After Chapter 11.3, you will be able to:
In some cases of biomedical or clinical research, we must move away from petri dishes full of cells or experimental animal models that can have all aspects of their living conditions controlled to research on human subjects. For ethical reasons, which we will discuss later, the level of experimental control is invariably lower than basic science research, and the relationships established by such research will therefore be weaker. In human subjects research, there are both experimental and observational studies.
In biomedical research, it is possible to perform experiments in which an independent variable is manipulated and an outcome is observed. In these experiments, we are still attempting to elicit a causal relationship. Because subjects are in less-controlled conditions, the data analysis phase is more complicated than in laboratory studies. In clinical and social sciences research, it is often still possible to conduct experiments by manipulating the environment or circumstances of the subject.
Randomization is the method used to control for differences between subject groups in biomedical research. Randomization uses an algorithm to determine the placement of each subject into either a control group that receives no treatment or a sham treatment, or one or more treatment groups. A proper randomization algorithm will be equivalent to a coin toss or die roll. Once each individual is assigned to a group, the intervention is performed and the results are measured. Ideally, each group is perfectly matched on conditions such as age and gender; however, as long as there is an appropriate randomization algorithm, the collected data may be analyzed without concern.
Because many of the measures in biomedical research are subjective, the perception of the subject and the investigator may be biased by knowing which group the subject is in. To remove this bias, the subjects and/or investigators may be blinded, which means they do not have information about which group the subject is in. In single-blind experiments, only the patient or the assessor (the person who makes measurements on the patient or performs subjective evaluations) is blinded. In double-blind experiments, the investigator, subject, and assessor all do not know the subject’s group. Without blinding, the placebo effect would be greatly reduced in the control group, but still be present in the treatment group.
Blinding isn’t only useful in drug trials; even sham treatments of acupuncture have been used to blind subjects in randomized controlled trials focusing on the use of acupuncture for musculoskeletal pain.
In biomedical research, data analysis must account for variables outside of the independent and dependent variables considered. Most often, these include gender and age; lifestyle variables, such as smoking status and body mass index (BMI), and other factors that may affect the measured outcomes. Some of these other factors can be inferred from the initial literature review, although other unexpected confounding variables may exist. Software programs can use binary (yes vs. no, better vs. worse), as well as continuous (amount of weight lost, percent improvement in cardiac output) or categorical variables (state of residence, socioeconomic status) to create a regression model. Regression analysis may demonstrate linear, parabolic, exponential, logarithmic, or other relationships, as we will discuss in Chapter 12 of MCAT Physics and Math Review.
We may wish to study certain causal associations for which an experiment cannot be performed for ethical or practical reasons. In such a case, we must draw on the available data and analyze it. Observational studies in medicine fit into one of three categories: cohort studies, cross-sectional studies, and case–control studies. These studies often look for the connections between exposures and outcomes. Observational studies do not demonstrate causality, although the tendency toward causality may be demonstrated by Hill’s criteria, which we will examine later.
Cohort studies are those in which subjects are sorted into groups based on differences in risk factors (exposures), and then assessed at various intervals to determine how many subjects in each group had a certain outcome. For example, a study in which 100 smokers and 100 nonsmokers are followed for 20 years while counting the number of subjects who develop lung cancer in each group would be an example of a cohort study.
Cross-sectional studies attempt to categorize patients into different groups at a single point in time. For example, a study to determine the prevalence of lung cancer in smokers and nonsmokers at a given point in time would be an example of a cross-sectional study.
Case-control studies start by identifying the number of subjects with or without a particular outcome, and then look backwards to assess how many subjects in each group had exposure to a particular risk factor. For example, a study in which 100 patients with lung cancer and 100 patients without lung cancer are assessed for their smoking history would be an example of a case–control study.
Hill’s criteria describe the components of an observed relationship that increase the likelihood of causality in the relationship. While only the first criterion is necessary for the relationship to be causal, it is not sufficient. The more criteria that are satisfied by a relationship, the likelier it is that the relationship is causal. Hill’s criteria do not provide an absolute guideline on whether a relationship is causal; thus, for any observational study, the relationship should be described as a correlation.
In addition to the measurement error found in basic science research, we must be aware of bias and error introduced by using human subjects as part of an experimental or observational model. As mentioned earlier, bias is a systematic error. As such, it generally does not impact the precision of the data, but rather skews the data in one direction or another. Bias is a result of flaws in the data collection phase of an experimental or observational study. Confounding is an error during analysis.
The most prevalent type of bias is selection bias, in which the subjects used for the study are not representative of the target population. People who volunteer for a study in a particular area may be significantly different from people who do not volunteer. For example, someone volunteering for a drug trial that requires clinical visits may be healthier or more likely to benefit from the study than someone who does not volunteer because they cannot make it to the hospital.
Selection bias may also apply in cases where one gender is more prevalent in a study than another, or where there are differences in the age profile of the experiment group and the population. Measurement and assessment of selection bias occurs before any intervention.
Detection bias results from educated professionals using their knowledge in an inconsistent way. Because prior studies have indicated that there is a correlation between two variables, finding one of them increases the likelihood that the researcher will search for the second. For example, high blood pressure (hypertension) and diabetes mellitus are more common in the obese population; thus, a physician may screen obese patients for hypertension and diabetes at a higher rate than healthy-weight patients, inflating the true value of the secondary measurement (although, as described in Chapter 12 of MCAT Behavioral Sciences Review, other biases against obese individuals actually tend to lead to lower rates of screening and preventative care).
The Hawthorne effect, or observation bias, posits that the behavior of study participants is altered simply because they recognize that they are being studied. Often these lifestyle alterations improve the health of the sample population. For example, patients in a study for a given weight loss drug may begin exercising more frequently or may make healthier diet choices, thus artificially increasing the perceived effect of the drug. Because the change in data is systematic and occurs before data analysis, this is an example of bias.
Confounding, sometimes inaccurately called confounding bias or omitted variable bias, is a data analysis error. The data may or may not be flawed, but an incorrect relationship is characterized. For example, consider the statement Having natural red hair leads to a decreased pain tolerance and higher opiate tolerance. There are two flaws with this statement. First, the statement implies a causal relationship as a result of what would almost certainly be an observational study. Second, consider whether or not this is realistic. How could red hair cause the findings described? According to current research, there is no likely causality between these two. However, a third variable, such as a gene mutation, could potentially cause both parts of this statement. If one measured the degree of red hair pigment and the degree of pain intolerance, there might be a very strong statistical relationship, but there is no causal relationship between the two. These “third-party” variables are called confounding variables or confounders, as illustrated schematically in Figure 11.2.