«EPIDEMIOLOGY AND BIOSTASTICS: AN INTRODUCTION TO CLINICAL RESEARCH Bryan Kestenbaum, MD MS University of Washington, Seattle TABLE OF CONTENTS. ...»
An important, but difficult concept is the distribution of unmeasured factors that are not presented in the baseline characteristics table, for example exercise and dietary factors. In the interventional study design, if the sample size is reasonably large then random assignment will balance not only measured participant characteristics, but also unmeasured characteristics that do not appear in the baseline table. So, it is expected that women assigned to estrogen in the randomized trial will have similar distributions of exercise patterns and dietary characteristics as those assigned to placebo. In contrast, there is no easy way to predict whether unmeasured characteristics will be balanced in an observational study and increasing the sample size will have no effect on this uncertainty.
IV. INFERRING CAUSATION FROM ASSOCIATION STUDIES
Many associations are not causal. For example, receiving last rites in the intensive care unit is strongly associated with impending death; however, it is unlikely that receiving last rites causes a higher risk of death. Estrogen use has been associated with lower risks of heart disease in observational studies, but not in intervention trials, possibly because estrogen use is a marker of other healthy characteristics that are also linked with lower risks of heart disease, such as a healthier diet or compliance with other medical therapies. Separating association from causation in clinical research studies is critically important, because uncovering causative factors in a disease process can lead to prevention and treatment. For example, an inference that the prone sleeping position caused sudden infant death syndrome (SIDS) led to successful prevention strategies that substantially lowered the risk of SIDS.10 An inference that LDL cholesterol levels caused coronary heart disease lead to the creation of specific drugs that lower LDL cholesterol, and to subsequent clinical trials proving their efficacy.
Inferring causality from epidemiological studies may not be easy. Causal inference is hampered in clinical research by the fact that (1) multiple exposures often influence a single disease outcome, for example, many people with low LDL cholesterol levels still develop heart disease, because they have other risk factors that play an important causal role, and (2) many exposures take a long time to influence the outcome; for example, the effect of diet on the risk of cancer.
B. Factors favoring an inference of causation Although we can never be sure that a particular exposure causes a particular outcome, a number of factors can be used to help decide whether an exposure of interest is likely to be a cause of a disease, rather than just being associated with it.
1. Evidence arising from randomized studies Studies that randomly assign subjects to one treatment group versus another are generally the most powerful way to show that an exposure is a cause of an outcome. Large randomized trials are usually free from confounding, that is, characteristics of subjects assigned to one particular treatment are usually very similar, on average, to those of subjects assigned to another treatment.
If outcomes differ between treatment groups, it is reasonable to conclude that the treatment is the cause of the difference. Unfortunately, randomized studies are limited to exposures that can be easily assigned to people, such as medications or devices.
2. Strength of association For both interventional and observational studies, a strong association between exposure and outcome increases the likelihood that the exposure is a cause of that outcome. Note that strength of association is not the same as statistical significance. For example, a case-control study observed that infants in the prone sleeping position had a 5-fold greater risk of SIDS (odds ratio 5.0, p-value=0.001). Although the p-value is important to rule out chance as a possible explanation for these findings, the strength of this association- that infants in the prone sleeping position were 5 times more likely to develop SIDS- was important for establishing the prone sleeping position as a cause of SIDS. One reason that strong associations often indicate causation is that there can only be so much bias and error in a well-conducted observational study. For the SIDS example, some errors in classifying SIDS versus other causes of death may have occurred, and there may have been other aspects of infants who sleep in the prone versus supine position that could have also influenced the risk of SIDS; however, it is unlikely that such potential errors would account for the entirety of such a strong observed association. In general, relative risks greater than 2.0 or less than 0.5 are considered to indicate strong associations.
3. Temporal relationship between exposure and outcome For an exposure to be considered as a cause of a disease, there should be evidence that the exposure was present before the disease developed. Consider a study that examines the association between cyclophosphamide chemotherapy and the risk of secondary bladder cancer.11 The study population includes patients who were free from bladder cancer at the beginning of the study, the exposure is the use of cyclophosphamide chemotherapy, and the outcome is newly diagnosed or incident bladder cancer that occurs at least two years after the initiation of chemotherapy. The investigators find that cyclophosphamide use is associated with a 4.5-fold greater risk of developing future bladder cancer. In this example, ensuring that the exposure (cyclophosphamide chemotherapy) clearly preceded the outcome (bladder cancer) in time strengthens the case for cyclophosphamide as a potential cause of secondary bladder cancer.
For second example, consider a hypothetical study that discovers high levels of a novel neurotransmitter, “DP-1” in the blood of people who have established major depressive disorder.
These data alone do not clarify whether higher circulating DP-1 levels were present before the development of depression, or whether depression was present before DP-1 levels increased. The alternative possibility that DP-1 levels might rise in response to depression diminishes the case for causal inference.
4. Exposure-varying association If a primary association between exposure and outcome is observed, the case for causal inference may be strengthened by additional evidence that the association differs predictably across different levels of the exposure. For the cyclophosphamide and bladder cancer example, the overall relative risk of secondary bladder cancer associated with cyclophosphamide use was 4.5.
The investigators next examined the risk of bladder cancer associated with different cumulative doses of cyclophosphamide. Their findings are presented in Table 2.3.
Cyclophosphamide dosage and the relative risk of secondary cancer.
Cumulative cyclophosphamide dosage (g) Relative risk of secondary cancer
This stepwise increased risk of bladder cancer associated with each higher cyclophosphamide dosage strengthens evidence for a causal relationship between cyclophosphamide and bladder cancer. The “dose-response relationship” need not be limited to a medication, and can apply to different levels of any exposure. For example, childhood streptococcal infections have been associated with the development of neuropsychiatric syndromes, such as Tourette’s disorder. A well-conducted observational study observed that children who had a streptococcal infection were
2.2 times more likely to develop a future neuropsychiatric syndrome.12 The investigators strengthened the case for a causal relationship by further showing that the risk of neuropsychiatric syndromes increased steadily with the number of previous streptococcal infections.
5. Biological plausibility Causal inference relies on translational and basic science knowledge to make sense of observed epidemiologic associations. Associations that have proven biologic plausibility based on experimental data are more likely to be causal than those not supported by scientific evidence.
Note that biologic plausibility derives from scientific evidence obtained from other studies. For
the example of LDL cholesterol levels and heart disease, multiple parallel studies took place:
basic science studies demonstrated LDL cholesterol deposition in the arterial wall, translational studies showed enlargement of atherosclerotic plaque size by angiography among patients with higher LDL cholesterol levels, observational studies indicated associations of higher LDL cholesterol levels with a greater risk of developing clinical heart disease, and interventional trials demonstrated a reduced risk of death and cardiovascular disease in patients treated with drugs specifically designed to lower LDL cholesterol levels. This example highlights the importance of interdisciplinary collaboration for producing quality science and for moving the medical research field forward.
1. Case reports and case series describe the experience of one or more people with a disease.
2. Case reports and case series are often the first data alerting to a new disease or condition.
2. Case reports and case series have specific limitations:
a. Lack of a denominator to calculate rates of disease b. Lack of a comparison group c. Select study populations d. Sampling variation Case reports and case series represent the most basic type of study design, in which researchers describe the experience of a single person (case report) or a group of people (case series).
Typically, case reports and case series describe individuals who develop a particular new disease or condition. Case reports and case series can provide compelling reading, because they present a detailed account of the clinical experience of individual study subjects. In contrast, studies that evaluate large numbers of individuals typically summarize the data using statistical measures, such as means and proportions.
A case series describes 15 young women who develop breast cancer; 9 of these women report at least once weekly ingestion of foods packaged with the estrogenic chemical bisphenol A (BPA). Urine testing confirms the presence of BPA among all 9 case women.
It is tempting to surmise from these data that BPA might be causally related to breast cancer.
However, case reports / case series have important limitations that preclude inference of a causal relationship.
First, case reports / case series lack denominator data that are necessary to calculate the rate of disease. The denominator refers to the population from which the diseased subjects arose. For example, to calculate the incidence proportion or incidence rate of breast cancer among women exposed to BPA, the total number of women who were exposed to BPA or the total number of person-years at risk is needed.
Disease rates are needed for comparison with historically reported disease rates, or with rates from a selected comparison group. Unfortunately, obtaining the necessary denominator data may not be easy. In this example, additional data sources are needed to determine the total number of BPA exposed women from whom the breast cancer cases arose. The case series data alone cannot be used to calculate the rate of breast cancer, because they do not include the total number of women who were exposed to BPA.
A second problem with case report / case series report data is the lack of a comparison group.
The 60% prevalence of BPA exposure among women with breast cancer seems unusually high, but what is prevalence of BPA exposure among women without breast cancer? This comparison is critical for addressing the hypothesis that BPA might be a cause of breast cancer.
A third limitation of case reports / case series is that these studies often describe highly select individuals who may not represent the general population. For example, it is possible that the 15 breast cancer cases originated from a single hospital in a community with high levels of air pollution or other potential carcinogens. Under these conditions, a fair estimate of breast cancer incidence among non-BPA exposed women from the same community would be required to make an inference that BPA causes breast cancer.
A fourth limitation of case reports / case series is sampling variation. This concept will be explored in detail later in this book. The basic idea is that there is tremendous natural variation in disease development in humans. The fact that 9/15 women with breast cancer reported BPA exposure is interesting; however, this number may be very different in the next case series of 15 women with breast cancer simply due to chance. A precise estimate of the rate of a disease, independent from chance, can be obtained only by increasing the number of diseased subjects.
Recall the list of factors that are used to judge whether a factor may be a cause of disease:
(1) Randomized evidence (2) Strength of association (3) Temporal relationship between exposure and outcome (4) Dose response association (5) Biological plausibility In general, case reports / case series rely almost exclusively on biologic plausibility to make their case for causation. For the BPA and breast cancer case series, there is no randomized evidence, no measure of the strength of association between BPA and breast cancer, no reported doseresponse association, and no evidence that BPA exposure preceded the development of breast cancer. The inference for causation derives completely from previous biologic knowledge regarding the estrogenic effects of BPA.