I have some cells, which I've divided into three sets, and given each set a different treatment - one is control, and let's call the others chocolate and beer (mmmmmmmmmm). I have measured each set's reaction to this, and I want to see whether either of the treatments cause the cells to react in a significantly different way to the control set.

Am I right in thinking that I need a student's t-test for control vs chocolate, and then control vs beer? And could I then do the same test for chocolate vs beer? Or is there some test which takes all three into account at once?

I'm obviously a bit of a stats newbie, and people here seem to have stats, so if anyone can help, it would be most appreciated :)

Matt

You need to do an ANOVA - which will control for familywise error (which you will get if you do multiple t-tests). Andy Field has a load of chapters about anovas and the various kinds. Sounds like you need independent groups - BUT have a look through, you may need repeated measures or something or a mixed model. He has all the info at

www.statisticshell.com

Would a one-way ANOVA with a post-test to compare to the control colum be right/wrong in this situation? Just out of interest, I'm rubbish at stats :p

Sounds about right. If its 1x3 then its a one way anova. if its 2x3 or more then you have to start looking at other options.

Yeah, I would say one-way anova with a post-hoc test to check for significant differences between groups if any crop up! Might be worth investing in a stats book or an SPSS guide if you're using SPSS. I have a really basic SPPS Survival Guide book and it's great, even now when I've been studying stats for about 7 years! Enjoy! KB

If you do post hoc tests i.e. t-tests afterwards to investigate further, make sure you use boneferroni correction i.e. cut your .05 p value by the number of tests you run, so you don't accidently think something is signficant when its not so if you run 2 t-tests

.05/2 = .025 so you need to work to p<.025 for significance. hope that makes sense!

2 things - am I right in thinking ANOVA is basically like a T-test which covers all combinations, so for A, B, C, it will test A v B, B v C, and A v C? 

So basically, I run an ANOVA, and look to see if P<0.05 for any combinations of groups. Then if it is, run t-test between those two groups?  I'm not sure I understand the bonferroni correction though...

your one way anova will tell you.......there is a significant difference between the 3 groups.

But it won't tell you where the difference is i.e. is group 1 signigicantly different from group 3?? and so on.

So you either do post hoc tests - these are done automatically if you select for them and they correct for familywise error rate themselves anyway, automatically.

Or you can do t-tests (this is called 'simple effects analysis') to work out the differences - if you do this you then have to manually do the 'boneferroni correction' which I described below.

Thanks for that Sneaks.  So I've done my ANOVA, got my p value, which is nice. I selected post-hoc bonferroni, but it just gives me some differences, which seem to be between one group and the other two for all cominbations, and tells me that the differences are significant at the 0.05 level.  It doesn't actually give me a p value for between two groups.

======= Date Modified 05 May 2010 17:30:22 =======
it should come up with a table (i'm assuming you're using SPSS) that has the difference between all combinations

e.g. 1 vs 2 and then 3, 2vs 1 and then 3, 3 vs 1 and then 2. and there will be a significance column on the table

Will look a little like the first table here http://faculty.chass.ncsu.edu/garson/PA765/anova26.jpg

Nowwww I see it.  So really I don't have to do a t-test because it's done it all for me, AND the Bonferroni correction means that the p-values are as they should be, and don't need dividing by 2 any more...

Nope, if you've done post hocs its all done for you. You may want to have a flick through andy field though, as there are different types of post hocs and they do slightly different things. They will probably come out with the same results, but if its for your PhD you may want to check to avoid Viva scrutiny. e.g. there is sidak, boneferroni and tukey hsd I think - may be more!

a ha - see page 4 of http://www.statisticshell.com/contrasts.pdf it has a few general ideas on the different post hocs you can choose.

I'm going to stick with Bonferroni because it's not going in my viva, and it's what my supervisor's used in the past! Thanks very much for your help with that :)

Matt

No probs - make sure you have done all the pre-checks though before you run the ANOVA e.g. normality, homogeneity of variance etc.