When researchers draw conclusions about something on the basis of regression analyses, in general, how strong might their conclusions be and how likely might they be affected by bias?
For example, in relation to a standard hierarchy of evidence, such as this one here Article Options for basing Dietary Reference Intakes (DRIs) on chron...
my feeling is that regressions sit around the middle of the pyramid by cross-sectional studies.
Does anyone have any opinions?
Hi Andrew,
I believe it is the study design which really matter most when you want to quantify the level of evidence rather than the analysis method used. A conclusion drawn from meta-analysis results sits at the top of evidence hierarchy.
Also the main causes of bias are others rather than the analysis methods used.
The most common statistical bias are:
Therefore, to judge the merit of scientific research evidence presented, I would better look at the study design used, Sample Size, possible effect of any of the above mentioned biases and appropriateness of the statistical method used.
Hi Andrew,
I believe it is the study design which really matter most when you want to quantify the level of evidence rather than the analysis method used. A conclusion drawn from meta-analysis results sits at the top of evidence hierarchy.
Also the main causes of bias are others rather than the analysis methods used.
The most common statistical bias are:
Therefore, to judge the merit of scientific research evidence presented, I would better look at the study design used, Sample Size, possible effect of any of the above mentioned biases and appropriateness of the statistical method used.
Andrew -
I think regression is a marvelous tool, but it can't be used blindly. Since correlation is not necessarily causation, there are various ways to misinterpret regression results. (The more predictors/independent variables there are and the smaller the relative sample size, the worse for that consideration, though sample size should also/primarily be based on the desired estimated variance of the prediction error.) An unfortunate selection of data may not be repeatable. Also, it is possible to be mislead, I'm guessing particularly in medicine or nutrition or anywhere that there are many factors to consider. But when you use regression with subject matter content in mind, and good data quality and quantity, I expect it can be very helpful. But because one variable or set of variables might be correlated with another, subject matter knowledge, I think, is very important, especially in more complicated models, but perhaps even in simple ones, I think.
Therefore, I don't see it as something you can rank with type of study, but rather as a tool* that goes with a lot of things.
* Relatedly, as John Tukey said, "The best thing about being a statistician is that you get to play in everyone's backyard."
Cheers - Jim
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression