Background and Purpose We sought to identify subgroups of stroke patients in whom thrombolytic therapy is particularly hazardous or efficacious.
Methods We conducted a post hoc subgroup analysis of a randomized, double-blind, placebo-controlled clinical trial of intravenous tissue plasminogen activator (t-PA) for stroke patients presenting within 3 hours after symptom onset. Before treatment, historical, physical, and laboratory findings were summarized. We identified variables that might predict outcome and/or differential response to t-PA therapy. Outcome was measured with four stroke rating scales administered 3 months after treatment. Statistical significance was assessed with a global outcome procedure that considers the results of all four scales simultaneously. Using regression analysis, we compared the information collected before treatment with the global outcome. Multivariable procedures were used to find information that could guide selection of patients for t-PA therapy.
Results No pretreatment information significantly affected patients’ response to t-PA. The power of the model to detect a treatment interaction was greater than 90%, and therefore the probability of a type II error is very low. Apart from t-PA therapy, outcome was related to age-by-deficit severity interaction, diabetes, age-by-blood pressure interaction, and early CT findings. These variables and interactions altered long-term patient outcome irrespective of t-PA treatment but did not alter the likelihood of responding favorably to t-PA therapy.
Conclusions Patients should be selected for t-PA thrombolysis according to the guidelines published in the report of the NINDS t-PA Stroke Trial. Further subselection of patients, such as by age or stroke severity, is not supported by our post hoc analysis.
There is evidence that intravenous thrombolysis with t-PA given within 3 hours of stroke onset improves long-term outcome.1 Appropriately selected patients treated with t-PA had a 30% to 50% greater likelihood of nearly complete functional and neurological recovery relative to placebo-treated control subjects. The American Academy of Neurology and the American Heart Association have recommended the use of this therapy in patients who meet the patient selection criteria published with the trial.2 3 In applying these selection criteria to individual patients, it would be useful to know whether any particular subgroups of patients have an increased likelihood to benefit or suffer harm from t-PA. Indeed, some critics of the original work have noted the lack of such subgroup data in the literature.4 The present study was undertaken to address these concerns. We compared final outcome in t-PA–and placebo-treated patients with available pretreatment information to identify subgroups that may or may not particularly benefit from t-PA treatment.
Subjects and Methods
The NINDS t-PA Stroke Trial was a placebo-controlled, randomized, double-blind, multicenter trial.1 The trial was conducted in two parts, each with identical methods including selection criteria, drug administration, and outcome measures. The first trial indicated 24 hours after treatment as the primary time point for analysis, but 3-month data were collected. The second trial indicated 3 months after treatment as the primary time point for analysis. The latter trial was the definitive test of the efficacy of t-PA at 3 months. The formal trial was confirmatory. To conduct subgroup analyses for this report, the two parts were combined because (1) all methods, including patient recruitment, treatment, and follow-up, were identical in the two parts; (2) the results of part 1 were not presented to any investigators until part 2 was completed, thus preserving the double-blind nature of part 2; and (3) the combined study group allows for greater statistical power. A total of 624 patients were randomized (312 t-PA–treated, 312 placebo-treated subjects). The subgroup analyses are presented as exploratory since the trial was not specifically designed for these post hoc analyses. Such exploratory analyses are useful in generating hypotheses for future studies and rarely may help guide physicians in subselecting patients for therapy.
Selection of Variables
We chose variables for study because they (1) would be available before treatment in most patients, (2) might influence outcome after stroke, or (3) might influence clinical response to t-PA either by magnifying or reducing t-PA efficacy or harm. The resultant 27 baseline variables selected for study are listed in Table 1⇓, and definitions of each are included in Appendix 2. For this study we defined outcome as the odds of an excellent recovery 3 months after stroke. Clinical response to t-PA was defined as the statistical interaction between treatment (t-PA or placebo) and each variable. When these definitions are used, outcome may be viewed as the final status of the patient, and outcome may be influenced by any of the candidate variables or by the clinical response to t-PA. Variables that change outcome but do not influence the clinical response to t-PA should not be used to select patients for treatment since, by definition, outcome is independent of treatment status. On the other hand, if a variable influences clinical response to t-PA, that variable could be used to guide subselection of patients for t-PA treatment.
From the 27 baseline variables, a multivariable model of favorable outcome was constructed. We assessed favorable outcome using the global test for multiple outcomes, as previously described.5 The global test for multiple outcomes is a summary statistic derived from the proportion of patients in each group (t-PA or placebo) scoring either normal or nearly normal on four clinical rating scales (Barthel Index ≥95, Modified Rankin Scale <1, Glasgow Outcome Scale <1, and NIHSSS <1) determined 3 months after stroke.5 6 Although the global test for multiple outcomes was not used in stroke research before the NINDS t-PA Stroke Trial, the method is well established in the statistical literature (for review, see Tilley et al5 ). This method is well suited for the present analysis because traditional methods of subgroup analysis would require multiple models for each outcome.7 In addition, use of the global test for multiple outcomes can increase the power to detect effects.5 However, because the analysis used generalized estimating equations, traditional stepwise procedures were not available. Instead, a modified step-down screening approach was used to construct the multivariable model.8 Details of the model-building procedure are given in Appendix 3.
We chose to include variables in our models using probability values of .2, .1, and .05 at different steps in the model-fitting process to increase our power to detect associations.9 When hypotheses are tested, type 1 errors are of concern, and therefore there is often an adjustment to the critical value for multiple comparisons (making α smaller) to protect against falsely rejecting the null hypothesis. In exploratory analyses, such as in this report, the goal is to avoid type II errors. Thus, we increased the critical value (made α larger) to increase the power to detect important subgroup effects. This approach is used for screening of variables to be included in a multivariable model.8 To assess the power of our model to detect significant interactions between treatment and variables, we performed power calculations twice, once for the earlier steps with more variables and again for the later steps with fewer variables. For each of the 3-month outcome scales (NIHSS, Barthel, Rankin, and Glasgow), we calculated the adjusted R2 of the final logistic model (step 5). To calculate the power, we used the smallest R2 found and a sample size based on the number of patients with complete observations on all covariates in the model.10
Table 1⇑ details the results of the univariate analyses including testing of interactions with t-PA treatment. These are the results of steps 1 and 2 of the model-building process. Treatment with t-PA was strongly associated with favorable outcome (OR, 1.86; Table 1⇑). By examining the univariate ORs, and using a critical probability value of .2, we found that 3-month outcome was associated with age, sex, smoking, problem drinking, diabetes, hypertension, atherosclerosis, atrial fibrillation, other cardiac disease, stroke subtype, baseline NIHSSS, presence of thrombus or early signs of infarction (hypodensity or midline shift) on baseline CT scan, admission and baseline blood pressures, and admission temperature. When we again used a significance of P<.2 as the criterion, race, diabetes, hypertension, baseline MAP, and baseline systolic blood pressure showed a significant interaction with t-PA treatment (Table 1⇑).
The selected variables and interactions were included in a multivariable model, and interactions among confounding variables were identified in steps 3 and 4 of the process. Age-by-baseline NIHSSS and age-by-MAP interactions were detected (ie, patients with older age and higher baseline NIHSSS or older age and higher admission MAP were less likely to have a favorable outcome), and the combined interaction term was therefore included in the final steps of building the model. At step 5, we determined that none of the two-way interactions significantly interacted with treatment. Table 2⇓ shows the final multivariable model after all 5 steps were completed. Treatment with t-PA remained strongly and independently associated with favorable outcome (OR based on multiple outcome, 2.02; P=.0001). Only age-by-NIHSSS interaction, diabetes, admission MAP-by-age interaction, and thrombus or hypodensity/mass effect on baseline CT scan were independently associated with favorable outcomes in this study. None of these terms, however, had a significant interaction with t-PA treatment (from step 5). That is, each of the variables and interactions in Table 2⇓ significantly influenced outcome, but none of them influenced the likelihood of differential response to t-PA. The proportions of favorable outcomes (NIHSSS <1 at 3 months) for subgroups of these variables are shown in Table 3⇓. The subgroups were created from the overall distribution with the use of quartiles. In nearly all subgroups, the proportion of patients with favorable outcome was greater in the t-PA–treated group. In the 49 patients aged >75 years and admission NIHSSS >20, there appeared to be no favorable response to treatment. However, closer evaluation of this subgroup, including analysis of outcome categories such as mild or moderate, suggested a treatment benefit (Figure⇓).
For the first power analysis, we noted that in steps 1 and 2 we included 22 variables and that the model adjusted R2 was .38. We found 496 subjects for whom we had complete data on all 22 of the variables. When we assumed α=0.2 in steps 1 and 2, there was a power greater than 90% to detect interactions that explain as little as 1.3% of the variability. In steps 4 and 5, we included 9 variables, the model adjusted R2 was .34, and we found 599 patients with complete data for all 9 variables. When we assumed α=0.1, there was a power greater than 90% to detect interactions explaining as little as 1.4% of the variability in the model.
We examined the NINDS t-PA Stroke Trial data set for indicators of differential response to thrombolysis. We could find no pretreatment information that could be used to predict a differential response to t-PA treatment. We found that t-PA treatment was independently and strongly associated with increased likelihood of favorable outcome 3 months after stroke.1 The absence of an interaction with treatment, after other explanatory variables are included in the multivariable model, suggests a persistent beneficial effect of t-PA treatment across all subgroups tested. Although tempting, it is fallacious to select a subgroup from Table 3⇑ in which the proportion of favorable outcomes was lower in the t-PA group and conclude that t-PA is not beneficial for that subgroup.7 The broader trend of t-PA benefit demonstrable across all subgroups, with the use of regression methods, makes the single isolated aberration within one category of a subgroup more likely to be a random occurrence related to small sample size within that category of the subgroup than a clinically meaningful, biological trend.7 That is, we found no evidence to justify withholding t-PA from any of the subgroups we studied.
This report contains multiple exploratory analyses. Exploratory analyses are useful for generating hypotheses for future studies, each of which must be properly designed and powered.7 Such analyses are limited in their applicability to patient care, however, because of the risk of both type II errors and spurious associations due to repetitive significance testing. Occasionally, such explorations yield useful clinical guidelines, and it was in this hope that these analyses were conducted. Also, the primary report has been criticized for not offering more detailed subgroup analyses.4 The NINDS t-PA Stroke Trial was designed with sufficient power to address the primary hypothesis, not the subgroup analyses. Nevertheless, the power analyses indicate that there were enough subjects available to detect significant treatment interactions with a power greater than 90%. This suggests that if such an interaction were present, there is a greater than 90% probability that we would have found it.
Numerous data items were collected before treatment. We chose 27 that were most plausible biologically to be related to either outcome or differential treatment response. There could be other variables that are associated with outcome or treatment response, although we could find no literature supporting inclusion of any other variables. In prior studies, the variables most commonly associated with long-term outcome after stroke include age, severity, and presence of comorbid risk factors (especially smoking, diabetes, heart disease, and hypertension).11 12 13 Therefore, it is unlikely that we missed an important variable. Our study is the first to include only patients who present within 3 hours of symptom onset. Prior studies of patients who present later after stroke generally agree with ours; long-term outcome is associated with the variables listed in Table 1⇑. Also, as is true for any stepwise regression procedure, the step-down procedure could potentially miss a significant association. This must be recognized as a potential limitation of these results.
Of most importance is the finding that no patient subgroups with differential response to t-PA could be identified. It is clear that those with more severe deficits at baseline are less likely to do well over the long term, but t-PA was effective for patients with severe as well as moderate deficits (Table 2⇑ and Figure⇑). We could not identify a threshold value of age, NIHSSS, or any particular stroke subtype that precludes t-PA treatment. Therefore, when considering older, sicker patients for possible t-PA therapy, physicians should keep in mind that such patients are more likely to do poorly but that t-PA may offer some potential benefit (Figure⇑). Further studies of this group are needed.
This study illustrates the utility of the statistical concepts of association and interaction. These concepts have not been used before in clinical stroke trials, although they are well established in the statistical and cardiovascular research literature.7 8 These statistical ideas help answer the most important question to the potential treating physician, “Can I select my patients more carefully and restrict use of t-PA to certain subgroups of patients?” The rigorous statistical approach to answer this question utilized a test of the association of the interaction term by means of a global test. A significant interaction between a variable and treatment implies that treatment benefit differs in the different subgroups identified by that variable. Tables 1 through 3⇑⇑⇑ demonstrate that the answer to the potential treating physician is, “Based on post hoc analyses of combined data from two trials and the positive association of treatment with a favorable outcome in each of the two trials, we would not recommend withholding t-PA from any subgroup of patients at this time.1 ” While the power was high to detect significant interaction terms, additional experience will be required to further reduce the chances of type II errors in these types of analyses.
Selected Abbreviations and Acronyms
|MAP||=||mean arterial blood pressure|
|NIHSSS||=||National Institutes of Health Stroke Scale score|
|NINDS||=||National Institute of Neurological Disorders and Stroke|
|t-PA||=||tissue plasminogen activator|
The following persons and institutions participated in the NINDS rt-PA Stroke Trial:
Clinical Centers: University of Cincinnati (n=150)—Principal Investigator: T. Brott; Co-investigators: J. Broderick, R. Kothari; M. O’Donoghue, W. Barsan, T. Tomsick; Study Coordinators: J. Spilker, R. Miller, L. Sauerbeck; Affiliated Sites: St. Elizabeth (South), J. Farrell, J. Kelly, T. Perkins, R. Miller; University Hospital, T. McDonald; Bethesda North Hospital, M. Rorick, C. Hickey; St. Luke (East), J. Armitage, C. Perry; Providence, K. Thalinger, R. Rhude; The Christ Hospital, J. Armitage, J. Schill; St. Luke (West), P.S. Becker, R.S. Heath, D. Adams; Good Samaritan Hospital, R. Reed, M. Klei; St. Francis/St. George, A. Hughes, R. Rhude; Bethesda Oak, J. Anthony, D. Baudendistel; St. Elizabeth (North), C. Zadicoff, R. Miller; St. Luke–Kansas City, M. Rymer, I. Bettinger, P. Laubinger; Jewish Hospital, M. Schmerler, G. Meiros.
University of California, San Diego (n=146)—Principal Investigator: P. Lyden; Co-investigators: J. Dunford, J. Zivin; Study Coordinators: K. Rapp, T. Babcock, P. Daum, D. Persona; Affiliated Sites: UCSD, M. Brody, C. Jackson, S. Lewis, J. Liss, Z. Mahdavi, J. Rothrock, T. Tom, R. Zweifler; Sharp Memorial, R. Kobayashi, J. Kunin, J. Licht, R. Rowen, D. Stein; Mercy Hospital, J. Grisolia, F. Martin; Scripps Memorial, E. Chaplin, N. Kaplitz, J. Nelson, A. Neuren, D. Silver; Tri-City Medical Center, T. Chippendale, E. Diamond, M. Lobatz, D. Murphy, D. Rosenberg, T. Ruel, M. Sadoff, J. Schim, J. Schleimer; Mercy General, Sacramento, R. Atkinson, D. Wentworth, R. Cummings, R. Frink, P. Heublein.
University of Texas Medical School, Houston (n=104)—Principal Investigator: J.C. Grotta; Co-investigators: T. DeGraba, M. Fisher, A. Ramirez, S. Hanson, L. Morgenstern, C. Sills, W. Pasteur, F. Yatsu, K. Andrews, C. Villar-Cordova, P. Pepe; Study Coordinators: P. Bratina, L. Greenberg, S. Rozek, K. Simmons; Affiliated Sites: Hermann Hospital, St. Lukes Episcopal Hospital, Lyndon Baines Johnson General Hospital, Memorial Northwest Hospital, Memorial Southwest Hospital, Heights Hospital, Park Plaza Hospital, Twelve Oaks Hospital.
Long Island Jewish Medical Center (n=72)—Principal Investigators: T.G. Kwiatkowski (6/92-), S.H. Horowitz (12/90-5/92); Co-investigators: R. Libman, R. Kanner, R. Silverman, J. LaMantia, C. Mealie, R. Duarte; Study Coordinators: R. Donnarumma, M. Okola, V. Cullin, E. Mitchell.
Henry Ford Hospital (n=62)—Principal Investigator: S.R. Levine; Co-investigators: C.A. Lewandowski, G. Tokarski, N.M. Ramadan, P. Mitsias, M. Gorman, B. Zarowitz, J. Kokkinos, J. Dayno, P. Verro, C. Gymnopoulos, R. Dafer, L. D’Olhaberriague; Study Coordinators: K. Sawaya, S. Daley, M. Mitchell.
Emory University School of Medicine (n=39)—Principal Investigator: M. Frankel (7/92-10/95), B. Mackay (11/90-6/92); Co-investigators: J. Weissman, J. Washington, B. Nguyen, A. Cook, H. Karp, M. Williams, T. Williamson; Study Coordinators: C. Barch, J. Braimah, B. Faherty, J. MacDonald, S. Sailor; Affiliated Sites: Grady Memorial Hospital, Crawford Long Hospital, Emory University Hospital, South Fulton Hospital, M. Kozinn, L. Hellwick.
University of Virginia Health Sciences Center (n=37)—Principal Investigator: E.C. Haley, Jr; Co-investigators: T.P. Bleck, W.S. Cail, G.H. Lindbeck, M.A. Granner, S.S. Wolf, M.W. Gwynn, R.W. Mettetal, Jr, C.W.J. Chang, N.J. Solenski, D.G. Brock, G.F.Ford; Study Coordinators: G.L. Kongable, K.N. Parks, S.S. Wilkinson, M.K. Davis; Affiliated Sites: Winchester Medical Center, G.L. Sheppard, D.W. Zontine, K.H. Gustin, N.M. Crowe, S.L. Massey.
University of Tennessee (n=14)—Principal Investigator: M. Meyer (2/93-), K. Gaines (11/90-1/93); Study Coordinators: A. Payne, C. Bales, J. Malcolm, R. Barlow, M. Wilson; Affiliated Sites: Baptist Memorial Hospital, C. Cape; Methodist Hospital Central, T. Bertorini; Jackson Madison County General Hospital, K. Misulis; University of Tennessee Medical Center, W. Paulsen, D. Shepard.
Coordinating Center: Henry Ford Health Sciences Center—Principal Investigator: B.C. Tilley; Co-investigators: K.M.A. Welch, S.C. Fagan, M. Lu, S. Patel, E. Masha, J. Verter; Study Coordinators: J. Boura, J. Main, L. Gordon; Programmers: N. Maddy, T. Chociemski; CT Reading Centers: Part A: Henry Ford Health Sciences Center, J. Windham, H. Soltanian Zadeh; Part B: University of Virginia Medical Center, W. Alves, M.F. Keller, J.R. Wenzel; Central Laboratory: Henry Ford Hospital, N. Raman, L. Cantwell; Drug Distribution Center: A. Warren, K. Smith, E. Bailey.
NINDS—Project Officer: J.R. Marler.
Data and Safety Monitoring Committee—J.D. Easton, J.F. Hallenbeck, G. Lan, J.D. Marsh, M.D. Walker.
Genentech Participants—Juergen Froelich, MD, Judy Breed, Fong Wang-Chow.
Definitions and Data Collection for Patient Subgroup Variables
1. Age: Recorded at entry.
2. Baseline NIHSSS: Performed by trained and certified examiner just before treatment with study drug.5
3. Weight: Actual first weight obtained after admission.
4. Percentage of correct dose: Protocol-specified dose of 0.9 mg/kg usually was based on estimated weight; this variable was calculated from actual dose administered compared with projected dose using actual measured weight.
5. Admission MAP: Calculated MAP obtained from first recorded blood pressure in participating center’s emergency department.
6. Admission systolic blood pressure: Taken from first recorded blood pressure in participating center’s emergency department.
7. Admission diastolic blood pressure: Taken from first recorded blood pressure in participating center’s emergency department.
8. Baseline MAP: Calculated MAP obtained at baseline just before study drug infusion.
9. Baseline systolic blood pressure: Recorded at baseline just before study drug infusion.
10. Baseline diastolic blood pressure: Recorded at baseline just before study drug infusion.
11. Time from stroke onset to treatment (start time): Derived from recording time of initiation of treatment with study drug and stroke onset time obtained from history.
12. Admission temperature: Taken from the first recorded temperature in participating center’s emergency department.
13. Race: Recorded at patient entry as black, white (non-Hispanic), Asian, Hispanic, or other. Only 37 Hispanics, 18 Asians, or others were enrolled (5.9%).
14. Sex: Male, female.
15. Smoking in previous year: Analyzed as either “Yes” or “No”; from patient or family history.
16. Drinking problems: “Yes” or “No” based on CAGE criteria.14
17. History of diabetes: Analyzed as either “Yes” or “No”; history from either medical record, patient, or family.
18. History of hypertension: Analyzed as either “Yes” or “No”’ based on history from medical record, patient, or family. Elevated blood pressure on admission was not sufficient to make diagnosis.
19. History of atherosclerosis: Analyzed as “Yes” or “No.” Derived from history of symptomatic coronary, peripheral, or cerebrovascular disease in medical record or from patient or family history.
20. History of atrial fibrillation: “Yes” or “No” from medical record, patient, or family history.
21. History of other cardiac disease: “Yes” or “No” from medical record, patient, or family history. Included specific inquiries regarding angina, congestive heart failure, myocardial infarction, and valvular heart disease.
22. Prior stroke: “Yes” or “No” from medical record, patient, or family history. Required focal neurological deficits lasting >24 hours.
23. Aspirin (NSAID): “Yes” or “No” if aspirin or any other NSAID was consumed within the previous 2 weeks.
24. Baseline stroke subtype: Clinical assessment based on initial history, physical examination, and baseline head CT scan. Coded as large-vessel occlusive, small-vessel occlusive, cardioembolic, or other, using the treating physician’s clinical estimate before randomization.
25. Early CT findings with (without) thrombus: “Yes” or “No” based on presence or absence of early signs of cerebral infarction on the baseline head CT scan. Early findings included frank acute hypodensity or midline shift. Scans were also examined for the presence of hyperdense acute thrombus in leptomeningeal arteries (dense artery sign).15 All baseline CT scans were reviewed by a single neuroradiologist blinded to clinical data, treatment group, and results of subsequent head CT scans.
26. Centers: Each of the participating centers (with all their participating hospitals) was considered a categorical variable in modeling outcome.
Step 1. To ensure that all possible candidate variables would be included in the preliminary steps, a liberal screening value of .2 was chosen. This choice keeps more variables “in the running” for possible inclusion in the final multivariable model.
Step 2. To test for variable treatment interactions, ie, to determine whether each variable was associated with a differential response to t-PA treatment, separate models of interaction were fit. Each model included one of the baseline variables, a variable representing treatment, and a variable representing the interaction between the two. If the interaction term probability value was <.2, then the variable and interaction term were used in the next step. Again, the liberal value of .2 was chosen to ensure that interactions were not prematurely rejected during the model-building process.
Step 3. To remove variables and interaction terms that might be redundant or explained by other variables in the model, a multivariable model was fit with the individual variables selected in steps 1 and 2. At this step, individual variables were retained in the model if associated probability values in the multivariable model were <.05, and interactions were retained if the associated probability values were <.1. These significance values were chosen in accordance with traditional multivariable regression methods.8
Step 4. To determine whether there were two-way interactions among the selected variables, all two-way interactions among the remaining variables (except treatment) were tested (eg, age-by-diabetes, age-by-NIHSSS, NIHSSS-by-diabetes) in a multivariable model. If the associated probability values were <.1, then the interaction was retained for the next step. On the other hand, if the probability value was >.1, then the interaction was not included since there was no evidence that the two variables interacted in a way that affected outcome. This analysis of the interactions was necessary to exclude spurious associations due to relationships among variables that had nothing to do with treatment.
Step 5. To test for three-way interactions, any selected two-way interactions from step 4 were tested as three-way interactions with treatment and three-way interactions among selected variables. Three-way interactions were retained in the final model if the probability values were <.1. If any of these final three-way interactions with t-PA treatment were significant at P<.1, then a clinical response of that variable with t-PA treatment was judged to be present.
This study was supported by NINDS grants NO1-NS02382, N01-NS02374, N01-NS02377, N01-NS02381, N01-NS02379, N01-NS02373, N01-NS02376, N01-NS02378, and N01-NS02380. We would like to gratefully acknowledge the forbearance of our families during the conduct of the NINDS t-PA Stroke Trial.
The persons and institutions that participated in this trial are listed in Appendix 1.
Almost all of the investigators of the current study are on the speaker panel for Genentech, Inc. Several of the investigators are consultants for an ongoing study of t-PA administered at 3 to 5 hours after onset that is funded by Genentech (ATLANTIS study). Joseph Broderick is a consultant for Genentech in the ATLANTIS study.
- Received June 11, 1997.
- Revision received August 5, 1997.
- Accepted August 5, 1997.
- Copyright © 1997 by American Heart Association
Report of the Quality Standards Subcommittee of the American Academy of Neurology: Thrombolytic therapy for acute ischemic stroke: summary statement. Neurology. 1996;47:835-839.
Adams HP Jr, Brott TG, Furlan AJ, et al. Guidelines for thrombolytic therapy for acute stroke: a supplement to the guidelines for the management of patients with acute ischemic stroke. Circulation. 1996;94:1167-1174.
Furlan A, Kanoti G. When is thrombolysis justified in patients with acute ischemic stroke? A bioethical perspective. Stroke. 1997;28:214-218.
Tilley BC, Marler J, Geller NL, Lu M, Legler J, Brott T, Lyden P, Grotta J. Use of a global test for multiple outcomes in stroke trials with application to the National Institute of Neurological Disorders and Stroke t-PA Trial. Stroke. 1996;27:2136-2142.
Lyden P, Brott T, Tilley B, Welch KMA, Mascha EJ, Levine S, Haley EC, Grotta J, Marler J, NINDS t-PA Stroke Study Group. Improved reliability of the NIH Stroke Scale using video training. Stroke. 1994;25:2220-2226.
Hosmer DW, Lemeshow S. Applied Logistic Regression. New York, NY: John Wiley & Sons; 1989.
Cohen J, Cohen P. Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences. 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates; 1983.
Nagelkerke N. A note on a general definition of the coefficient of determination. Biometrika. 1991;78:691-692.
Carlberg B, Asplund K, Hägg E. The prognostic value of admission blood pressure in patients with acute stroke. Stroke. 1993;24:1372-1375.
Loewen SC, Anderson BA. Predictors of stroke outcome using objective measurement scales. Stroke. 1990;21:78-81.
Katz S, Ford AB, Chinn AB, Newill VA. Prognosis after strokes, II: long-course of 159 patients. Medicine. 1966;454:236-246.
Tomsick TA, Brott TG, Chambers AA, Fox AJ, Gaskill MF, Lukin RR, Pleatman CW, Wiot JG, Bourekas E. Hyperdense middle cerebral artery sign on CT: efficacy in detecting middle cerebral artery thrombosis. AJNR Am J Neuroradiol. 1990;11:473-477.