Session by Vinjar Fønnebø, a medical doctor in Norway who teaches statistics

Definition: Looking at data that has more than one variable. Single out one variable and see what influence it has on the outcome you’re looking at. It helps you reduce the “noise” from the other variables. It can also be used to explore data – hypothesis generating analysis. If you want results though, you will use a different approach.

This is used on experimental studies that ask “does it work?” “efficacy”. Does it work in real life is more about “effectiveness.”

Confounding variables. In any research situation a misrepresentation of a true causal association can arise when…

• individuals are not randomized with regard to exposure
• OR individuals and/or evaluators are not blinded as to the alotted exposure

Quantitative research is usually asking is this better than this or does this influence this in this way.

In the example shown, age was a confounding factor. The multivariate analysis helps to get to more accurate representation of the data. Population distribution is really important. Age was a confounding factor in the example data because it confused the association between ethnic group and incidence.

How do you deal with confounding variables? By using multivariate analysis.

A factor can only confound if it is:

• Associated with the exposure
• AND associated with the outcome (in this case disease)

A factor is only called a confounding factor if a different, similar, relative rate is found in every level of the confounding factor. This is really important. It has to be both different, but consistently similar within the strata of the confounding factor.

Most multivariate analyses have a purpose of controlling for confounding factors. Something to consider when reading research is that you should have thought through how if the confounding variables meet the criteria above. It’s not appropriate to use this “just for the fun of it” to see what it does in the data. You should choose the confounding factors BEFORE you analyze your data. Which variables do I think could be confounding based on my reading and research?

How do deal with confounding variables:

• Either avoid it in the planning phase of your research.
  • Make the compared groups similar on possible confounding variables.
• Adjust for it in the analysis phase
  • Stratified analysis
  • Multivariate techniques

You have to get to know your data and justify why you need those techniques. Vinjar thinks these techniques are used too much in the literature. Some researchers tend to give their data to a statistician to make sense of it and then don’t really understand the results.

Do not be too impressed by fancy intricate statistical techniques to “control” confounding variables. They could be a “cover-up” of poor study planning. We shouldn’t be impressed because it isn’t good research. They have had some big challenges in their research.

If you do quantitative research in your dissertation you will be asked a lot of questions about this in your defense.

Interaction/Effect Modification
Even if individuals are randomized with regard to exposure AND the individuals and/or the evaluators are blinded as to the allotted exposure, you can still have a crude estimate of relative rate that can be completely misleading.

In one example, diet is probably an effect modifier in the connection between smoking and heart disease. We think that smoking causes heart disease, but in Japan there is hardly any heart disease and the people generally smoke. Why? Diet is probably the effect modifier.

Sometimes the “standardized” relative rate can be just as misleading as the crude rate. If you standardize the “confounding” variable and the rates are not similar, then the standardized rate is misleading. In the example the numbers were different across all the age groups. Then you can’t summarize it together. You have to present each age group separately.

A factor can only be interactive if: the association between exposure and outcome (disease) is different in the levels of the interaction factor.

How to deal with interaction:
Avoid it in the planning phase. Be aware of possible interaction factors and give your study sufficient power within every subgroup of the interaction factor. OR you can limit your study to one subgroup of the interaction factor.

Or adjust for it int he analysis phase. Present your data strata-specifically. Use only multivariate methods for “testing” of interactions. BUT be aware of your low power to detect interaction. I.e. the analysis may not look statistically significant but it may still be a really important factor. You need to tabulate your data and see what it looks like.

We should think through the data and the variables and not just do it mechanically (i.e. with software).

Interaction may be difficult to detect; therefore; always look closely at your data before shoving them into fancy statistical programs. Remember that the computer is extremely stupid!!!  What comes out will make sense only if what you put in makes sense.

Think about these two concepts when thinking and planning your study. Then you will be able to reach the RIGHT summit.  Confounding is “clean dirt” that you can do something about. Interaction is very difficult to deal with. Often you don’t see the interaction. It’s easy to overlook. Be aware of this when you read other scientific work. Can you think of any confounding variables or interaction that the author didn’t think of?

Recommended Books

• Statistical First Aid – buy used on Amazon.com. it gives an overview like: If you have this type of data, then this is the test you use.
• Statistics in Small Doses – only if you’ve taken a stats class. It is made up of questions and answers.

Software

• Microsoft Excel
• Epi-Info
• STATA
• SPSS
• SAS