Statistics Questions

Permalink Reply by Harry Lim on October 7, 2011 at 9:33

Hi Andy,

I carried out a Condition (4 levels) x Distance (2 levels) repeated measures ANOVA, with bonferroni adjustment. There was a significant main effect for condition, but pairwise comparisons on SPSS did not show any significant effects between any conditions. I wanted to find out if the right thing to do is to carry out multiple t-tests, and then use a corrected p value to test for significance?

Permalink Reply by Professor Andy Field on October 17, 2011 at 13:56

In programmes like SPSS/R you can get it to do the correction for you. The trouble with post hogs is that they can be relatively insensitive, which might be why you're missing the difference that is driving the main effect. In general you're probably better off just looking at the confidence intervals around the group means, computing some effect sizes for differences between conditions, and using those for interpretation. Also, if your condition * distance interaction is significant then ignore the main effects altogether.

andy

Permalink Reply by Angel Carrasco on October 24, 2011 at 8:27

Maybe I run the risk of being annoying but I have realized I posted this in the wrong place so I'll copypaste plus add some clarification:

Dear Andy and everybody,

To keep it short, I have tried to replicate research with a mixed design in which anxiety patients and controls (that was the between-subjects IV) completed a task which had different levels of anxiety-provoking or neutral stimuli (the within-subjects IV) and their response latency was measured. In fact, I have two versions of the task one of which has two within-subjects IVs. However, I do not have patients but undergraduate students. I initially thought about dychotomizing my sample using a median split in an anxiety measure but I have learnt since that everybody is against this nowadays and for good reasons. I have been reading about interaction effects in multiple regression which seems like the best alternative but I haven't found yet clear instructions for repeated measures designs. Maybe I could make an ANCOVA but my primary interest is not in controlling for the effect of anxiety, but the interaction with the IV, measured in a continuous way by a questionnaire. I would be very grateful about some advice, even if it is something like to carefully read Aiken and West's book. Thanks in advance.

A.C.

Update: The problem is, of course, not simply to calculate and interpret an interaction effect in a repeated measures ANOVA. The interaction that interests me is one between a continuous variable and a repeated measures categorical effect. So everybody's saying that I should use RM and not doing median splits, but I can only find instructions for factorial between-subjects designs. Thanks again for reading this.

Permalink Reply by Michael Crites on October 26, 2011 at 17:29

Hey There,

I am a graduate student, currently working on my Master’s
thesis. I am attempting to analyze differential carryover effects for a
within-subjects design. I have three conditions, counterbalanced of course, and
I would like to see how early exposure to one or two conditions affects the
other(s). If anyone knows of anything to point me in the correct direction on
how to set up and analyze this in SPSS, or any other statistical programs, I
would greatly appreciate it. Also, I have been trying to find research that
have already analyzed a similar design but have not had any luck there. So if
this approach rings any memory bells, then please pass along that info.

Thank you very much for your time.

-Mike

Permalink Reply by Krista Ritchie on November 9, 2011 at 15:38

so sorry - i just sent an e-mail with the below questions to all members of this group. Clearly a first time user.

Two questions regarding specifics of Multiple Imputation in SPSS

A: The default setting for m is 5. Does the program make m imputations required for your data set (sort of like iterations) or does it create m data sets that are specified. I have generally read that m is usually between 3 and 10. If we set m to ten, will it stop when it needs to (e.g., maybe 8) or does it compute 10 imputations?

B: Currently we have 5 imputed data sets. What do we use as our "new" data set? Do we aggregate these sets somehow? What is the next step so we can get back to hypothesis testing? :)

Permalink Reply by Jeremy Miles on November 9, 2011 at 23:15

I think that SPSS makes m datasets (I'm not sure, I haven't used SPSS for this). You then analyze these m datasets, and combine the results, there should be a way of doing this in SPSS.

SPSS makes as many datasets as you tell it to. Although 3 or 5 is often used as the standard, you sometimes need many more. I've needed 50 on one occasion, and I've heard of people needing hundreds. When you analyze the data, you'll get a t-statistic, this tells you if you had enough imputations.

I suggest the book Missing Data, by Paul Allison.

Permalink Reply by Krista Ritchie on November 10, 2011 at 15:24

thank you for the response. I will check out the book, and keep increasing m until the t-statistic comes out significant. I would like specific direction about combining the results.

Permalink Reply by Krista Ritchie on November 17, 2011 at 14:03

Anyone able to give direction about combining the results from multiple imputations of missing data from SPSS?

Permalink Reply by Kimmie-Louise on December 4, 2011 at 17:35

Hi,

I'm confusing the hell out of myself so could someone please clarify if a mixed factorial ANOVA is the same as a repeated measures ANOVA.

The design I've currently been forced to look at has 120 participants split into 4 groups. It is a 2 (gender) x 2 (treatment condition) x 3 (time) repeated measure design, where there are: 2 between subjects factors (gender and treatment) and 1 within subject factor (time). Dependent variable is reported alcohol consumption.

Apparently a mixed ANOVA is a mixture of between group and repeated measures variables and should have at least one d.v (alcohol consumption) and at least 2 i.v's with at least one between groups. (course and gender) and one within groups (time). This allows you to test two between groups (course and sex) and one repeated measures (time) though it's a more complex design. As my question falls into this design I was just wondering if some bright spark decided to use different names for the same test or if they are completely different. Thanks :)

Permalink Reply by RH on December 7, 2011 at 16:04

Hi

Technically a repeated measures factorial ANOVA is where all your IVs are repeated measures (i.e. there are no independent groups) . The Mixed ANOVAl is where you have at least one repeated measures IV and at least one between-subjects IV.

The confusion comes because in SPSS you have to use the repeated-measures option under the GLM to do the mixed ANOVA.

In short, to analyse your experiment do Analyse>GLM>Repeated Measure. Then put the columns which are your between-subjects factors into the 'Between-Subjects Factors' box.

Permalink Reply by Oscar on December 8, 2011 at 10:19

repeated-measures and mixed-designs ANOVA are different types of techniques. mixed-design ANOVA is more flexible because it allows you to bring in both between-subjects and within-subjects factors. repeated-design ANOVA is better suited to handle variables that get repeated. time is the most common one but it could also be over space or conditions (for example, if i were to apply a treatment on the arms of people. they're correlated because they belong to the same person and the dyad is the unit of analysis, not the individual arms).

in your case you identified your design correctly as as a mixed-design ANOVA

Permalink Reply by David Kronemyer on December 4, 2011 at 19:06

Good morning Dr. Field, I started off with 32 predictor variables and successfully completed factor analysis to reduce them to four constructs. No real interest in assigning factor loadings to each one of the predictors. Rather I want to use the four factor scores as new predictor variables in a regression analysis. At p. 636 text states this can be done but doesn't set forth an operational procedure. What SPSS output does one use? Are they coefficients, do you have to multiply them by something (what?), or add them together? Several of the original predictor variables used different measurement scales. At p. 633 the text says that as a result, different factor scores can't be compared. Do you have to normalize them in order to use them as new predictors? Or is this obviated by using the "Anderson-Rubin method" described at p. 635? I could use some advice from Jane Superbrain, thanks in advance. DAVID

Replies to This Discussion

Follow us: