5.1.1 Selection bias

Selection bias occurs when individuals or groups in a study (the ‘study population’) differ systematically from the population of interest (the ‘target population’).

In AMR studies, selection bias occurs often when patients are selected because they are sick with a bacterial infection, and their clinician has requested testing. The results of this testing are then often included in a study or surveillance system. If these findings are used to make inferences about the level of AMR in the general population, this is would likely lead to an overestimate of the frequency of AMR. Selection bias can also arise even if the findings are not inferred to reflect the general population, but only hospital patients. For example, in most countries, tertiary referral hospitals in capital cities receive patients with critical or complex illnesses from all over the country, who are transferred from the first hospital or health facility they attended. These hospitals are also often the best equipped to conduct a study on AMR.

So, if a study is performed on AMR in a tertiary referral hospital, and the results are used to draw conclusions about the frequency of AMR in hospitals in the same country in general, the findings might be biased by the fact that the patients in this study are sicker on average (and perhaps more likely to have resistant infections) than the average hospital patient in the country. On the other hand, tertiary referral hospitals often have better infection prevention and control practices than other hospitals, and so despite the fact they have sicker patients on average, they might have lower frequencies of AMR because of their improved infection prevention and control. It’s not always straightforward to predict how selection bias might affect the results!

Selection bias can also affect studies in animal populations. For example, it is common to sample broiler (meat) chickens for resistant pathogens at slaughterhouses. This makes a lot of sense – it is an efficient way to sample a large number of birds, and as the meat is then sold for human consumption, it is an important time point for evaluating the risk of transfer of AMR infections from food animals to humans. However, typically only healthy animals reach the stage of being sent to the slaughterhouse – sick birds may die or are euthanised before reaching the end of the production cycle. If those birds were sick with a bacterial infection, they might have been more likely to have had a resistant infection that did not respond to any treatment attempted by the farmer or veterinarian (including antimicrobial medicines in animal feed or water). This means that the results of a study of AMR in broiler poultry when samples were collected from slaughterhouses might underestimate the frequency of AMR in broilers overall.

Selection bias can be a particular issue in settings where people (patients or farmers) have to pay a fee for their samples to be tested. Depending on how expensive the fee is compared to typical incomes, it can be the case that sampling and testing for AMR is only performed on a subset of patients (or animals owned by farmers) who can afford it. These wealthier patients (or the animals the farmers raise) might be healthier on average and have different risks of AMR infections compared to the general population.