selecting ANOVA over T test

[jeffreyg]

Hello,
I am a graduate student in science preparing for my defence and taking a bit of heat regarding the statistical test I chose for an experiment.
I am comparing feeding in gene KO and control mice (11 each). I measured food cosumption in these 2 groups at 5 time points over a 24h period (after fasting). 1, 2, 4, 8 and 24hrs. I conducted unpaired t-tests on the two groups at each time point. The only point that has p<0.05 is the 1hr time (which makes sense in the context of genetic KO animal). Im now being told that this is incorrect as there are multiple time points and as such i need to complete ANOVA so that im not reporting a type 1 error. My question is why do I need to complete ANOVA (which would compare the time points to each other) If all I care about is weather the KO ate less then the WT at any given point... If I do need to do ANOVA how would I go about it.

I hope this is clear,
thanks in advance
Jeff

[Philyuko]

That looks like repeated measures ANOVA

If you run all the tests seperately then you're not helping yourself - you are basically saying that each test is independent, unrelated, separate. With the ANOVA, you are saying from the get-go "Hold up 'ere pal, I'm going to do a bunch of tests; they're all important."

Now - if you do a planned comparison - select Bonferoni - then you can see differences between groups - not just a single over-all measure.

Bottom line - this is classic ANOVA - and you seem to have walked into the classic trap. If you google ANOVA/t-test and something like "why should I bother?" then it'll practically have your name and this data on it ... which is great because you'll get all the examples and explanations you could ever shake a stick at.

[jeffreyg]

Thanks philyuko,

I have set this up using prism. the bonferoni tests are comparing each time point to eachother and not WT and KO within 1 time point. WT is being campared to WT between individual time points. I dont care how food consumption varied between time points for a given genotype. I want to know if the KO ate less then WT at that given time point.

have i got my data set up wrong?

thanks

[Philyuko]

WT?

and don't know prism ... should mention that ...

[Philyuko]

oh wait ... KO is "knock out" - a gene removed? and WT is "wild type", right?

I'm an applied linguist in a psychology department so not too familiar with all the terms from the "Real" sciences ...

[Philyuko]

Still looks like Repeated measures ANOVA
within factors are time intervals
between factors are type (KO and Wild)
Looks like it's the between that you're interested in (type) but there may be an interaction of type and time_interval.

Either way - it's an ANOVA (repeated)

If you do five different t-tests (KO vs WT) then you're getting five bites at the cherry. Now, you could take Alpha (.05) and divide by five (.01) and call that alpha ... although in your field I think you set alpha lower from the start ... anyway, you get the point. But, really, I say it again, the ANOVA seems like the way to go - you'll get the right result and one that favors the math much more so.

[Philyuko]

wow ... one more ... I meant "mixed" ANOVA

This is the only paper I've had to do any extended kind of mixed ANOVA - and it was quite a while ago ... may help though ...

http://csep.psyc.memphis.edu/mcnamar...naturesLSA.pdf

[jeffreyg]

I see. thanks for the help and the resources. It appears my feeding difference is no longer significant when examined with the proper stats test.

Thanks again for the clarification

jeff

[Philyuko]

There are "trend tests" that you could still look up ...

I would also strongly recommend that you look carefully at the effect size ... at the end of the day, "significant" don't mean nuffing (as they say) ... it's only the effect that anyone one cares about (outside of student courses) ... you can have high sig with low effects and pretty good effects without "significance" ... in other words, "significant" isn't all that significant (but don't tell anyone) ... so, maybe your story lies in the effects ...

Also - just look at the graph ... does it LOOK like a story? If it does then you do have a story - you just have to find the right analysis for it. Are the results in the predcited direction? If so, again, Ps can become very flexible. At the end of the day, stats don't tell you that much more than your eyes ... though they may seem more convincing.

[jeffreyg]

trend test? effect test? Does prism have these tests or do I need the pencil and paper?

[Philyuko]

don't know prism ...

you'd have to look it up ...

sorry ;-(

but ESizes are not too hard by pen/paper ...
Still, I'd be doing partial eta squared ...

good luck!