Bounmy
Confounding means effects of other variables that we are not mainly interested on the dependent variable.
It is natural to get rid of those confounding effects on dependent variable, in order to obtain the real effect of our main independent variable.
Adjusting is the way to do it.
Confounding is where the effect of one variable changes depending on the state of another variable in cases where the experimental design failed to allow separation of these effects.
Nitrogen at 3 levels
Potassium at 3 levels
If each level of nitrogen is tested at each level of potassium then I have a factorial experiment where I can identify the main effects of nitrogen and potassium, and I can calculate how they interact. If I have an experiment where I have three treatments (Treatment 1: low nitrogen, low potassium), (Treatment 2: medium nitrogen, medium potassium), and (Treatment 3, high nitrogen, high potassium) I have now confounded the effects of nitrogen with the effects of potassium, or in layman's terms I can no longer determine what effect nitrogen has, nor what effect potassium has. All I can say is that Treatment 1 was (was not) different from Treatment 2. I can't say why there was a difference.
Limited to layman's terms, I can't see a way to "adjust" anything that will enable you to change this outcome other than to rerun the experiment using a better experimental method.
Confounding is quite different if you are dealing with observational data or designed experimental data.
An experimental design (fractional factorial amongst others) may deliberately introduces confusion -- usually between the main effect of a factor and interactions of others factors. This kind of confounding is not a problem because the confusion is (supposed to be) made on purpose - interactions being known to be weak or absent.
It is more problematic with observational data. To make it simple, suppose you have a dependant variable Y and two explanatory variables X1 and X2 -- of course there is a link between X1 and X2 - otherwise there is no problem.
You may have the following pattern : significant link (correlation if all continuous) between Y and X1 alone, significant link between Y and X2 alone, link between Y and the couple (X1,X2) weakened -- may be still significant but "much less", even no more significant at all with either X1 or X2.
If the link between Y and (X1,X2) is still "significant", the coefficient of X1 is then said to be the effect of X1 corrected from the effects of X2. Another complicated way of saying that is : the coefficient of X1 is the way X1 affects Y, if X2 were held constant - I am not very fond of this one.
If the link Y-(X1,X2) is now no more significant, you are then unable to disentangle the effects of X1 from those of X2, and conversely. No escape, more data are needed.
For instance suppose you have data on cholesterolemia for a group of young women (35
- Badges
- Science topic