class: center, middle, inverse, title-slide .title[ # Clinical Trials ] .author[ ### Yue Jiang ] .date[ ### Duke University ] --- ### Why clinical trials? <img src="img/trials/totalcv.jpg" width="60%" style="display: block; margin: auto;" /> - .vocab[Prevalence]: Total number of cases at a given point in time - .vocab[Incidence]: Number of *new* cases over time .question[ What can we tell from this chart? ] .small[Mozaffarian D, et al. on behalf of the American Heart Association, Circulation (2015)] --- ### Why clinical trials? <img src="img/trials/estradiol.png" width="80%" style="display: block; margin: auto;" /> --- ### Why clinical trials? Estrogen levels: - Pre-menopausal women: 20 - 400 pg/mL - Post-menopausal women: 5 - 25 pg/mL - Men: 10 - 60 pg/mL **Theory**: high estrogen levels are protective for heart disease --- ### Potential study designs .vocab[Cross-sectional study]: measure estrogen and heart disease at the same point in time - Three studies measured atherosclerosis via angiogram, accompanied by asking women about estrogen use - In all three studies, women taking estrogen had less artery blockage than women who were not taking estrogen .vocab[Case-control study]: compare the estrogen use of women who have experienced heart attacks against a matching set of women who haven't - 5 studies used hospital-based controls (controls in the hospital for non-CVD reasons) and 6 studies used population-based controls - Studies with hospital-based controls found no association - However, studies with population based controls did find less estrogen use among heart attack victims .small[Stampfer MJ and Colditz GA, Prev. Med. (1991)] --- ### Potential study designs .vocab[Cohort study]: identify a large group of women who do not have heart disease and ask them about their estrogen history. Follow these women and observe which group has more heart attacks. - 16 cohort studies of heart disease and estrogen use in women were performed - 15 studies found that women taking estrogen had fewer heart attacks than non-estrogen users - Combining studies using meta-analysis, it was estimated that estrogen was associated with reduced risk of a heart attack of about 50\% (wow!) .small[Stampfer MJ and Colditz GA, Prev. Med. (1991)] --- ### Potential study designs > *“...the protection is biologically plausible and the magnitude of the benefit would be quite large...”* .small[Barrett-Connor and Bush, JAMA (1991)] Clearly, these studies strongly support the theory, and physicians were encouraged to put post-menopausal women on estrogen replacement therapy to avoid CVD. .question[ Do you think estrogen is protective for heart disease in women? ] --- ### Further studies **HERS**: The Heart and Estrogen/Progestin Replacement Study - 2,763 post-menopausal women with previous heart disease, women followed for 4.1 years - No statistically significant difference in heart attacks, bypass surgery, congestive heart failure, peripheral artery disease, or stroke **ERA**: The Estrogen Replacement and Athersclerosis Trial - 309 women with angiographically verified disease, women followed for 3.2 years - No statistically significant difference in artery diameters **WHI**: The Women's Health Initiative - Two trials: 16,608 women with an intact uterus and no history of CVD followed for 5.2 years, 10,739 women with hysterectomy and no CVD followed for 6.8 years - No beneficial effect of HRT on CV outcomes. In fact, **higher breast cancer risk** among women with intact uteruses. .small[Hulley et al., JAMA (1998), Herrington et al., NEJM (2000), Rossouw et al., JAMA (2002); Manson et al., NEJM (2007)] --- ### What happened? .question[ Why did the initial studies find such a strong effect, but further studies find no effect? ] -- > *“These important observations need to be confirmed in a double-blind, randomized clinical trial, since the protection is biologically plausible and the magnitude of the benefit would be quite large if selection factors can be excluded."* .small[Barrett-Connor and Bush, JAMA (1991)] > ...women who take HRT differ from those who don’t in many ways, virtually all of which associate with lower heart disease risk .small[Gary Taubes, NYT] --- ### Clinical trials Clinical trials are often .vocab[experimental studies] as opposed to .vocab[observational studies]. 1. Clinical trials have defined .vocab[subjects] 2. Clinical trials involve direct .vocab[intervention] 3. Clinical trials are .vocab[prospective] 4. Clinical trials are performed under conditions controlled by researchers, often involving a .vocab[placebo arm] and .vocab[randomization] of assigned treatment groups --- ### History of clinical trials <img src="img/trials/ship.jpg" width="40%" style="display: block; margin: auto;" /> --- ### History of clinical trials (also vocab terms) - 1848: Bartlett argued that no proof of efficacy could be obtained without a .vocab[control] group - 1865: Sutton describes first use of .vocab[placebo] (mint water for rheumatic fever) - 1904: Cushny and Peebles performed a .vocab[crossover] trial for a sleep drug - 1915: Greenwood and Yule describe .vocab[random allocation] of patients to treatment groups - 1926: Fisher formally introduces .vocab[randomization] - 1927: Ferguson et al. use .vocab[blinding] in their sutdy of cold vaccines --- ### Randomization Randomization eliminates conscious bias on behalf of the intervener; groups become alike *in expectation* (they may still be unbalanced in practice along certain covariates, but this is ok*). Importantly, with randomization, we may make .vocab[causal claims]. There are two competing goals in randomization: to keep imbalance low throughout and in the final allocation, and to keep predictability of treatment low. Variance of estimators may be lowered by .vocab[stratifying] on certain factors. .question[ Why might we want to lower the variability of our estimators? ] --- ### Blinding To further eliminate bias, randomized trials are often .vocab[blinded] or masked - .vocab[Open]: no one masked - .vocab[Single-blind]: participant masked - .vocab[Double-blind]: participant and intervener masked Randomization and blinding help eliminate *unconscious bias* in terms of care, mnagement, and evaluation .question[ How might we assess blinding? ] --- ### Two types of analyses .vocab[Intention-to-treat] (ITT): every patient randomized in the study will be analyzed, including those who are non-compliant. .vocab[Per-protocol]: only analyze patients that were fully compliant. .question[ Which approach is more conservative in terms of minimizing type I error rate? ] --- ### Speaking of type 1 error rate... The .vocab[Bonferroni-Holm] method 1. Sort p-values in ascending order `\(p_{(1)} < p_{(2)} < \cdots < p_{(m)}\)` 2. Compare `\(p_{(1)}\)` to `\(\alpha / m\)` (just like normal Bonferroni) 3. If significant, we've reduced our multiple testing problem by one test 4. Compare the next smallest p-value `\(p_{(2)}\)` to `\(\alpha/(m - 1)\)` 5. ...and so on; stop at the first non-significant p-value. The Bonferroni-Holm method also strongly controls family-wise error rate without being as conservative as the Bonferroni method <img src="img/trials/holm.png" width="80%" style="display: block; margin: auto;" /> --- ### Power and sample size analysis .question[ What is the definition of power? What affects it? ] .question[ Suppose you've been asked how many patients are needed for a survival outcome comparing two groups (let's say using a log-rank test). What information additional information would you need from them? How would you go about doing this calculation? ]