Different Types of Bias in Statistics and Health Research
In the last blog post, we discussed the definition of bias. In this context, we talked about the difference between bias vs. consistency and random error vs. systematic error. In this blog post, we are going over the different types of bias in statistics that are most prevalent in health research. We are going to talk about selection bias, performance bias, detection bias, attrition bias, and reporting bias. We will also give you lots of examples in order to grasp the concept of the different types more intuitively. Also, check out our next blog posts where we talk about more forms of biases that can occur in big data and in more general statistical analysis.
Different Types of Bias in Statistics
Selection bias
Let us refer back to the Cochrane handbook.
“Selection bias refers to systematic differences between baseline characteristics of the groups that are compared”
But what are baseline characteristics? Baseline characteristics is data about age, gender, or race of cases participating in a study. These baseline characterises can be nicely investigated with a summary statistics table which compares the different groups. Check out my blog post about creating such tables. In these tables, statistics such as mean, median, and minimum and maximum are displayed for continuous data. For categorical data, the percentage of each level is displayed. Optionally, one can display the missing values as well. Often, the last column of such a table is a p-value which compares if groups differ or not. So, when one gets a p-value < 0.05 then they can conclude that there are systematic differences between baseline characteristics of the groups that are being compared.
Now, let’s come back to selection bias again. Imagine a clinical controlled trial where there are two groups. One group receives the drug and the other group receives the placebo. The study design should be executed in a way such that there is no or little imbalance of baseline variables. Imbalance is not particularly a bad thing. However, all of the imbalance in the two groups should occur by chance rather than by selection bias.
How Do We Avoid This Kind of Selection Bias?
Random Sequence Generation
So, an imbalance in treatment groups can occur by chance (not particularly bad) or bias (bad). In order to avoid that it is being introduced by bias, random sequence generation is important. When random sequence generation is being used, then a random process is allocating study participants to either drug or placebo. In order for random sequence generation to work perfectly, it is desirable for the physician and participant to not know what random sequence generation is being used. It should be unpredictable and completely random. This can be achieved by a computer. Sometimes, study designers assign participants to groups according to their birth date. Even numbers get the drug and odd number the placebo. This is not desirable because physicians and participants know what treatments they are going to receive and the random sequence generation is so not random anymore. So, if a physician wants to assign one particular individual to drug or placebo then they would be able to do that based on this participants birth date. This would destroy the random sequence generation and bias is being introduced in our study design.
Allocation Concealment
Another way to avoid selection bias is to use allocation concealment. Notice, this is not the same as blinding. Allocation concealment occurs after the random sequence generation. After the computer assigned drug or placebo to participants, physicians and participants know to which group they belong with the lack of proper allocation concealment. However, with allocation concealment, they do not know to which group they belong. The assignment is being concealed to them before they are actually receiving the drug/placebo until the end of the study. This has the advantage that after the drug or the placebo has been assigned to the patient, physicians cannot re-assign participants to groups because they do not know to which group they belong in the first place.
Random sequence generation and allocation concealment do not guarantee balanced groups with respect to baseline characteristics. It only ensures that if an imbalance occurs it is not due to bias. So, if researches have imbalanced groups, they have to decide whether it is due to bias or chance. After that, they have to decide whether they want to adjust some baseline variables or not. This is unfortunately not black and white. A significance test only shows the correctness of the randomization. However, it does not show whether this variable has affected the result or not. So, whether to adjust variables or not is up to the researchers and their domain knowledge. It is important to think whether an imbalance in groups can affect the outcome or not and then act accordingly. Selection bias is probably the most important and complex bias among all the different types of bias in statistics.
Performance Bias
“Performance bias refers to systematic differences between groups in the care that is provided, or in exposure to factors other than the interventions of interest”
This kind of bias occurs when no blinding is used. Imagine the two groups know which kind of treatment they are receiving, drug or placebo. This can alter their behavior and also has a psychological effect which can affect outcomes. On the other hand, if researchers know which kind of treatment a participant is receiving, then the doctor’s care level for participants who are receiving the drug and for participants who are receiving the placebo might differ. The different care levels will change the outcome of the experiment and bias was introduced. Blinding is not always possible. For example, physicians might notice the side effects of a drug. So, know which participant gets what kind of treatment. Another example is when participants receive major treatments that are impossible to hide, like surgery.
Detection Bias
“Detection bias refers to systematic differences between groups in how outcomes are determined”
This type of bias is also related to blinding. Imagine the physician knows which kind of treatment participants received. If the researcher has an agenda and really wants the drug to succeed, outcomes might be evaluated differently as opposed to not knowing which group received what treatment. On the other hand, it also alters outcomes if participants know which treatment they are receiving. Imagine researchers want to test a pain drug and participants know if they are getting the placebo or the drug. A doctor might then ask them what pain level they are experiencing. Participants who know that they received the actual drug might assess their pain level as lower as opposed to if they had not known about their treatment.
Attrition Bias
“Attrition bias refers to systematic differences between groups in withdrawals from a study”
This kind of bias can happen when there is incomplete outcome data either through withdrawals of participants or death. This is problematic because researcher might focus more on participants that are still part of the experiment. As a result, outcomes of these participants might have more weight in the conclusion of the experiment. In addition to that, if there are too many drop-outs, the sample size shrinks, and the analysis lacks the statistical power.
Reporting Bias
“Reporting bias refers to systematic differences between reported and unreported findings”
Again, if researchers have an agenda and want an experiment to go a certain way, then they might omit results that are not desirable to them. This can clearly lead to differences in outcomes.
I hope you have enjoyed the discussion about the different types of bias in statistics. Again, if you want more detail about bias then check out this blog post. Remember that there are all kinds of different types of bias in statistics that I have not discussed in this blog post. However, the different types of bias in statistics we just discussed is a good place to start. In my next blog post, I am talking about more general forms of biases.
Links you might be interested in which explain the different types of bias in statistics: