Welcome to the Recognizing Bias in Research section of the EBM Express Course. EMB Express is designed to build your understanding of evidence-based practice in short, manageable blocks of content. 

 

In this module, we will examine bias in clinical study designs and explore how confounding occurs, as well as ways to prevent it. 

 

Let’s start with bias. 

 

Bias refers to prejudice or the absence of neutrality. It represents a systematic deviation from the truth in research findings that impacts conclusions, and can occur during either the research process or design phase. 

 

We will look at different biases through the lens of the various study types. 

 

The four study types that we will be covering are Case Control Studies, Cohort Studies, Diagnostic Test Studies, and Screening Studies. 

 

In case control studies, you might find Selection Bias or Recall Bias. Selection bias occurs when individuals or groups in the study differ systematically from the population of interest. And recall bias is a systematic error that occurs when participants do not accurately remember previous events or experiences, or they omit details.  

 

An example of selection bias is when participants in a flu vaccine trial are healthy young adults, but those most likely to receive the intervention may be elderly with comorbidities. In this case, the two groups differ systematically. For recall bias, parents of children with cancer may be more inclined to remember infections that occurred earlier in their child's life compared to parents of children without cancer. This can create a misleading or inaccurate association between childhood infections and the development of cancer.  

 

Cohort Studies might contain Sampling bias or Migration Bias. Sampling bias occurs if the study population is not representative of the entire population. This can affect external validity. Migration bias occurs when patients drop out of a study during follow-up and are systematically different from those who remain. This can cause over- or underestimation of disease prevalence.  

 

An example of sampling bias is a clinical trial for prostate cancer, where less than 3% of the participants are Black men, despite their having a 1.5 times greater chance of developing the disease and being 2.2 times more likely to die from it. In a study about diet and depression, individuals with more severe depression may struggle to adhere to the diet plan. Consequently, they might be more likely to withdraw from the study, resulting in Migration Bias. 

There are four common biases that can occur in diagnostic test studies. Spectrum bias occurs when a diagnostic test is used on a group other than the population for which it was intended. This can have varying effects on sensitivity and specificity. Verification bias occurs during investigations of diagnostic test accuracy when testing strategies differ between groups of individuals, leading to differing ways of verifying the disease of interest. Incorporation bias occurs when the results of an index test are used as part of the reference test in a diagnostic study. The estimated sensitivity and specificity of index tests are at risk of being falsely raised or lowered when the index test results include part of the reference standard.  Ideally, the index test and the reference test should be independent of each other. Lastly, Diagnostic Review Bias occurs when the person interpreting the index test results is aware of the reference standard results.  

 

An example of Spectrum Bias is when a clinical breast examination is used to diagnose women with breast complaints; it correctly identifies breast cancer 85% of the time and gives a true negative result 73% of the time. However, when used as a screening test for women without symptoms, its ability to identify breast cancer drops to 36%. In comparison, its ability to correctly identify those without the disease increases to 96%. As for Verification Bias, a study examined the use of D-dimer tests for diagnosing pulmonary embolism. Patients with positive test results had ventilation–perfusion scans. Those with negative results were monitored instead. This method might have missed some cases of PE that did not present with symptoms, especially if those symptoms had resolved by follow-up. 

 

An example of incorporation bias is when some studies evaluating prostate-specific antigen used PSA results to determine prostate cancer presence, resulting in verification bias and inflated sensitivity estimates. PSA levels should have been excluded from the criteria to confirm the disease. As for Diagnostic Review Bias, a review assessing bias related to reference standards in studies of radiographers’ reading of plain radiographs found that reading performance is often inflated when the observer knows the reference-standard report before commenting on the radiograph. 

 

There are three biases that you may encounter in screening studies. Lead Time is the time between finding a medical condition through screening and when it would usually be diagnosed after a patient shows symptoms and visits a doctor. Lead time bias occurs when it is believed that people with a disease live longer than they actually do because their diagnosis happens earlier than usual. Length time bias is the preferential selection of people whose disease progresses slowly, even without intervention. Thus, comparisons of survival time between screen-detected and symptomatically detected individuals are biased toward the screen-detected group. Overdiagnosis refers to the detection of a disease through screening that would not have caused any issues for the patient if it had not been detected. 

 

 An example of Lead Time bias is a study showing women 80 years and older with breast cancer had improved survival rates when they received regular screenings. Lead time bias can inflate survival rates in screened women, complicating claims of screening benefits without proper adjustment.  Length time bias would occur with screening tests that detect slow-growing tumors, which may not cause symptoms for a long time. Fast-growing tumors, however, often show symptoms before the next screening. As a result, screening tends to identify tumors with a better chance of successful treatment. As for Overdiagnosis bias, Screening tests can detect histologic prostate cancer in men, but for most, the cancer will not become invasive. 

 

Now that we have covered Bias in research, let’s take a look at Confounding. 

 

A confounder is a variable that affects both the exposure, or the factor being studied, and the outcome, or the result being measured. It can make the exposure seem to cause the outcome when it does not. This means a confounder can create the illusion of a link between the exposure and the outcome, or it can hide an actual link that is present. 

 

An example of confounding is when early research indicated hormone replacement therapy reduced heart disease risk, but this changed when income and education were considered, revealing potential confounding factors. 

 

Confounding mostly occurs in observational studies as they are not randomized to ensure equivalent groups for comparison or to eliminate imbalances due to chance. Confounding may also occur in randomized studies if they are poorly designed.  

 

The best way to prevent confounding is to use randomization. However, that may not be enough when imbalances in prognostic factors are anticipated or occur by chance. Stratification and statistical adjustment can help minimize confounding by dividing treatment and control groups into smaller, more homogeneous subgroups where the confounding factor remains constant. 

 

In summary, bias occurs in the process or design of research and affects the conclusions of studies. Different biases can occur depending on the type of study. Confounding suggests an association where none exists or masks a true association. It generally occurs in observational studies and is best prevented by randomization.