Bias in observational study designs: prospective cohort studies

Answers

Answers b, c, d, e, and f are true, whereas a is false.

A prospective cohort study was used to investigate the association between smoking behaviour and diagnosed invasive breast cancer in postmenopausal women. Prospective cohort studies are observational by design and have been described in a previous question.2 The participants were postmenopausal women originally recruited to the Women’s Health Initiative Observational Study, which had been designed to investigate causes of morbidity and mortality. Cohort studies are prone to various types of bias. Bias is a systematic error, rather than random variation or lack of precision, in the recruitment of participants, the measurement of their risk factors and outcomes, or reporting of the results. Observational studies are recognised to be prone to three broad types of bias, including selection bias, information bias, and confounding. The most common types of bias within these categories are described below.

The Women’s Health Initiative Observational Study originally recruited 93 676 women aged 50-79 from 40 clinical centres throughout the US between 1993 and 1998. The cohort would probably have been representative of the population of postmenopausal women in its characteristics because it was sampled from a large number of clinical centres. However, the cohort was still prone to selection bias (f is true), a general term used to describe a group of biases and effects that result in a sample that is not representative of the population from which it was selected. Selection bias results in a lack of external validity—that is, the extent to which the study results can be generalised to the population that the sample is meant to represent. Selection bias includes non-response bias, the healthy entrant effect, and attrition bias, which are all described below.

Non-response bias would have occurred if the women who accepted the invitation to be part of the study were different from those who did not. However, any differences in characteristics (including demographics, risk factors such as smoking behaviour, and prognostic factors) would be difficult to quantify because limited information, if any, would be available for those who refused to be part of the cohort. Non-response bias should not be confused with response bias, described below.

In total, 93 676 women were recruited to the Women’s Health Initiative Observational Study. Women were not recruited if they had conditions that were predictive of survival less than three years or had complicating conditions such as alcoholism, drug dependency, or dementia. For the above analysis, 12 075 (12.9%) women with a history of cancer (except non-melanoma skin cancer) were excluded at baseline. When investigating the epidemiology of invasive breast cancer, it was important that the women selected were healthy and free of cancer at baseline. This enabled the temporal association between the risk factor of smoking behaviour and diagnosis of invasive breast cancer to be determined. In particular, it ensured that smoking behaviour and exposure to other pertinent risk factors occurred before the diagnosis of invasive breast cancer, thereby permitting a potential causal role to be investigated. Because women were healthy and free of cancer at baseline, the cohort would have been healthier than the general population of postmenopausal women. The results of the study would therefore have been prone to the healthy entrant effect (d is true)—a reduction in rates of morbidity and mortality in the initial stages of the study compared with the general population of postmenopausal women. Rates of morbidity and mortality would have started to increase and resemble those in the population several years after the start of the study.

As is typical of most cohort studies, the Women’s Health Initiative Observational Study recruited a large number of women who were followed for a substantial period of time. The median length of follow-up in the above study was 10.3 years. This ensured that a sufficient number of women received a diagnosis of invasive breast cancer, thereby enabling the association with smoking to be investigated. It would have been difficult to maintain contact with all of the women because of the size of the cohort and length of follow-up. Of the 93 676 women originally recruited to the Women’s Health Initiative Observational Study, 443 (0.5%) were lost to follow-up. Therefore, the above study was prone to attrition bias (b is true), also known as loss to follow-up bias. Attrition bias would have occurred if the women lost to follow-up differed in a systematic way to those not lost to follow-up. In particular, it would have been a problem if the reason for loss to follow-up was related to the risk factor of smoking or the outcome of diagnosed invasive breast cancer. Although the proportion of the cohort in the above study that was lost to follow-up was minimal, it may still have biased the results. There is no exact proportion of a cohort that when lost to follow-up results in attrition related bias becoming a concern.

Of the original cohort of 93 676 women, 1168 (1.25%) had missing values of smoking status at baseline. If the participants with complete data on smoking behaviour at baseline were not representative of the population of postmenopausal women, then the women with missing data would have introduced selection bias.

The collection of data in the above study was prone to information bias, a type of bias that occurs during data collection. It would have occurred if the measurements of the risk factors (including smoking) or outcome (diagnosed invasive breast cancer) were systematically distorted. Such bias in data collection can be unconscious or otherwise and can come from the investigators or participants. Response bias would have occurred on behalf of the women if their reported smoking behaviour was systematically different from their actual smoking behaviour (e is true). For example, women may have under-reported their smoking behaviour because they were aware that it can affect their health. Assessment bias, also known as observer bias, would have occurred if the outcome of invasive breast cancer was diagnosed incorrectly. For example, diagnosis could have been influenced by knowledge of the study research hypotheses. Response and assessment biases are part of a group of biases collectively known as ascertainment bias, sometimes referred to as detection bias.

Confounding would have occurred if a variable, such as alcohol intake, obscured the association between smoking behaviour and diagnosed invasive breast cancer. Alcohol intake would then be a confounding variable or confounder. To be a confounder, alcohol intake would have to be a risk factor for diagnosed invasive breast cancer. Alcohol intake would also have to be associated with smoking behaviour. For example, if women who smoked were also more likely to drink alcohol, the increased risk of invasive breast cancer associated with being a current smoker might have partly been the result of alcohol intake. Confounding was accounted for at the analysis stage—the association between smoking and diagnosed invasive breast cancer (as measured by a hazard ratio) was adjusted for all other potential confounders measured at baseline. The adjusted hazard ratio provided a true indication of the association between smoking behaviour and diagnosed invasive breast cancer, with all other confounding variables being equal.

The results of the above study were still prone to confounding (c is true), however, because it was unlikely that all potential confounders were measured and adjusted for. Furthermore, confounding variables must be measured accurately. In the above study, information on smoking behaviour and other potential confounders was collected at baseline. However, if cohort members changed their smoking behaviour during follow-up and after baseline measurements, perhaps as a result of knowledge of the study research hypotheses, this would have introduced further confounding.

Allocation bias is mainly of concern in clinical trials, not cohort studies (a is false). It occurs when there is a systematic difference between participants in how they are allocated to treatment groups. For example, researchers may allocate people who they think would show the greatest benefit to a particular intervention, perhaps because they favour the intervention and wish to show that it is more effective than the control treatment. Allocation bias is eliminated by randomising trial participants to treatment so that all participants have the same probability of being allocated to each treatment group.3