Table of Contents
Systematic error, also called bias, usually leads to an inaccurate estimate of the variable we are studying.
To err is human, to forgive is divine, as the saying goes. But, what does it mean?. If you read the sentence you could think that making mistakes is part of human nature. And may be a grain of truth in it, because few are the times that we, the so-called human beings, do something that is not riddled with errors, even though we try hard not to make any.
Regarding forgiveness, it is also true that it is more a divine than a human trait. Though there are some errors that deserve no forgiveness, neither human nor divine.
Random errors, as the name suggests, are due to chance. When we want to study a variable in a population we usually have to be satisfied with a selected sample from that population. Well, with random sampling there is always a certain probability that the sample is not representative of the population from which it comes. This probability of error will be larger the smaller the sample size and the greater the variability of the studied variable in the population.
Another source of random error is the variability of the measurements we do, either by their biological variability, by the instruments that we use or by observer subjectivity or variability. For example, let’s think that we are going to do a study about the prevalence of tuberculosis in our sample by studying the tuberculin skin reaction and that the day of measurement we get our glasses broken. Any resemblance to reality will be purely coincidental.
The other kind of errors is systematic errors, also called bias, which usually lead to an incorrect estimation of the effect we are studying. These are not due to chance, but to errors in study design, whether related to the participants (selection bias) or to the measurement of the variable (information bias).
Types of bias
Selection bias occurs typically when we choose an unrepresentative sample from the population. Let’s consider that we want to know the prevalence of a disease and we take a sample from patients attending our clinic. Obviously, the result will be biased and the prevalence in the population over-represented.
Also, selection bias may occur in other situations. For instance, if we pick out a control group with a disease related to the one of study, our results will be incorrect. It can also occur when the probability of being lost to follow-up is not equal in the two groups.
For example, let’s suppose we’re studying two interventions and there’s the same percentage of lost to follow-up patients in both groups, although one group tends to miss the respondents and the other the non-respondents. Although the response rate is the same, the truth is that the most effective intervention is the one in which the non-respondents patients are more susceptible to get lost to follow-up. And something similar happens with no-answers when we do polls. If you ask something about a socially shameful disease topic, you’ll always underestimate the real prevalence.
Meanwhile, information bias occurs when, in a systematic way, we do the measure in a wrong or different way in both groups. In general, it is caused by using tests with low sensitivity or specificity, by using incorrect diagnose criteria or by committing inaccuracies or errors in data collection. Let’s suppose we study weight in our patients and the scale is out of calibration. Or that we study height and we measure one group barefoot and the other wearing shoes.
There are a couple of differences between the two types of errors, random and systematic. As we’ve already mentioned, random error depends on the sample size, so it tends to be lower with increasing sample size. However, this does not happen with systematic errors that are perpetuated however much we increase sample size.
Moreover, random errors can be controlled with relative easy, if they are not too large, during data analysis, whereas systematic errors are much more difficult to correct when analyzing the results. This is why you have to be very careful during the design phase and try to avoid them.
And with that we are done for today. Just realize that bias family is a very large one. Although almost all errors can be included in any of the mentioned above, there are many more types of bias described, many of them specific of a particular type of study design. But that’s another story…