Classification and Bias of Clinical Research, with a Randomized Clinical Trial Case Study

1. Classification and Bias of Clinical Research, with a Randomized Clinical Trial Case Study Rick Chappell, Ph.D. Professor, Department of Biostatistics and Medical Informatics University of Wisconsin Medical School

2. Good Ethics is Good Science: “If a research study is so methodologically flawed that little or no reliable information will result, it is unethical to put subjects at risk or even to inconvenience them through participation in such a study. … Clearly, if it is not good science, it is not ethical.” - U.S. Dept. of Health and Human Services, Policy for Protection of Human Subjects (45 CFR 46, 1/1/92 ed.)

3. Types of Studies Classified by Temporal Point of View • I. Instantaneous Studies - Surveys • II. Longitudinal Studies • A. Retrospective Studies • Historical Observational Cohort • Case - Control • B. Prospective Studies • Prospective Observational Cohort • Clinical Trial • C. Hybrid Designs

4. A Schematic for Temporal Classification Retrospective Prospective Observational Cohort Observational Cohort Randomization Case - Control Clinical Trial Now Instantaneous: Survey

5. I. Instantaneous:Population-Based Studies • Synonyms • Survey • Population-Correlation Study • Ecological Study • Two or more populations are instantaneously compared through the prevalences of both exposure and disease. • As summarized units get smaller (country  region  neighborhood  individual), a survey approaches a historical observational cohort study.

6. Population-Based Studies Advantages Instantaneous. Easy access to a large and varied population. Good for hypothesis generation. Disadvantages Intervention is usually not feasible. Very little information on causality: IARC standards require individual-based evidence.

9. II. Longitudinal:Individual-Based Studies • A longitudinal study observes exposures and events for individuals over a period of time. • There are two types, depending on whether one is looking forwards (prospective) or backwards (retrospective) from the present.

10. Longitudinal Studies:A. Retrospective • Historical Observational Cohort • Synonyms- survey, retrospective cohort study. • Examines outcomes among patients with past exposures. • E.g., track down 1950s asbestos miners & determine current status. • Case - Control (Breslow and Day, 1980) • Synonyms - case referent, retrospective study. • Examines past exposures among a group of patients with current outcomes. • E.g., interview mesothelioma patients & determine past exposures.

11. Historical Observational Cohort Studies Advantages Quick results - no wait. Easy to get large samples by ‘mining’ databases. Yields wide range of sequelae. Useful for investigating raretreatments orexposures. Disadvantages No opportunity to customize data collection. No possibility for blinding. Many possible biases: Confounding Selection Information

12. Case - Control Studies Advantages Cheap, quick - record searching can be automated. Useful for pilot studies. Useful for investigating rare disorders. Disadvantages Gives narrow picture of risks due to treatment or exposure. Biases: Confounding Selection Recall Yields only estimates of relative, not absolute risk.

13. Hypothetical Historical Cohort Study Exposed Group 100 Patients 10 Events Rate = .1 Odds Ratio 2 Control Group 100 Patients 5 Events Rate = .05

14. Hypothetical Case-Control Study Event Group 100 Patients 10 Exposures Event Rate per Exposure = ? (Not 100/200). Non-Event (Control) Group 100 Patients 5 Exposures Odds Ratio 2

15. Longitudinal Studies:B. Prospective • General Advantages • Can collect detailed exposure, treatment, disease, and demographic information. • Blinding is possible. • Recall and information bias may be eliminated. • Useful for investigating raretreatmentsor exposures. • Classification depends on the presence of intervention.

16. Prospective Studies • Prospective Observational Cohort • Synonyms - prospective trial, ‘clinical trial’. • No intervention. • Randomized Controlled (“Phase III”) Clinical Trial • Synonyms - prospective interventional cohort study, experiment, prospective trial, clinical trial. • Experimenters directly intervene in patient treatment, usually on a randomized basis with controls.

17. Prospective Observational Cohort Study Additional Advantage Passive observation; no need to dictate treatment. Disadvantages May take a long time to accrue cases and wait for results. Potential confounding bias due to lack of randomization and suitable controls.

18. Clinical Trials Additional Advantages “The most definitive tool for evaluation of the applicability of clinical research” - 1979 NIH release. Biases may be eliminated. Good design may make analysis simple. Disadvantages As above, may take a long time. Must be ethically and laboriously conducted. Requires treatment on basis (in part) of scientific rather than medical factors. Patients may make some sacrifice (Meier, 1982).

19. Phases of a Clinical Trial • Biochemical and pharmacological research. • Animal Studies (Gart, 1986 & Schneiderman, 1967). • Phase I (Storer, 1989) - estimate toxicity rates using few (~ 10 - 40) healthy or sick subjects. • Phase II (Thall & Simon, 1995) - determines whether a therapy has potential using a few very sick patients.

20. Phases of a Clinical Trial (cont.) • Phase III - large randomized controlled, possibly blinded, experiments • Phase IV - a controlled trial of an approved treatment with long-term followup of safety and efficacy.

21. Sackett’s Levels of Evidence for “Evidence-based Medicine” (Cook, et al., 2002) … continued

22. Levels of Evidence (cont.) … continued

23. Levels of Evidence (cont.)

24. A critique (by example) of clinical trials and evidence-based medicine “Parachute use to prevent death and major trauma related to gravitational challenge: systematic review of randomised controlled trials” (Smith & Pell, 2003). Fig. 1:

25. Abstract Objectives: To determine whether parachutes are effective in preventing major trauma related to gravitational challenge. Design: Systematic review of randomised controlled trials ... Main outcome measure: Death or major trauma, defined as an injury severity score > 15. Results: We were unable to identify any randomised controlled trials of parachute intervention. Conclusions: As with many interventions intended to prevent ill health, the effectiveness of parachutes has not been subjected to rigorous evaluation by using randomised controlled trials. Advocates of evidence based medicine have criticised the adoption of interventions evaluated by using only observational data. We think that everyone might benefit if the most radical protagonists of evidence based medicine organised and participated in a double blind, randomised, placebo controlled, crossover trial of the parachute.

26. Longitudinal Studies:C. Hybrid Designs • Prospective Treatment, Historical Controls • Currently treated series of patients is compared with a previous series. • See Gehan & Freireich (1974), Gehan (1984). • Advantages • Doesn’t assign treatments. • No need to recruit controls.

27. Longitudinal Studies:C. Hybrid Designs (cont.) • Prospective Treatment, Historical Controls • Disadvantages • Same as in Historical Observational Cohort except that characteristics of treated patients (only) can be collected. • Selection bias likely because of time lag between groups.

28. Hybrid Designs • Prospective Treatment with Both Prospective and Historical Controls • Uses both types of controls to maximize efficiency and minimize bias • See Pocock (1976a and 1976b).

29. Bias in Clinical Studies • Definition: Bias is a systematic error in estimation which is not reduced by increasing the study sample size (as opposed to random variation). • See Sacket (1979) and other articles in the same issue; Rose (1982); and Lachin (1988). • Classification is based on whether bias occurs at the time of patient Selection; or at the time of Information collection; or at the time of Publication. • They are all variants of Confounding, in which a third variable is related to both treatment and outcome.

30. I. Selection Bias • Prevalence - Incidence Bias • Prevalence (observed occurrence) of a trait  Incidence (rate of onset). • Cause: gap between exposure, selection of subjects. • Not a problem with irreversible events such as mortality, if detectable. • E.g., hypertension may disappear with onset of CV disease and can be overlooked as a risk factor. • See Neyman, 1955. • (Any retrospective study, especially case-control.)

31. Selection Bias • Admission Rate Bias • Patients may differ from noninstitutionalized subjects in size or direction of effects. • E.g., systemic weakness vs. arthritis: • Negative relation among inpatients; • Positive relation among outpatients. • See Berkson, 1946. • (Any nonrandomized study with a mix of patient sources, especially case-control.)

32. Selection Bias • Nonrespondant (Volunteer) Bias • Nonparticipation may be related to the subject of investigation. • E.g., smokers ignore surveys more often than do non-smokers (Seltzer, 1974). • For general methods to analyze data with ‘nonignorable nonresponse’ see Little and Rubin (1987) and Rubin (1987). • (Case-control, though drop-outs can effect any study not analyzed ‘intent to treat.)

33. Example: Where to add armor to fighter planes? In World War II, the U.S. Air Force conducted an investigation into where armor could most effectively be added to fighter planes. Researchers examined returning aircraft, mapped the locations of bullet holes, and recommended that the most commonly pierced areas be reinforced. Their recommendation neglected the most vital part of the aircraft, which was intact in all returning aircraft: the area surrounding the pilot’s head!

34. II. Information Bias • Detection Signal (Diagnostic Suspicion) Bias • In unblinded studies, an exposure may be considered a risk factor for an endpoint, and such patients preferentially observed. • In blinded studies, an exposure may make an endpoint more detectable. • E.g., estrogen causes bleeding from uterine cancer to be more easily detectable. • (Any unblinded study except case-control; also clinical trials with sensitive endpoints.)

35. Reports of Original Studies JAVMA191, 12/1/87 “High-rise syndrome in cats” Wayne O. Whitney, DVM & Cheryl J. Mehlhaff, DVM Selection and/or detection bias

36. Information Bias • Exposure Suspicion Bias • An outcome may cause the investigator to look for a particular exposure. • The temporal reverse of detection signal bias. • E.g., arthritis and knuckle-cracking. • (Case-control studies.)

37. Information Bias • Recall (family information) Bias • Similar to exposure suspicion bias, but errors originate with the subject or his/her family. • E.g., in a study of prescription use among women with fetal malformation, 28% reported unverifiable exposure vs. 20% of the controls (Klemetti & Saxen, 1967). • (Case-control studies.)

38. III. Publication (Reporting) Bias • Even a perfect study leads to bias if dissemination depends on the direction of its result. • Causes: • Commercial reasons; • Researchers’ personal motivations; • Editorial Policy ! • Vickers, et al. (1998) show that the problem is widespread: in some countries, 100% of publications show treatment effects.

39. Publication (Reporting) Bias • A version of the multiple comparisons problem (Miller, 1985), or ‘testing to a foregone conclusion’. • E.g., ORG-2766 protected nerves from cytotoxic injury in 55 women with ovarian cancer - NEJM lead article (van der Hoop, et al., 1990); a subsequent negative study of 133 women - ASCO Proceedings abstract (Neijt, et al., 1994). • (All Studies.)

41. A type of reporting bias: Multiple Comparisons (“Data Dredging”) • A “p-value” is interpreted as the probability of attaining a result as extreme that observed given that the result is false (under the null hypothesis); it can be viewed as the false positive rate under the null hypothesis. • This assumes that only a single test is conducted. If many tests are performed, it is possible to “sample to a foregone conclusion” and produce a falsely low p-value. • For example, if twenty-five independent tests are conducted, the probability of at least one p-value being less than .01 is .22. • Often only the significant result is reported, and the 24 others ignored.

46. IV. Confounding (General) • Caused by any situation in which: • A third variable exists which isn’t known or at least isn’t accounted for; • It is associated with the “cause” and • It is also associated with the “effect”. Then: • The supposed cause-effect relation will be confounded by the third variable. • (Any nonrandomized study)

47. Do Storks Bring Babies?

48. Population of Oldenburg, Germany, 1930-1936 (Ornithologische Monatsberichte44, Jahrgang, 1936, Berlin) Humans (1000s) Storks (1000s)

49. A Case Study in Randomized Clinical Trials: the Women’s Health Initiative Slides adapted from a presentation by Jacques E. Rossouw, National Heart, Lung, and Blood Institute.

50. Rationale for WHI: Prevention of Chronic Disease in Women • Clinical trials to test promising but unproven therapies • Hormone therapy to prevent CHD • Diet to prevent cancers • Calcium+Vitamin D to prevent fractures • Observational study to identify/quantify risk factors • Community intervention studies to test strategies for health promotion