Close Home Forum Sign up / Log in

Moderation analysis with multiple regression

S

I've been going around in circles with this so any suggestions/feedback would be great.

I have:
3 predictors/IVs
2 moderators
2 criteria/DVs

I've centered the data, calculated interaction variables, etc the bit I'm unsure about is the actual multiple regression.

Initially I'd done individual regressions for each IV/mod/DV combination but I've realised that I'll be losing lots of power as will need to divide p by the number of tests. I don't think I can do 3IVs/2mods for each DV, which would mean I get two equations, or can I? Or do I just group the predictors and calculate different equations for all the possible mod/DV combinations which would then give me 4 equations?

Any ideas are welcome!

Are your moderators moderating just 1 IV or are they moderating all 3?

If you're doing it in SPSS I'd be tempted to run it separatly, so control from IV 1 and IV2 and then create interaction terms and run a 3-way interaction on DV1, then the same on DV2

And then repeat with various cominations if you want to see the moderation of specific variables.

For a 3 way interaction in SPSS you'd want to put in hierarchical regression

Block 1

Iv
M1
M2

Block 2

IV x M1
IV x M2
M1 x M2

Block 3

IV x M1 x M2


But if you want to moderate all IV's you'd probably be better off getting your head around MPlus or other SEM software

S

======= Date Modified 14 Sep 2011 20:05:27 =======
======= Date Modified 14 Sep 2011 20:05:09 =======
Thanks for that Sneaks! I'm afraid moderators moderate all IVs but I only have access to SPSS so I'll just have to go with the above.

I hadn't thought of your way at all so thank you!


S

Erm.. how do you control for variables? I'd thought you had to enter them in a block alone first but you can only have 3 blocks so I wouldn't have enough blocks. Anyone?

you can have more than 3 blocks - you can just stick your controls in block 1 and move the rest down.

I'd conduct an analysis like that for each of the 3 IVs on each of the 2 DVs so 6 analyses in total. I'd then put any non-sig ones in the appendix to make it easier to read (if this is for your PhD).

If you have any sig results you can use the spreadsheets here to make the graphs http://www.jeremydawson.co.uk/slopes.htm

S

I did that, I set it up with 4 blocks but output only gave me 3 models and there's something wrong with the last one. It should have IVxM1, IVxM2 and M1xM2 but it's not taking into consideration M1xM2. There's an error saying 'Tolerance = .000 limits reached' any ideas?

how many participants do you have? maybe there's not enough power?

S

Just over 150 but it's only taking 139 into consideration, probably due to missing values as I couldn't replace all of them due to their nature. I feel like crying but instead I'm going to email sups and have a cup of tea before I come back and think about it again.

You could reduce your control variables. Maybe run it with pairwise? (I've read that's a bit dodgy but probably worth a shot), or you could boostrap it if you have the newer version of SPSS to make any findings more robust.

S

I don't think I can reduce my controls and I don't have bootstrapping. How would you run it pairwise though? Stats is def not one of my strengths...

I think there's an option in one of the buttons on the regression pop up that says listwise or pairwise.

Your uni will probably have the bootrapping add on on their lab computers?

S

I'll have a look at the options now for the pairwise function, I'm afraid I'm 2hrs away from uni at the moment so I can't use their comps :( btw, thank you for all this advice, I REALLY appreciate it!

if it is a power issue then you might be best just testing 1 IV at a time with the 2 moderators, so not including the other 2 IVs in every analysis.

S

I've been thinking about dropping the controls altogether. Is that what you mean?

well if you have controls (e.g. gender, age) you could drop them and then talk about their possible effects later.

You also have IV1 IV2 and IV2. My understanding is that you want to see whether the relationship between each IV is moderated by 2 moderators.

So rather than chucking IV1 in and seeing if M1 and M2 moderate it *whilst controlling for IV2 and IV3*, i'd drop the latter bit, and just run each IV moderated separately soo..

Model 1 = IV1, M1, M2, DV1
Model 2 = IV1, M1, M2, DV2
Model 3 = IV2, M1, M2, DV1
Model 4 = IV2, M1, M2, DV2
Model 5 = IV3, M1, M2, DV1
Model 6 = IV3, M1, M2, DV2

But I don't know what the variables are so that might be a bit rubbish if your varibales are all supposed to interlink.

20448