UKC

hi everyone,

some stats help sought....

i've got data that i've been advised to use ANOVA to analyse- but i'm not familiar with this technique and unclear which variety of ANOVA would be appropriate.

In summary- i've got two categories of people- one with, and one without the condition of interest; two time points- at entry to study, and after one year; and a score on a test for each person at each point. I had done t tests looking at the differences in means between groups at each time point, but a reviewer has suggested 'two-way ANOVA looking for a group by time interaction' would be more compelling.

But is this a straightforward 2 way anova? or a repeated measures 2 way anova? or a mixed model anova? I've tried reading online tutorials on this but still can't quite get my head round it- any help gratefully received!

best wishes
gregor

In reply to no_more_scotch_eggs:

Hi
See if there is a trial version of minitab to do this, it will guide you through it via the wizard

In reply to no_more_scotch_eggs:

What type of data are the test scores?

In reply to Anoetic:

But do you see an advantage for this type of data over the two t-tests that NMSE has used?

In reply to Dr.S at work:

scores on a cognitive test- so while *technically* ordinal, its the accepted practice to treat them as continuous variables

In reply to Dr.S at work:

not sure i see the advantage but its been recommended by a reviewer..!

In reply to no_more_scotch_eggs:


harrumph, down with this sort of thing!

In reply to no_more_scotch_eggs:

Its only a suggestion:

"We thank the reviewer for this suggestion but do not feel that utilising an ANOVA as suggested will add further clarity for the reader"

In reply to Dr.S at work:

i don't know if i'm feeling that brave.... another reviewer has suggested that the whole section should be dropped as its overanalysing rather flimsy data! i can't please both, but if i go with one of, rather than neither, at least i can keep an ally...!

and he may have a point- i think its likely that there were differences in the groups on the variable at baseline, which widened at follow up; t tests of patients vs controls at each time point don't capture that, nor do paired t tests over time. so i think he's right, but hasn't said which particular sort of ANOVA should be used...

In reply to no_more_scotch_eggs:

Repeated measures ANOVA, explained far better than I can do here: https://statistics.laerd.com/statistical-guides/repeated-measures-anova-sta...

In reply to ianstevens:

i'd read that- but i didn't think that looked right- as in the examples, there is only one population, with measures either at different time points, or exposed to different interventions (cake variety, in the examples given...)

i've got two groups, at two time points- so of the independent variables, 2 are linked- the same people at two time points- and two are separate- patients/controls.

so i *think* its a mixed models ANOVA- but would like someone that actually knows about this to tell me if i'm in the right area..!

In reply to no_more_scotch_eggs:
What are the mean and SD for the two groups at the two time points?

In reply to no_more_scotch_eggs:
its a mixed design

you have a between subjects factor 2 groups (people with or without)
and a within subjects factor of time

so you get a Main effect of group, a mian effect of time and an interaction, with athe control for multiple t-test for free

have you got SPSS or R ? or you could look at one of the many online versions.

and the advantage is you can see the interaction of the two factors - plot the graph are the lines parallel?

In reply to Dr.S at work:

group 1 time 1: mean 9.75 SD 2.42
group 1 time 2: mean 9.33 SD 2.64
group 2 time 1: mean 11.3 SD 1.7
group 2 time 2: mean 11.4 SD 1.58

In reply to lithos:

thanks lithos- no access to SPSS at present- hoping there would be an online tool i could use, as time is pressing!

In reply to no_more_scotch_eggs:

and how big are your groups?

In reply to Dr.S at work:

small

12 and 10 at each time point

In reply to no_more_scotch_eggs:

you can probably download minitab express and use it for free on a 30 day trial, pretty easy to use.

In reply to no_more_scotch_eggs:
eyeballing that data i'd say no effect of time but prob effect of group,
doesnt really look like an interaction from the graph.

info about factorial anova http://davidmlane.com/hyperstat/factorial_ANOVA.html
this explains it and how to do it in spss http://www.statisticshell.com/docs/mixed.pdf
another example but uses spss http://psych.unl.edu/psycrs/statpage/2x2mg.pdf

https://www.wessa.net/rwasp_Mixed%20Model%20ANOVA.wasp is an onine tool

BTW aka factorial anova or split-plot (split splot as i always called it!)

wouldn't worry too much about sample size for that, though you can do some tests - read the above

In reply to no_more_scotch_eggs:

BTW why do you say technically ordinal ? - I guess they are interval data and hence appropriate

In reply to lithos:

ordinal as they are just a summation of points gained by answering questions- the difference between say 5 and 10 isn't necessarily the same as the diffrence between 10 and 15, etc

its the same for all rating scales; but the convention appears to be to treat them in effect as ratio data, and i'm not going to dispute that...!

In reply to lithos:

does this site look like its hosting the right test...?

http://vassarstats.net/anova2corr.html

In reply to no_more_scotch_eggs:


yep, looks fine, you have 2 rows and columns iwith 10 and 12 sub rows.

those pages above should help interpret the output. The one I linked has R at its back end so the uotput may be more informative (but less easily read!)
A, B are your main effects and A X B is your interaction term.

Ok i take you word for it but some cognitive tests (IQ) are interval/ratio (theres a difference - the 0 point but who cares!)

In reply to lithos:

thanks Lithos, i'll have a good read of these, see if i can get my head round what its saying...!

In reply to no_more_scotch_eggs:

Was just about to recommend vassarstats. The reviewer is correct, by the way: No multiple t-test for repeated measurements against the same control (yes you can do corrections, but that only made sense when you were calculating the statistics by hand as it saved some work), no repeat t-tests on correlated samples, etc....

CB