THE QUALITY OF MEASUREMENTS: What Then Is Meant By Bias?

WHAT THEN IS MEANT BY BIAS?

Biases detract from validity. They distort the estimation of a parameter and do so systematically. For example, the true mean height of a population is 175 cm, but may be estimated from a sample to be 170 cm because the observer is tall and by parallax underreads by one gradation on the scale. The point estimate is then systematically biased downwards by up to 5 cm.

Sackett in his article "Bias in case-control studies" provides a telephone directory of differently named biases.

Rothman provides a more logical approach to bias and considers three main types of bias which partly overlap:

1. Selection bias:

This occurs as a result of the sampling process. The effect estimated for the sample population will be different from that potentially applicable to the target population. Two examples of this are due to voluntary self selection bias where the sick present more frequently for surveys, or the healthy worker effect where sick workers leave the factory at a higher rate leaving relatively healthier people behind who are then surveyed.

Diagnostic bias is another form of selection bias in which people with suggestive symptoms or signs are diagnosed as a result of a suspected relation of these symptoms and signs with an exposure. Rothman gives the example of oral contraceptive use and venous thromboembolism. Because doctors are aware of a possible relationship are more likely to diagnose this condition in patients taking contraceptives, and therefore more likely to hospitalise them, a study of those hospitalised with this diagnosis could lead to an exaggerated association.

In general, selection bias arises when the relation between exposure and disease is different for those participating in the study from those who are eligible but do not participate.

2. Information bias:

This arises out of the measurement process and results from differentially available existing information, diagnostic standards or protocols, and subjective recall bias of study participants to questionnaire items about past events. Such bias leads to misclassification. An example would be doing a study where in one group mortality data based on post-mortems is compared with another group where mortality data drawn from a population death register.

The key question is whether misclassification is differential or not across comparison groups. With differential misclassification the results are uninterpretable in that the direction of bias is indeterminate. With non-differential misclassification estimates of an existing effect are always biased in the direction of the null value (no effect) and sometimes even beyond (a negative effect).

In general, subjects have already been selected but information on them is imperfectly available, either differentially or nondifferentially, leading to misclassification.

Example of Non-differential misclassification. Table 1 shows the true figures for exposure among cases and controls.

Table 2 shows the same number of people studied but due to inaccuracy in the record keeping system 20% (50) of the controls are misclassified as cases among the exposed and 20% (120) of the controls are misclassified as cases among the non-exposed. There is then 20% misclassification of disease status which is the same for both the exposed and the unexposed, that is, nondifferential misclassification of diagnosis with respect to exposure.

The same would apply to non-differential exposure classification of the cases and controls.

The main point is that when you do not find an effect in an occupational health study - a negative study - it could be due to non-differential misclassification. When you do find an effect, if anything it is likely to be bigger than what you find.

3. Confounding bias:

This is unique in that it can be easily controlled for at the data analysis stage by standardization of the groups to be compared provided data has been collected on the confounding variable. In this way "apples can be compared with apples - instead of with oranges". There are other more complex statistical methods of controlling for confounding too.

Particularly important sources of bias relate to biological variations (for example, circadian rhythms) and it is important in studies to obtain comparable measurements for everyone. For example if one was studying byssinosis, different times at which FEV1 were measured would be important as this increases perceptibly throughout the working day. Many biological parameters are subject to circadian or other rythms. Weight or respiratory symptoms may vary seasonally. Systematically different measurement times could cause bias.

General Introduction to Occupational Health: Occupational Hygiene, Epidemiology & Biostatistics by Prof Jonny Myers is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 2.5 South Africa License