Interpretation : Chanfe And Bias Flashcards
Non random error
The word bias means “deviation from the truth” and is used to describe non-random error in epidemiological studies. This non-random error is called systematic error.
Bias, or systematic error, leads to an incorrect estimate of the effect of an exposure on the development of a disease or outcome of interest. The observed effect will be either above or below the true value, depending on the nature of the systematic error. Unlike random error, systematic error is not dependent on the sample size. If we have systematic error in a study, then even if the study is very large (and consequently the statistical confidence interval on the estimate of effect is very small) the estimate of the effect will be biased.. Statistics cannot help us “solve” the problem of bias and much of the effort in designing an epidemiological study is about identifying potential sources of systematic error and taking steps to minimize their impact.
More than 50 types of bias have been identified in epidemiology but, for simplicity, these biases can be grouped into two major types: selection bias and information bias.
Selection bias
Selection bias occurs when there is a systematic difference between the characteristics of the people who take part in a study and the characteristics of those who were eligible but did not take part, or who took part but dropped out during the study. In all instances in which selection bias occurs, the result is an observed relation between exposure and disease that is different among those who are in the study and those who are not. For instance, selection bias can occur when individuals nominate themselves to take part in a research study (selfselection bias). People who volunteer to participate in a study are likely to be different from the rest of the population in a number of demographic and lifestyle variables (volunteers tend to be more health conscious, better-educated, etc); some of these variables may also be risk factors for the outcome of interest.
Bias in case control trials
In case-control studies (Lecture 5), controls are recruited to provide an estimate of the exposure prevalence in the general population from which the cases come. This is relatively straightforward to accomplish in a nested case-control study in which the cases and the controls will arise from a clearly defined population - the cohort. In a population-based casecontrol study, a reference population can also be defined. From that population, a random sample of all incident cases can be obtained and controls will be randomly selected from the disease-free members of the same population (e.g. through random digit dialling from the phone book or random selection from the electoral roll). Sometimes, it is not possible to define the population from which the cases arise. In these circumstances, hospital-based controls may be used because the reference population can then be defined as ‘hospital users’. Hospital controls may also be preferable for logistic reasons (easier and cheaper) and because of less potential for recall bias (see below). But the selection of suitable hospital controls is not an easy task since admission to hospital might be related to ‘exposure’ status.
Neighbour or friend controls may also lead to bias as they may be more similar to the cases than the general population (see note on overmatching in Case-control lecture notes).
Bias in cohort studies
For cohort studies (Lecture 3), selection bias can arise from non-response (e.g. missing data), refusal to participate and loss to follow-up, such that the cohort and its data are no longer is representative of the population from which the cohort was drawn. Of particular concern is the situation when missing data are related to either the exposure and outcome measures. The choice of exposure groups can also be a source of selection bias for cohort studies. For instance, in occupational cohort studies in which workers are recruited as the ‘exposed’ group and the general population is used as the comparison group, it is usual to find that the overall morbidity and mortality of the workers is lower than that of the general population. This is because relatively healthy people are able to remain in employment. In contrast the general population includes those individuals who are too ill to be employed. This type of selection bias is called the healthy worker effect. This bias can be minimised by restricting the analysis to people from the same factory but having a different job.
Randomised intervention trials
Randomised intervention trials (Lecture 4) are subject to selection bias through withdrawals from the study (loss-to-follow-up). The results of a trial may be affected if loss to follow-up is related to exposure status or outcome status.
Information (or measurement) bias
Information (or measurement) bias occurs when measurements or classifications of disease or exposure are inaccurate (i.e. they do not measure correctly what they are supposed to measure). Errors in measurement may be introduced by the observer, by the individual, or by the instruments (e.g. questionnaire or sphygmomanometer) used to make the measurements. We can refer to all these inaccurate measurements as misclassification. Misclassification of exposure or disease status can be of two types, non-differential (or random) misclassification and differential (or non-random) misclassification.
Non-differential misclassification
Non-differential misclassification occurs when an exposure or disease classification is incorrect for equal proportions of subjects in the compared groups. In other words, nondifferential misclassification refers to errors in categorisation of disease that are unrelated to the individual’s exposure status, or misclassification of exposure unrelated to the individual’s disease status. In these circumstances, the misclassification is random (i.e. all individuals have the same probability of being misclassified); however, this random misclassification (also called random measurement error) gives rise to a weakening in estimates of the strength of the association between exposure and disease Non-differential misclassification will make the two groups more alike and leads to underestimation of the strength of the association, when a true association exists. In other words, bias from non-differential misclassification will bias the estimate of effect towards the null hypothesis.
The implications of non-differential misclassification depend heavily on whether the study shows an effect or not. This bias is a greater concern in interpreting studies that seem to indicate the absence of an effect. In these circumstances, it is crucial for the researchers to consider the problem of non-differential misclassification to determine to what extent a real effect might have been missed. On the other hand, it is incorrect to dismiss a study reporting an effect simply because there is substantial non-differential misclassification, since an estimate of effect without the misclassification may be even greater.
Differential misclassification
This occurs when 1) errors in classification of disease status are related to exposure status or 2) errors in classification of exposure status are related to disease status. This differential misclassification can bias the estimates of the association in either direction (under or over estimation of effect size) and, hence, it can be responsible for associations which prove to be spurious. There are two main types of differential misclassification, responder bias and observer bias.
Responder bias
Responder bias. This occurs when the way in which study participants supply information about exposure differs according to outcome status, or the way in which study subjects supply information about outcome differs according to exposure status. As an example, in a casecontrol study, cases’ recall of their past exposure to risk factors may differ from the recall of the controls. If patients with breast cancer are more likely to remember to have ever used oral contraceptives than healthy controls, a spurious association between oral contraceptives and breast cancer will result. This is a particular type of responder bias called recall bias. (Note: recall bias may also be non-differential if recall errors equally affect exposed and unexposed individuals, or cases and controls. But the term usually refers to differential misclassification).
Observer bias
Observer bias. Similarly, observers who know the exposure status of an individual may be consciously or unconsciously predisposed to assess outcome variables according to the hypothesis under study. This type of bias is known as observer bias. For instance, in a trial of non-pharmacologic interventions to lower blood pressure, observers may underestimate the blood pressure of those in the treatment group or overestimate the blood pressure in those in the control group.
One way of minimising this type of bias is to keep the exposure status, or the disease status (depending on the study design), of the individual concealed from the observers (ensure observers are masked/blind). But sometimes this is just not feasible as it may be obvious whether a study subject has the exposure, or the disease, or not.
How to identify bias in epidemiological studies
Besides being able to identify potential sources of bias in a particular study, it is also important to be able to estimate their most likely direction and magnitude. Some strategies may be introduced deliberately into the study to assess the effect of a potential bias. For instance, in a mortality study the vital status of people who were lost to follow-up may be ascertained through routine vital statistics registrars and their mortality compared with that of the people who did participate in the study to ascertain whether selection bias occurred. It is essential that the same procedures be applied to any participant irrespective of his/her exposure or disease status. But often this type of analysis is not possible, and it is always better to strive to minimise bias from the very beginning of the study design process.
Random error
In general, the magnitude of the random error will depend on the size of the sample we take – the larger the sample the closer the estimate is likely to be to the true proportion of smokers. In your statistics course, the concept of sampling error for means and proportions is introduced. This leads us to calculate confidence intervals and significance testing to quantify the random error in the estimates. The concept of random error is equally applicable to the rate ratios, risk ratios, odds ratios, rate differences, and risk differences we estimate using epidemiologic study designs. Whatever measure of effect is estimated, we must calculate the confidence interval and p-value for that estimate
