Stats help with repeated measures for a single group

PoleStar

Hi guys

I originally posted this int he maths forum but prob the wrong place so will try here!

Im stumped, Im a medical research with basic stats knowlege normally limited to simple one way ANOVA and t-tests etc ha!

I have a single group of 5 subjects who do not differ. They each undergo an intervention which is the same in all subjects. A single variable is measured before the intervention i.e. at baseline, and at 2, 4, 6, and 8 hours after the intervention.

I need to know if the intervention changes the level of the variable from baseline. I.e. basically a paired t-test to compare 2, 4, 6, and 8 hours with baseline. I assume though because there are multiple time points that this is not valid.
Im using SPSS by the way!

Is it valid to do a one way anova with Dunnetts test to compare all time points with the baseline which I will set as control? Or is this not suitable because I would be treating all times as separate groups and thus the analysis would not be paired? Is there a way to pair the values within subjects?

Checking SPSS help suggests using the GLM repeated measures however the one similar example I found has a between subjects factor of gender and thus Im not sure if this is ok. I also checked minitab help but most of the examples are very complicated with different factors etc.

Also, while I tried GLM with repeated measures, it only gives comparisons of time 1 with time 2, time 2 with time 3 etc i.e. how the variable changes over time, where I want pairwise comparisons with baseline, i.e. time 1 with time 2, time 1 with time 3, time 1 with time 4 etc

Any ideas?

And how to do this with SPSS. I also have minitab but am stumped as to how to use it!

2Scoops

PoleStar wrote:

Is it valid to do a one way anova with Dunnetts test to compare all time points with the baseline which I will set as control?

It’s valid but, with such a small sample size, you’re shooting yourself in the foot and I wouldn’t be surprised if an unpaired test said there were no differences even if there probably were. A repeated measures design will improve your power to detect a real difference substantially.

PoleStar wrote:

Also, while I tried GLM with repeated measures, it only gives comparisons of time 1 with time 2, time 2 with time 3 etc i.e. how the variable changes over time, where I want pairwise comparisons with baseline, i.e. time 1 with time 2, time 1 with time 3, time 1 with time 4 etc

It sounds like you did contrasts through the repeated measures window system. Change the contrast type to simple and list your control group as the reference category. Alternatively, go to the options tab in the same window and compare the estimated marginal means which are post hoc tests.

PoleStar

Thanks that seems to be working better now

One problem though

In SPSS it states you need to check Mauchlys test of sphericity as if this is rejected you need to use one of the other tests to check for signicicance eg Greenhouse Geisser.

However, in the output, it gives for example, a value of 0.000 for Mauchly's W, not chi squared, a df of eg 9, but no sig.

Does this mean there is some problem? Or is it due to
a) perhaps this test of sphericity is irrelevant as it is a single variable that is being measured
b) because of the large number of timepoints

I think I remember seeing something about this before although cant remember.

And then, which test of within subject effects do I use?

MORE CONFUSION

Ha thanks

2Scoops

PoleStar wrote: »

In SPSS it states you need to check Mauchlys test of sphericity as if this is rejected you need to use one of the other tests to check for signicicance eg Greenhouse Geisser.

However, in the output, it gives for example, a value of 0.000 for Mauchly's W, not chi squared, a df of eg 9, but no sig.

Not sure what you mean here, I'd have to see the output. Usually though, violating sphericity is not a big deal as long as the epsilon values are close to 1. What is the Greenhouse-Geisser epsilon value? If it's way off, you can use multivariate analysis instead.

PoleStar wrote: »

Does this mean there is some problem? Or is it due to
a) perhaps this test of sphericity is irrelevant as it is a single variable that is being measured
b) because of the large number of timepoints

No. It simply means the variances at each time point are not similar - a common problem with smaller samples.

PoleStar wrote: »

And then, which test of within subject effects do I use?

Not sure what you mean here either. Contrasts should test the main effect of time point. Or, you could use LSD and the Bonferroni correction with the EM means and see what you come up with.

PoleStar

Hi thanks again for your help.

Im curious now as to what you work at as you really seem to know SPSS!!

Anyhow what you said about using contrasts made sense, when I use a simple contrast to compare to the first timepoint then I get the comparisons I need so I guess I dont need to use the bonferroni or LSD options right?

As for the Mauchly question, Its difficult to explain, shall I send you the output file?

Thanks!

2Scoops

PoleStar wrote: »

Im curious now as to what you work at as you really seem to know SPSS!!

Jack-of-all-trades.

PoleStar wrote: »

Anyhow what you said about using contrasts made sense, when I use a simple contrast to compare to the first timepoint then I get the comparisons I need so I guess I dont need to use the bonferroni or LSD options right?

Right, but you should know that contrasts are more likely to draw criticism from reviewers if you're intending to publish this investigation. If the data holds with the post hocs, use them instead.

PoleStar wrote: »

As for the Mauchly question, Its difficult to explain, shall I send you the output file?

I'll have a look if you think it will help, but it looks like you're on the right track in general.

PoleStar

Thanks for your advice.

Contrasts do seem to work fine. Obviosuly using something like the bonferroni is not desirable because for some of the variables there will be 9 time points which equals no chance of getting significance!

I have attached the Mauchly's test of spehricity from the output here.

As you can see, where there should be a reading for Approx Chi-Square and the Sig, all I see is a full stop. In this particular test, I had 5 time points and 5 subjects, placed as always in 5 columns.

Nothing funny about the data so not sure why this is occuring.

Interestingly, when I just took 3 of the time points, it will give a reading. Doesnt make sense to me. But means then that for the test of within subject effects Im not sure which p value to take, the one for spehricity assumed, or one of the others which are adjusted, Greenhouse etc.

Thanks!

2Scoops

PoleStar wrote: »

I have attached the Mauchly's test of spehricity from the output here.

As you can see, where there should be a reading for Approx Chi-Square and the Sig, all I see is a full stop. In this particular test, I had 5 time points and 5 subjects, placed as always in 5 columns.

Not sure why this is - are there any missing datapoints? I think it might be because Mauchly's W statistic is so low that nothing else can be calculated... maybe.

PoleStar wrote: »

But means then that for the test of within subject effects Im not sure which p value to take, the one for spehricity assumed, or one of the others which are adjusted, Greenhouse etc.

Now for the bad news! :pac: If Mauchly's was not significant you could use the sphericity assumed p value. If it was significant but Greenhouse-Geisser or Huynh-Feldt were close to 1 (i.e. >.8) you could use one one of those correction factors. However, since they are so low it looks like the assumption of sphericity is badly violated. This may mean that you can't really trust the repeated measures output - but it's difficult for me to tell by how much. Try multivariate analysis and see if it can hold significance. Is it possible to enroll more subjects in this experiment - it would help out a lot, I think.

RubenJ

Not very familiar with GLM in SPSS so maybe this is already what 2scoops has suggested but you could do a mixed model with subject as the random effect and take into account the fact that the measurements are serially correlated.

my 2cent