I know this might have been discussed a gazillion times. I fail to understand the difference between the two when doing a question. Can anyone explain the difference in simple words???
How stratification helps in figuring these out? Do we consider the p value or RR?
Confounding, in really imprecise terms, means that you're unable to tell the effect of x on y because you didn't properly control for z. For example, we might find that men have shorter life spans, on average, in comparison to women. Is it necessarily due to gender or is it possible that men and women have different behaviors that would cause a divergence in average lifespan? Maybe men or more likely to smoke, drink alcohol, engage in risky behavior, or other things that would reduce the average male lifespan. Maybe women engage in activities like exercise, healthy habits, or other things that would increase their lifespan, on average. We don't know because we didn't control for these things (either we didn't use a statistical technique to explicitly account for them, or we had groups (M/F) that were unbalanced with regards to these attributes). If we did account for these other factors that differ between genders, and they really effect the life span, we might find that gender really doesn't have any influence, on the average. (So stratification can help with this...stratify by smoking status, for example.)
Effect modification is a case where one of the independent variables(Z) changes (or modifies) the relationship between the dependent variable (Y) and a different independent variable (X). For example, if we look at STEP 1 scores (Y) as a function of hours studying (X) and whether or not a q-bank was used (Z), we can see this effect (most likely). The increasing the number of hours studied will increase the STEP 1 score, on average. This increase is likely to change depending on whether or not students also used a q-bank. To simplify this, pretend the relationship is a straight line. To see the effect modification we will have 2 separate lines. One line is for students who used a q-bank, and one line is for students who didn't use a q-bank. In the event of effect modification, these lines will be non-parallel and intersect at some point-- they will have different slopes. It's reasonable to assume that the line relating STEP1 scores to hours studied will increase more rapidly for students who use a q-bank when compared to student who don't use a q-bank. This would demonstrate effect modification, since q-bank use changes how the number of hours studied impacts STEP 1 scores. It's important to note that this would also work with another quantitative variable (instead of a qualitative), but it's easier to visualize this way. Also, effect modification does not imply that one independent variable is correlated with another. It only means that the effect of x on y changes depending on the value of z (assuming both x and z are independent variables...they are also interchangeable in the interpretation). This also applies to non-linear functions, but again, straight lines and slopes should convey the point more easily.
So, to quickly tie these back to p-values and RR:
P-values are used to determine statistical significance...you could use one to determine if an effect modification is statistically significant at some threshold.
RR might be different between two groups (our first example), but we don't really know why or if this difference is due to unbalanced characteristics between the groups. We should control or account for these other potential variables.
For our second example, RR of an event when looking at drug x compared with placebo might be different for men and women. In other words, the effect of drug x on the risk of an event occurring might depend on you gender (or age or some other variable).
Let me know if this helps.
