 |
 |
Hi fiona The main problem is that you have lots of ivs, and spss thinks they are different, but they are all the same - e.g. Reaction time. You would probably like the effect of a covariate to be the same for each effect (I'm thinking here of a between person covariate like sex). If the effect is different you lose a bunch of df, and hence power. The bigger problem though is that you are not easily able to interpret your effect - all a manova tells you is that the variables are different (and because the effects of the covariates are different its not possible to compare them really). A multilevel model tells you how different the different variables are. What we might call 'modern' stats programs, like R (actually now I can't think of anoter) don't really have a manova command, they expect you to use a miltilevel model. Oh, now I think of it one more advantage of going multilevel is missing data. If a person misses a test or is a wild outlier and has to be deleted (which is pretty common in reaction time stuff) and you want to use manova then you need to delete the whole person. If you do a multilevel model you only need to discard that one measurement. Another thing (while I'm on a roll) is that you can have hundreds of measures per person - the more the better - in a multilevel model. Manova would get upset about that. If you want more help with the glm output, andy field's book "discovering statistics" is good. The new edition (third) also covers multilevel models (is that out yet? I don't think it is but I might not have noticed). There's also a nice book called "multilevel models: its just regression" (I forget the author). Jeremy P.s. I wrote this mostly on my phone in a dentist waiting room, so forgive typos pls. On 22 Jan 2009 15:28:40 +0000, [log in to unmask] <[log in to unmask]> wrote: > Hi, > > Yes I had suspected I might be looking at something horribly involved :( > > Regarding the previous reply suggesting adding several DVs into the > GLM-repeated measures option in SPSS - what would be the reasons for not > doing it that way? I have just quickly run that and it's given me a > stupendous amount of output (which might be one reason!), but will this > this give me increased error or something similar? > > Fiona > > On Jan 22 2009, Jeremy Miles wrote: > >>Hi Fiona, >> >>It sounds to me like you might need a multilevel regression to analyze >>these data, and if you have multiple measures (that's called a >>multivariate multilevel model) it gets really hard, so you probably >>want to treat them separately. >> >>It's very similar to a regression analysis, except that you are >>allowed to "stack" people, so a person becomes more than one row, and >>then you have a variable that identifies which is the person. Can make >>your data look like this using the 'restructure' command. >> >>You can then have two kinds of covariates - between person covariates >>are covariates that are a function of the person, and they have the >>same effect on every measure; and within person covariates which can >>be measured separately for each (say) task. >> >>It's fairly fiddly though. If you're at Herts there's a guy in the >>stats department whose name I've forgotten (Neil something?) who (I >>think) does a bit of that sort of thing. The book that Josephine >>suggested (thanks, your fiver is on the way) covers it very briefly, >>but not really in enough depth to be able to do it. A good book is by >>SInger and Willett (2003), although I've forgotten the title. :) >> >>Jeremy >> >> >>2009/1/22 Fiona Essig <[log in to unmask]>: >>> Hi all, >>> >>> Hope someone can help with this as I seem to be going round in circles. >>> I'm looking at Task Switching using some novel verbal based tasks. On >>> the task people have to manipulate two, then three, then four verbal >>> categories. They are scored on RT and accuracy. On this particular study >>> there are two different types of task (basically using different types >>> of verbal categories). So we are looking at what effect the number and >>> type of categories being used has on Rt and accuracy. >>> >>> The study is repeated measures - IV1 = Test Type (A & B), IV2 = Number of >>> Categories (2, 3 or 4), and I have 2 DVs of RT and accuracy. >>> >>> Try as I might I can't think of how to get this into one model. I've >>> been reading through Tabachnik & Fidell and I'm looking at profile >>> analysis / doubly-MANOVA design (and I'm a bit shaky on that!), but I >>> can't see how to include the second IV of test type as well. I wonder if >>> I may be able to change the way I have arranged my dataset but I really >>> can't see what to do differently. Currently I have 12 main variables of >>> interest representing the various combinations of the IVs and two DV >>> measures for each person. >>> >>> Oh and there are a bunch of covariates thrown in for good measure as >>> well! >>> >>> Previously I have used separate repeated measures ANOVAs (with adjusted >>> alpha) for RT and for accuarcy but it's not really telling me what I >>> want to know. As well as looking for main effects and interaction of the >>> IVs I want to see how the two DVs relate to each other through these >>> various combinations. I could do a profile analysis for both DVs on each >>> test type separately - is this as close as I'm going to get? >>> >>> Many thanks in anticipation :) >>> >>> Fiona >>> >>> -- >>> Tel: 01707 284 761 >>> E-mail: [log in to unmask] >>> >>> Fiona Essig >>> Research Student >>> Room E384 Research Huts >>> School of Psychology >>> University of Hertfordshire >>> College Lane >>> Hatfield >>> Herts AL10 9AB >>> >> >> >> >> > > -- > Tel: 01707 284 761 > E-mail: [log in to unmask] > > Fiona Essig > Research Student > Room E384 > Research Huts > School of Psychology > University of Hertfordshire > College Lane > Hatfield > Herts AL10 9AB > > > > -- Sent from my mobile device Jeremy Miles Psychology Research Methods Wiki: www.researchmethodsinpsychology.com