Panayiotis Louca, you'll increase your chances of getting some good help if you show what you've already tried (e.g., post your syntax) and explain clearly where things are going wrong.

If you don't know now to even start, note that there are many examples online showing how to perform hierarchical linear regression with SPSS--e.g., https://www-01.ibm.com/support/docview.wss?uid=swg21476749. Try a Google search on to find many more.

If any of your explanatory variables are categorical, you'll also want to Google for info on how to treat those variables.

HTH.

Hello Panayiotis,

For the question of relationship with systolic pressure, I agree with Bruce Weaver that hierarchical (linear) regression is a good choice. For the second question, of relationship with hypertension status, hierarchical logistic regression would be a good choice.

In SPSS, for the linear regression, just enter the "control" variables as the first stage, then enter your metabolite measure as the IV for the second stage.

Good luck with your work!

Both of the previous two responses are quite useful. The main thing I would add is that the term "hierarchical" refers to the order in which you enter more than one set of independent variables As David Morse indicates, you begin by entering the control variables, and then conduct a further regression that tests the effect of the key variable(s), above and beyond the control variables.

The order of entry that the two Davids* are advocating--control variables first, then the focal variable--is very common in fields like Psychology. But in medicine, epidemiology and health sciences generally, it is quite conventional to look at the crude effect of the focal variable first, then at the adjusted effect. Panayiotis Louca is in a department of Health and Life Sciences, and has said this is the approach he wants to use. In that case, he should enter just the focal variable on step 1 and add the control variables on step 2. HTH.

* Not to be confused with The Two Ronnies. ;-)

Bruce Weaver, thank you for the insight. Of course, the real issue is whether the key variable remains significant after controlling for the other effects, but it is interesting that different fields approach this issue from different directions.

No worries, David. As someone trained in psychology who later worked for folks doing medical research, I had to learn some new terminology and get used to some different conventions. ;-)

Another interesting possibility is that the crude effect of X is small and not statistically significant, whereas the adjusted effect is larger and statistically significant. Epidemiologists describe this as negative confounding, whereas folks in Psychology, Sociology and related fields usually call it suppression.

One more example: Moderation (Psychology) vs effect modification (Epidemiology) to describe interaction. Note that moderation was originally used to describe an interaction where the X-Y association became weaker as variable Z (the moderator) got larger. If the X-Y association became stronger as Z increased in value, that was called augmentation (or synergy), not moderation. Nowadays, that distinction has been lost, as far as I can tell: Most people use moderation as a synonym for interaction. Effect modification, on the other hand, is completely neutral as to the type of interaction. That's why I prefer it.

Bruce Weaver, I was trained in sociology, where mediation was a matter of direct and indirect effects (which could indeed show suppression), and interaction was (and mostly still is) interaction.

Oh well...

I would just add a tad to the good advice that you have been given if your main goal is prediction. I would first examine this paper to be clear on which of the 2 types of model you want.

https://www.researchgate.net/publication/48178170_To_Explain_or_to_Predict

If you choose prediction then I have added a paper that might also be of value to you. Hope this helps. Best, David Booth

Thanks to all above. Your inputs are very much appreciated.

I agree with David Morse