11/26/12
Bias, Part One
Look, we all make mistakes. We’re human. If we stick our fingers into an electrical socket, and we get shocked, it will more than likely be the only time we do it. I know I only did it once. It didn’t matter if the electricity was on or off anymore after that. The one shock was enough for me.
Researchers are also human, so they will also make errors. In research, there are two main types of errors. One type is random errors, those that happen “just because” and are more related to the inherent nature of what you’re testing or who you’re sampling. While there is nothing you can do about them, you can minimize them with things like randomization of subjects or stratifying the results. The second type of error is more insidious and can be easily overlooked. So let’s talk about it… Let’s talk about systematic errors, also known as bias, after the jump.
 |
| There are biases and bi-asses, get it? |
I need to warn you right now that this will be a four-part lesson because I’m going to include plenty of examples for each kind of bias. Systematic errors come in three varieties: Selection, Measurement, and Intervention biases. So this post will be about selection bias. The second one will be about measurement bias. The third one will be about, what else, intervention bias. And the fourth will touch on confounding.
SELECTION BIAS
Let’s say that you’re conducting a rather large study on the effectiveness of a new medication against malaria. If it were a prospective study, then you would give the medication to a group and placebo (or nothing) to another group. The treatment and control groups would then be followed to see which group develops malaria in higher proportion, or faster, or a worse kind of malaria, or is cured from malaria. At any rate, you’re just looking for some sort of outcome.
If it were a case-control study, then you would interview a group of sick people and a group of healthy people. You would then determine if those who are now healthy were more likely to have taken the medication than those who are now sick. When we look at study designs, we’ll talk about risk ratio and odds ratio. Anyway, in both these scenarios, you can probably tell that how you pick your subjects is key, right?
 |
| How you choose your prey is also key. |
If you were studying a treatment for gonorrhea, who would be most likely to volunteer for the study: A person who got it from an affair or a person who got it by accident at an all-night kegger? Seriously, these are the things you need to ask yourself in your study design because it goes to who is most likely to volunteer. A person with HIV/AIDS who knows that an infection of any type will likely be a very bad thing for them might be more likely to volunteer. If that is the case, then was it your treatment that cured them or some other medication they’re getting for their condition?
 |
| If it's homeopathy that they're taking, then it's your medication that did the trick. |
Let’s say that you live in a town that is 50% Republican and 50% Democrat. Further, let’s say that 90% of Republicans have a telephone, while only 10% of Democrats have a telephone. If you were to conduct a random-digit dialing poll about the next election, who would you think was going to win? Remember, the split is down the middle in the general population. See where the bias comes in?
 |
| The master of bias. |
One way to control for volunteer bias is to tweak the definition of a case or the requirements of a volunteer BEFORE you do the study. For example, you might want to keep out all people who are taking antibiotics if you are measuring the effectiveness of your own antibiotic at preventing infections, because it will be hard to know if it was your antibiotic that did the job. (This is also a co-intervention bias, which we’ll discuss when we talk about intervention biases.)
In the polling example, you might want to call nine times more Democrats than Republicans to get a sample that is 50-50 when it comes to party affiliation. That would balance out your sample, increase your sample size, and minimize selection bias.
NON-RESPONDENT BIAS
Another type of selection bias is non-respondent bias. This is a problem when people with significant differences respond to surveys or follow protocols differently. For example, let’s say that you’re looking at the malaria drug again. Who would be most likely to respond and participate in your study? People who live in Nigeria or people in Canada?
 |
| It was a rhetorical question. |
Selection bias also appears when we think there is a cause and effect in a population when the effect would probably appear without us doing much. A classic example is looking at the association of vaccines and respiratory problems in babies born prematurely or with very low weight. If you give them a vaccine and they get sick from a respiratory condition, you really need to take a look at the control group because the condition may be common for babies in the study group.
Another example I can think of is a medication for chronic-obstructive pulmonary disorder (COPD). What if you saw that several of your study participants had respiratory problems and even died during the study? Would you blame the medication? You’d have to look at the incidence of respiratory problems or death in the control group before you make a decision. (Sadly, this is not the case in the court of public opinion, where association almost always means causation.)
CONTROL, CONTROL, CONTROL
As much as you can, you need to control for these types of biases in your study design and, to an extent, in your analysis. If you go trying a vaccine against pancreatic cancer, and you only enroll vegans who live in a far away island, then the peer review process is going to rip the study to shreds. Likewise, if you are looking at the association between coffee and cancer, you better look at how many smokers are in each group. (We’ll talk about the coffee-cancer example when we talk about confounding, another type of bias, sort of.) Or if you’re looking at an outbreak of vomiting in a school, for crying out loud, don’t count as cases only those people who ate one specific item of food, leaving others out.
 |
| Speaking of control... Well, at least he's safe. |
My grandmother used to say that the best cook still has an entire tomato fall into the soup. It sounds better in Spanish; trust me. At any rate, any and all researchers and their studies are susceptible to bias. In the next lesson, we’ll look at measurement bias, also with examples. We’ll cover why both a thermometer and a human can really screw things up.
Thank you for your time.
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment
Keep the comments clean.