Hi everyone,
I need some help--and I'm hoping that all of you who are better at statistics than me will be able to help the fog lift!
Background to my problem:
I have done a three-way ANOVA (with each of the three between subjects factors having two levels, so 2 x 2 x 2) with robust estimates of standard error. Or at least I think I have. I used the the "Parameter estimates with robust standard errors" set at "HC3" to generate my parameters as recommended by Cai and Hayes (2007) and others. (For amateurs like me reading this: https://www.ibm.com/support/knowledgecenter/en/SSLVMB_25.0.0/statistics_reference_project_ddita/spss/base/syn_unianova_robust.html, might help you).
Just to clarify, I have used robust estimates, as my Levene's test was highly significant (p
Hi Bruce Weaver . Thanks so much for replying so quickly and helpfully! I appreciate it greatly; and I'm sure many others on Research Gate appreciate your helpful posts too.
To follow up, for others who may be reading this with the same problem and have a similar level of ability with software and statistics:
Even with Bruce Weaver's help and the documents provided, I've really struggled to do it with the LMATRIX in spss. So I've learned how to do it in R, after all.
I'd strongly recommend doing it in R.
If you don't know how to use R, here's a few bits of advice on how to do a three way ANOVA with robust standard errors there:
1) Download R studio -- it's much easier to use than R if you're an SPSS user usually. It's a half-way house between SPSS and R.
2) If you don't know how to use R, I'd recommend doing any recoding and exclusions in SPSS, and then importing the file to R just to do the three-way ANOVA with robust SEs.
3) When you open R, you need to install the 'car' package. Here's the manual on the car package (https://cran.r-project.org/web/packages/car/car.pdf). Here's a guide to installing r packages: https://www.r-bloggers.com/how-to-install-packages-on-r-screenshots/
(You may have some trouble with this; I did--lots of error messages; there are forums that advise on these--if this happens to you, you are not the only one.)
4) The general form of the code for running a three-way anova once you have the car installed is as follows:
> name name
Hello Douglas Flint . I only have time for a quick look at your question right now, but I think what you've run into is the same thing I did when I first tried using the ROBUST option for UNIANOVA (or equivalently, GLM). I expected that the F-tests in the Between-Ss effects table would use the robust covariance matrix, and ditto for tests obtained via the EMMEANS sub-command. But I was disappointed to discover that they did not--those tests were not affected by the presence of a ROBUST sub-command. The only thing that is affected by including a ROBUST sub-command, it seems, is the table of parameter estimates.
Somewhere, I have a pair of syntax files showing results for such a problem using both SPSS (UNIANOVA) and Stata (-regress-). In Stata, I follow up with -margins- commands to get the comparisons I get via EMMEANS in SPSS. In Stata, the comparisons obtained via -margins- DO use the robust covariance matrix if I incude a vce(robust) option, or vce(hc3), etc. So the short answer is that I think the way SPSS currently handles this is not ideal, and if you have access to some other software that handles it more appropriately (IMO), it might be worth using that software for this analysis. I can't find the aforementioned example right now, but if/when I do, I'll share it here.
Cheers,
Bruce
Hi Bruce Weaver , that's really good of you to get back so quickly on this.
That's bad news. At least I don't feel so bad about being stumped by the output.
Unfortunately, I don't have access to another statistics programme at the moment.
Does anyone else know of a line of syntax that can force SPSS to use the robust estimates of standard error in generating the Between-SS Effects Table so I can just report a three-way ANOVA using heteroskedasticity consistent SE?
Failing that, does anyone have any suggestions as to how I could report the output that I actually do have? For example, could I say I ran multiple regression, and just report the parameter estimates for the terms of interest?
(Then again, I suppose I can't say that, as I don't have any statistics for the model in general; and I don't really want to stray from an ANOVA, as I think that's what the reviewers and readers will expect...)
Hi Douglas Flint. I have now found the example I mentioned above. It turns out that if you use UNIANOVA's LMATRIX sub-command to carry out the contrasts of interest, it will give two sets of results, one using the usual asymptotic covariance matrix, and one using the robust covariance matrix. I've attached my syntax file and an Excel file showing the exported results. It's not the same 2x2x2 ANOVA design you have, but it should give you the general idea. Getting the contrasts you want via LMATRIX is less straightforward than getting them via EMMEANS with COMPARE, but it can be done.
For those who are interested, I've attached the corresponding Stata DO file. (The .DO file and the .sps file are both plain text, for those who do not use SPSS or Stata.)
HTH.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA