Why don't you do a multiple regression including all the cofactors?

And please do not exclude a potential confounder just because the relationship between this variable and the response is not significant. Firstly, this would mean to interpret a "non-signifcant" result (non-significant -> non-relevant; that's not a logical conclusion!), and secondly there can be a very clear relationship after adjusting for other confounders (you wouldn't find this with your method). So again, my advice is to fit a multiple regression model including all experiemtal and potentially confounding predictors. Then report the findings for the experimental factors as "statistically adjusted for age and sex (and whatever)".

Hi Jochen,

Thank you for your answer. But I don't understand this comment: "my advice is to fit a multiple regression model including all experiemtal and potentially confounding predictors. Then report the findings for the experimental factors as statistically adjusted for age and sex (and whatever)". How can I say that my variables are adjusted for sex and age if haven't adjusted them at all?

Also, I'm doing a multiple linear regression with the counfunding variables but also another regression with the adjusted variables. This way I can say, for instance, only Sex (a counfounding variable) has a significant relationship with my osseus variable, and when I adjust it by sex, my interest variables keep having a non-significant relatioship. Do I explain my self?

If your regression model is, for instance,

yi = b0 + bsex*1sex,i + bage*agei + bx*xi + ei

then you get a coefficient bx that gives you the effect of the predictor X on Y, adjusted for sex and age. The model considers sex and age and x, alltogether. You get a standard error and a p-value and a confidence interval for bx, and as sex and age are also in the model, these statistics are adjusted by age and sex.

Several examples were explained in:

1. Landau, S. and Everitt, B. S.(2004). A Handbook of Statistical analyses using SPSS.

2.Howitt, D. and Cramer, D.(2008). Introduction to SPSS.

Regards,

Zuhair

Hi Ana, it depends on what you want to achieve. If you use a multiple regression model, you take into account all confounders and the results for your variable of interest will be reported accounting for these confounders (what Jochen suggested). However, I suspect that you actually try to perform rather a normalization (remove effects of the confounders before doing other analyses with the data). If this is the case and if you are familiar with R you may try out the function 'RemoveFactorsByANOVA' from the MetabolomicsBasics package. There you can specify a model with many factors and define those factors you would like to normalize the data for. Bests, Jan

Hi Ana!

Your #1 is sound; checking out that independent variables have at least some relationship to the dependent variable is a wise procedure. Altman (1991) recommended to include variables with p-value

Thank you John, Jan and Jochen,

I have a better picture of the problem now. I think what Jochen suggest is appropriate/enough for the regression I'm trying to do.

And Jan is also right since I'm trying to remove the effect of the confounders so I can evaluate the relationship between my variables of interest and my dependent variable.

I intend to do also a correlation between them, and in this case, I think it's appropriate to adjust my variables (remove the effect of the confounders) and then do the Pearson coefficient test. So this is a different question. It makes sense that after I make a regression, as Jochen suggest, and see that my variables of interest do not have a significant relationship with my dependent variable but the confounders do, I adjust them and do the Pearson test? I hope that after adjusting the variables, my adjusted-variable of interest does have a significant relationship with my dependent variable.

So, all together I think I'll do:

1. A regression including my variables of interest and the confounders.

2. As the confounders sadly are the only ones that have a significant relationship with my dependent variables, I'll adjust them so I can remove the effect of the confounders.

3. I'll make a correlation between my adjusted variables, and so they may bear a significant relationship

I still don't get you entirely.

1 is all you need to do. If the experimental factors will explain anything that is not yet explained by any of the confounders, you will see this. 2 and 3 won't give you any additional information, as this is all already considered in 1.

To do a multivariate regression is actually 'adjustment' for all the variables you put into your model. As @Jochen says; you don't need more than #1 above.

You describe sociodemographic variables as confounders in this case. If you're interested in one particular independent variable, you can say that your result is adjusted for the sociodemographic variables you put into the model.

Best

JK