gastrointestinal counterpart, their biologic behavior, and the parameters that predict risk for adverse outcome. Classic leiomyomas and leiomyosarcomas were specifically excluded. The tumors occurred in 32 women and 16 men, who ranged in age from 31 to 82 years (mean, 58 years). Forty tumors arose from the soft tissue of the abdominal cavity, and the remainder arose from the retroperitoneum. They ranged in size from 2.1 to 32.0 cm and varied from tumors composed purely of rounded epithelioid cells to those composed of short fusiform cells set in a fine fibrillary collagenous background with some cases showing a mixed pattern. Tumors displayed variable amounts of stromal hyalinization, myxoid change, and cyst formation. The tumors expressed CD117 (c-kit receptor) (100%), CD34 (50%), neuron-specific enolase (44%), smooth muscle actin (26%), desmin (4%), and S-100 protein (4%). Tumors were evaluated with respect to several parameters: size (<10 cm or >10 cm), cellularity (low or high), mitoses (0 to 2 per 50 high-power fields, >2 per 50 high-power fields), nuclear atypia (1 to 3+), cell type (epithelioid, spindled, or mixed), and necrosis (absent or present). These parameters were then evaluated in univariate and multivariate analysis with respect to adverse or nonadverse outcome, the former defined as metastasis or death from tumor. Follow-up information was obtained for 31 patients (range, 4 to 84 months; median, 24 months). One patient presented with an adverse event and, therefore, was excluded from subsequent analysis. Twelve patients (39%) developed metastases or died of tumor. In univariate analyses, cellularity, mitotic activity (>2 per 50 high-power fields), and necrosis were associated with statistically significant increases in the risk for adverse outcome. Despite the relatively small sample size, in a multivariable analysis mitotic activity (relative risk, 7.46; P = .09) and necrosis (relative risk, 3.75; P = .07) displayed trends toward independent predictive value. No association was noted between histologic pattern and outcome. Although only 39% of tumors behaved in a malignant fashion, this figure probably represents a conservative estimate because long-term follow-up (>5 years) was available for only a limited number of patients. Stratification of patients who have extragastrointestinal stromal tumor into those with 0 to 1 adverse histologic factors versus those with 2 to 3 offers the advantage of separating patients into two groups that have a markedly different risk for adverse outcome in the short term (0.02 events versus 0.54 events per person-year; P < .001, respectively). Extragastrointestinal (soft tissue) stromal tumors are histologically and immunophenotypically similar to their gastrointestinal counterpart but have an aggressive course more akin to small intestinal than gastric stromal tumors.