Bias is similar to accuracy in that it observes the variation from the true value. Often however we are uncertain about the true value. This makes assessment of bias difficult. Obviously, if the accuracy is 100%, there is no bias (unbiased).
There are many types of bias described, but they can be separated into three main groups:
- Selection bias
- Confounding bias
- Observational bias
Selection bias occurs when there are differences between the study groups. For example, an orthopaedic surgeon measures time to fracture healing in two groups of patients with tibial fractures. One group is treated by external fixation and the other by intramedullary nailing. If the surgeon selects the treatment method, it is likely the two groups are not comparable. Perhaps the external fixator group contains mostly patients who had open fractures and the intramedullary nailing group patients with closed tibial fractures following football injuries. Obviously, as the groups are not comparable, this could have an effect on the outcome of the study. Selection bias can be reduced by randomisation. For example the surgeon can, when a patient with a tibial fracture is admitted to hospital, randomise the treatment to either intramedullary nailing or external fixation. The randomisation process (rather than the surgeon) decides what treatment the patient will receive.
This really is a special form of selection bias. This form of bias is observed when another factor (than the measured variable) influences the outcome. The factor causing the confounding bias is called the confounding factor. For example, we know that smoking has adverse effect on fracture healing. Suppose, in a trial, we compare two types of osteotomy and use the time to fracture healing as the outcome measure. It could well be that there are more smokers in the one group than in the other. Obviously, this would influence the outcome of the trial. In this example, smoking is the confounding factor. When we know something is a confounding factor, stratification can be used to reduce confounding bias. Stratification is a special form of randomisation. The mathematics of which is beyond the scope of this book. However, suffice to say that the stratification process ascertains there are the same proportion of smokers in each study group.
Observational bias is observed when there is error in measurement of disease or misclassification. For example, this can occur when the results of conservative and operatively treated tibial fractures are compared. The fact that patients in one group had an operation whilst the other group had not; can lead to observational bias. Patients might think they are better off (or perhaps worse off) because they had an operation. Also, in assessing the outcome, the surgeon might be influenced by the fact that the patient had an operation rather than conservative treatment. This is more likely if the surgeon performed the operation him or her self.
One way of trying to reduce observational bias is by requesting another surgeon to measure the outcome. However, this does not eliminate bias. The other surgeon may well show bias that could be ‘in favour’ or ‘against’ the surgical procedure. An orthopaedic registrar performing the measurements for his / her consultant is unlikely to ‘upset the boss’ and therefore also biased.
Another way to reduce observational bias is by blinding. Ideally, the subject of investigation as well as the person measuring the outcome are blinded. This is called double blinding. Obviously, this is not always possible in orthopaedics. Radiographs often show what procedure has been performed, in which case we will have to suffice with single blinding. In animal experiments, a sham operation could be performed to try to reduce observational bias. A sham operation is an operation in the control group; the surgical exposure is the same, but no procedure is performed. Obviously, this is unethical in human studies. A sham operation is the surgical equivalent of a placebo (non functional drug given to the control group).
Another type of observational bias occurs when we measure the outcome of hip replacement with a functional scoring system (such as the Harris hip score). Such scoring systems normally assess pain, function and range of motion. Function is often assessed as ‘walking with a frame’, ‘crutch’, ‘stick’ or ‘unaided’. It could well be that the patient’s hip is functioning extremely well, but that the patient is walking with a stick because there is also severe arthritis in the knee. Consequently, the outcome measure is not measuring what it supposed to measure (function of the hip replacement). The observational bias so introduced, can be reduced by trying to improve upon the outcome measure.