Raw data is now rarely presented in clinical trials on nutrition. Most of the time I see data that has been "adjusted" for one or more variables. I do not remember seeing any explanation as to the technical details of these adjustments.
How exactly is this adjustment occurring? What is the math? And why was that math developed? What are the underlying assumptions behind it?
My concern is that we may be adjusting away important correlations if some of our adjustment assumptions are incorrect.
Can someone please help me with this?
hi,
Follow this link for a better understanding on this concept:
https://rstudio-pubs-static.s3.amazonaws.com/622030_3f744c08016446f68ce02cad67c2436a.html
Dear Cate,
In randomized studies, all potential confouders are equally distributed among groups. A confounder is any variable that affects both the exposure variable, like the treatment, and the outcome, like overall survival. However, in non-randomized studies, those confouders are not equally distributes. Adjusting is performed by including those confounders in multivariable regresiones (Cox, logística, linear, etc). Mathematically, this is achieved, basically, by partial derivatives, that enables to analise the influencie of the variation o X1 variable on Y outcome while X2, X3, etc., are stationary. Mathematically, this approach Is accurate. Notwithstanding, it is impossible to collect, and include, all confounders in any study type.
Thank you, both.
So a follow up question then: If something is thought to be a variable that increases mortality, but actually decreases mortality, then how would that impact the conclusions? Specifically, for the sake of argument, lets say that high total cholesterol actually reduces mortality but is considered to increase it. Is that something that could cause problems?
Dear Cate,
Science is constantly constructing knowledge, and we must deal with apparent contradictions. If the data collection and analyzes were carried out properly, that is what it is. However, we must keep in mind that, especially for retrospective studies, an imbalance of such confounders between groups generates a substantial bias even when we adjust the models for such variables. There are basically two ways to minimize, or even corroborate, such "strange" findings: the first would be to perform subgroup analyses, and the second would be propensity matching.
In subgroup analysis, we can subdivide our cohort according to characteristics where there is an interaction with such a confounding variable, that is, one alters the effect of the other. Example: elderly vs. non-elderly, women vs. men, etc.
In propensity (Propensity Score Matching or inverse Propensity Weight), a logistic regression is constructed using mainly confounding factors as explanators of the group/treatment variable. After obtaining the results, we can either adjust the models using the propensity score or, I prefer, use only the paired instances. In the latter, many software pairs two participants as closely as possible to those we use as explainers, greatly reducing the imbalance bias.
I hope this helps.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Materials Science
- Materials
- Inorganic Materials
- Metals
- Alloys
- Steel