Associations, Predictions, Causes, and Interventions
Idea: Relationships among associations, predictions, causes, and interventions run through all the cases and controversies in this course. The idea introduced in this session is that epidemiology has two faces: One from which the thinking about associations, predictions, causes, and interventions are allowed to cross-fertilize, and the other from which the distinctions among them are vigorously maintained, as in "Correlation is not causation!" The second face views Randomized Control Trial (RCTs) as the "gold-standard" for testing treatments in medicine. The first face recognizes that many hypotheses about treatment and other interventions emerge from observational studies and often such studies provide the only data we have to work with. What are the shortcomings of observational studies we need to pay attention to (e.g., systematic sampling errors leading to unmeasured confounders-see next class)?
Guidelines for annotations
Notes and annotations from 2007 course
Initial notes from PT
Common readings and cases: Ridker 2007 (Cardiac risk factors), Stampfer 1991, 2004 (Hormone replacement therapy)
Ridker et al. show that the conventional risk factors for heart disease in women (as combined in the Framingham score) identify many women as of intermediate risk who are higher or lower risk. The new Reynolds Risk Score does a much better job, primarily it seems by including the risk marker cReactive Protein. Both scores are based on observations not randomized trials. (But see Shunkert for recent assessment of the role of CRP.)
The case of hormone replacement therapy as a protection against heart disease (Stampfer 1990) is another, more significant instance of mismatch of observational results and RCTs -- see Stampfer 2004 & Pettiti for analyses of the discrepancy.
Try to get a handle on the different kinds of explanation for this and other discrepancies, including physician bias in who gets prescribed a treatment, residual confounders, and reverse causation. Gordis may help here.
Supplementary Reading: Alzheimer Research Forum 2004, Davey-Smith & Ebrahim 2007,pp2-8, Jick 2000, Petitti 2004, Shunkert 2008
Jick presents evidence that statin treatment was associated with lowered risk of dementia but the Alzheimer Research Forum presents the more recent assessment (using RCTs) that statins are not protective against dementia. The discrepancy seems to be undetected bias in which patients get prescribed statins.
Davey-Smith & Ebrahim provide a quick review of a number of cases.
Annotations on common readings
Davey-Smith and Ebrahim
First I started with the Petitti reading, which I felt was too difficult (and I’m not entirely sure it was the right one). Then, I tried the Alzheimer’s reading which was uninteresting to me. So I moved on to the first part of Davey-Smith and Ebrahim which is pointing to an interesting theme in this weeks offering for supplemental readings.
The theme, it appears to me, is a cautionary tale of causality. From these readings I can see that much research happens to research the wrong things, or looks at the wrong causes and much time and resources are spent on these errant studies. What Davey-Smith and Ebrahim seem to be pointing to is that the job of the epidemiologist is to look at all possible solutions to the puzzle in front of them, because even the one that looks most likely or obvious might be wrong and a waste of time. They warn about reverse causality (and give an example of cancer lowering a person’s blood pressure years before the cancer is detected, whereas it could be interpreted that people with low blood pressure are more likely to get cancer).
The authors also note that it is important to take all sorts of possible confounding variables into account. And to look at the data collection system which can, either on purpose or by accident, be hypothesis affirming if the epidemiologist hasn’t properly exhausted all possibilities. Bias plays an important role as well and needs to be considered in study design and result interpretation. (MC '09)
Paul Ridker, Development and Validation of Improved Algorithms
for the Assessment of Global Cardiovascular Risk in Women. The Reynolds Risk Score
Ridker’s article provides readers with a comprehensive overview of a study that was initiated in 1992 “to develop and validate cardiovascular risk algorithms for women based on a large panel of traditional and novel risk factors.” Ridker et al gives a step by step account of how, and with whom the study was conducted.
Healthy women were studied in an effort to determine their future risk for various diseases such as atherothrombosis, stroke, myocardial infarction, and so on. Participants in the study were derived from the Women’s Health Study (WHS). The women selected had to meet certain health requirements in order to be included in the study; “Women eligible for the current analysis were those who provided an adequate baseline plasma sample”, “and had complete ascertainment of all blood covariates of interest.”
Researchers divided the women into two groups, Group A and Group B, and risk prediction algorithms and outcomes were measured using Entropy, the Yates slope, and the Brier score. These women were followed over a median of 10.2 years for Group A, and approximately 8 years for Group B. As an additional measure, researchers looked at two “components of accuracy: discrimination and calibration.”
The article goes on to further illustrate the methods of testing used; it provides summary statistics, the variables that were used in calculations, and the resulting outcome of the study in its entirety.
As a result of this study, researchers were able to develop, confirm and show greater “accuracy of two clinical algorithms for global cardiovascular risk prediction that classified 40 percent to 50 percent of women at intermediate risk into higher – or lower-risk categories.” These results could bode well for the prevention of cardiovascular disease in women. (CH 09)
Estrogen replacement therapy and coronary heart disease: a quantitative assessment of the epidemiologic evidence - Stampfer & Colditz
This article begins with a discussion of post-menopausal estrogen use and its ability to diminish the risk of developing coronary heart disease (CHD) in women.
An overview of the different designs and methods used to study the effects of estrogen on CHD are given, and the circumstances by which final conclusions were drawn. This was done in several ways: computer searches, article reviews, cohort and cross-sectional angiography studies, among others. There were several variables that affected results, such as whether or not cohorts had existing conditions like osteoporosis that could also be treated with estrogen; age was another important variable considered.
Five hospital-based case-control studies provided results that were inconsistent; but, these are difficult to interpret because of problems in selecting appropriate controls. Six population-based case-control studies revealed decreased relative risks in estrogen users, but only one was significant statistically. Three cross-sectional studies of women with or without stenosis on coronary angiography clearly showed less atherosclerosis among estrogen users. Fifteen out of 16 studies found decreased relative risks, most of which were statistically significant.
The Framingham study saw an increased risk, which was not statistically significant when angina was eliminated. Additional analysis of the data showed an insignificant protective effect among younger women and a non-significant increase in risk among older women. The majority of evidence suggests a protective effect of estrogens and can not be readily explained by confounding factors. This benefit is consistent with the effect of estrogens on lipoprotein subfractions.
Age has been suggested as a potential modifier of the estrogen effect, especially since a trend toward benefit was observed in the Framingham study for younger but not older women. Benefits at all ages in the studies were observed. It was found that there was a moderate increase in protection among younger women. However, the opposite was found in both studies; all age groups experienced a clear benefit. Substantial benefits were observed in a population with a median age of 73.
The studies indicated support for the view that postmenopausal estrogen therapy may significantly reduce the risk of CHD. It was found that in cohort and angiographic studies, the finding were more consistent. Overall, the various studies and trials have shown that the relative risk was dependent upon the different variables present during the course of analysis. But the general agreement is that there is a low relative risk of developing CHD in users of estrogen. (CH 09)
Annotated additions by students
Jick et al Article and Breitner Response
The published research by Jick et al regarding the application of statins in reducing dementia appears to be a nice straightforward case-control study that establishes causality. This, despite the fact that statins were prescribed originally to lower cholesterol levels, thereby reducing risk of heart disease. Certainly, the Lancet is the gold standard among epidemiological publications and its publication of this study serves to further bolster its findings and assertions. However, the case that is so neatly laid out is seriously challenged four years later, at a conference where new findings are presented. Breitner shares three studies that contradict the original study, and identifies methodology as the culprit. Due to the timing of the statin drug exposure and the subsequent time involved in observing outcomes, case-control studies such as Jick et al conducted were indeed pointing to causation, while prospective studies found no such evidence.The point is that association and causality should not be presumed lightly, and a different methodological approach to the same research subject, or questions, may uncover very different findings. (AH 10.06.09)
