Such a comparison is biased, because in the general population there are many people who cannot work because they suffer from poor health. As a consequence, the death rates in workers are often lower than those in the general population – the so-called healthy worker effect. It would be better to compare the death rates of workers in a specific job to those of workers in other jobs.

Also cohort studies may suffer from selection bias. In this regard it has been suggested that the comparison of survival of haemodialysis and peritoneal dialysis patients will always suffer from selection bias, as these patients are inherently different from the beginning. Randomization may address this problem, but it has been shown that randomization for treatment modality is very difficult. In cohort studies as well as clinical trials the primary sources of selection bias are non-response, loss-to-follow-up and withdrawal from the study. Many times such loss of data is not co-incidental. Especially those subjects who are either in very bad or in very good health are at a higher risk of drop-out. As a consequence the subjects who remain for data analysis may no longer be representative for the source population from which the original sample was selected. If that selection would to a certain extent be different in the one treatment group compared to the other, this will result in a biased comparison.

The key to decrease bias is to identify possible areas where bias may occur at the design stage of a study and to change the study design accordingly. In this respect it is important that the study subjects should be clearly defined. In case-control studies bias will be reduced if there is a defined population with full ascertainment of cases and the controls are a random sample of the population from which the cases were derived. As stated, in an intervention study randomization of subjects to treatment groups should prevent the selection of subjects with a better prognosis for the preferred treatment. Multiple methods of follow-up can help to reduce loss-to-follow-up. It is also useful to compare the subjects who were later lost to see how they compare with the other study subjects. Finally, knowing the reasons for withdrawal from a study will help to interpret its results.