I'm working on a database with almost 300 patients, of which I have their age, sex, BMI and their level of diabetes (which, for the purpose of the study, is stratified into mild, intermediate or severe diabetes). They also have to do a cognitive questionaire, and the focus of the study is to see if the level of diabetes results in cognitive damage/impairment. However, after making the three groups, I noticed that there are statistically significant differences in terms of age, sex and BMI. Hence, I think it would be best to adjust the cognitive score by the other variables. How could I do that? Thanks in advance!
2
-
2$\begingroup$ Good that you caught that. Could you please edit the question to say more about the outcome variable for cognitive damage/impairment? From what I know about such measures, a standard linear regression might not be the best way to adjust the outcomes for these other variables. Also, do you have the HbA1C (or other continuous) values that were used to categorize the level of diabetes? In general, it's best not to categorize continuous variables; you can always illustrate estimates at particular HbA1C values after you build the model. $\endgroup$– EdMCommented Aug 26 at 19:08
-
$\begingroup$ Thank you for your answer. In regards to the HbA1C values, unfortunately we don't have them for this specific study. Also, we are using the categorization levels of diabetes based on another study done in a center with which we have already published. Please excuse my lack of knowledge but, how would you do the standard linear regression? $\endgroup$– Alex HorrilloCommented Aug 27 at 20:32
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
You don't adjust the outcome (cognitive score) values themselves. You include the other variables (age, sex and BMI) along with the main predictor of interest (level of diabetes) into a multiple regression model. That way you adjust for potential associations of the other variables with the outcome, to evaluate how level of diabetes are associated with cognitive score when the other variables are taken into account.
Standard statistical software provides tools for doing multiple linear regression. You specify the outcome variable and the predictors of interest, along with a properly formatted data set. Details depend on the software that you use.
A few cautions. First, linear regression works best with outcome values that can be considered continuous. If your cognitive score scale just has a handful of discrete values, you might be better off with ordinal logistic regression, which only cares about the order of the outcome values. This web page has links to examples in several software packages.
You also need to think about how to model your predictor variables. With only 3 levels of diabetes, that's simplest to use as a categorical predictor; similarly for sex. BMI and age are continuous variables. If you just include them as such, you are assuming a strict linear association between each of them and the outcome values. That isn't always the case. It can help to model them flexibly, with something called a regression spline, that removes that strict assumption.
This page suggests many introductory statistics textbooks. I suspect that there are many more now available freely online. For more advanced topics (like ordinal regression and splines), Frank Harrell's Regression Modeling Strategies is a valuable resource, but it assumes a fair amount of background knowledge.
It would make sense to get local advice from an experienced statistician, as there are many potential pitfalls that might make it difficult for others to reproduce your results or even to get your results published. One particular concern (besides the implementation details noted above) that might require local advice is what you can hope to learn from such a model. If you are just evaluating associations among current levels of the variables, the model can't tell you much about how or whether changes in the predictor variables have direct effects on cognitive scores.
