I'm not as familiar with the second book, but I don't think it's
super-specific to Stata. Basically, you need to know how to interpret
the multinomial logistic regression and that's the same whichever
program you're using.
I'm not sure about the SPSS algorithms, but the Stata ones tend to be
kind of slow and a little flaky - they fall apart relatively easily, I
think SPSS's might be better, and quicker (but i'm not sure about
that).
Here's a nerdy joke;
Knock knock.
Who's there?
[wait]
[wait]
[wait]
[wait]
[wait]
[wait]
Multilevel model running in Stata!
The number of participants is more important than the number of
trials. 100 participants sounds good. If you'd have had 20 or 30
participants, I'd be nervous. You can run the analysis with one trial
per person, so that part doesn't matter (but more trials per person is
better).
I'll be at Essex in July (I teach a course on regression in the Summer
School), but you probably don't want to wait that long.
J
On 19 December 2012 14:02, Kevin Glover <[log in to unmask]> wrote:
> Hi Jeremy,
>
> There are two books by J Scott Long, 1997 and 2005 - the latter is specific
> to Stata. I can get both from the library, and what else would I do over
> Christmas but read about regression?
>
> Essex has Stata on campus but I'm a 'distance learner' so would have to
> purchase a licence myself. Not impossible, but SPSS 19 is free! And I have
> never used Stata.
>
> Based on bitter experience, I have been trying to figure out the analysis
> method at the same time as designing the experiment rather than afterwards,
> so sample size is a guess, but I think at least 2000 trials (20 sentences x
> 100 participants). Yet more bitter experience tells me that beyond 20
> sentences fatigue sets in.
>
> Thanks again for your help with this, much appreciated.
>
> Kevin
>
>> Date: Wed, 19 Dec 2012 13:09:11 -0800
>
>> From: [log in to unmask]
>> Subject: Re: Predicting the 'Winner'
>> To: [log in to unmask]
>>
>> On 19 December 2012 12:01, Kevin Glover <[log in to unmask]> wrote:
>> > Hi Jeremy,
>> >
>> > Thanks - your simplification looks good. The characteristics of the
>> > participants should not be a factor - items only. And fortunately I do
>> > have
>> > SPSS 19, courtesy of my university (Essex). So now I just need to
>> > understand
>> > tricky-squared, or I have to ask you to point me to an idiot's guide (my
>> > understanding of logistic regression has so far been acquired from
>> > YouTube
>> > videos). Thanks again.
>> >
>>
>>
>> Hi Kevin,
>>
>> You might want to take into account participant characteristics (if
>> only gender) because it might account for some variation in responses.
>> That means it soaks up a bit of error variance and thereby increases
>> power. Also, by soaking up a bit of error variance it can help the
>> model out and mean it's more likely to converge (what's your sample
>> size, by the way? And in a multilevel model you have two sample sizes
>> - the number of level 2 units (which is people in your case) and the
>> number of level 1 units (which is trials).
>>
>> Multinomal logistic regression is a relatively straightforward
>> (relatively!) from binary logistic regression - it's just that you
>> have a dummy coded outcome variable. The best description I've seen is
>> in a book by J Scott Long, called Regression Models for Categorical
>> and Limited Dependent Variables (or something like that).
>>
>> I've never seen a nice description of multilevel logistic regression
>> models, and I haven't seen many non-nice descriptions. There are all
>> kinds of horrible complications, such as that you don't really have
>> error variance in a multilevel model, so it doesn't work to estimate
>> the correlations of error variance. which is what you usually do in a
>> multilevel model (and there are other issues that I can't even
>> remember). My usual approach here is to say "I have done what I am
>> supposed to do, and
>> SPSS employs some fairly clever people who understand this, and make
>> sure that doing what I'm supposed to do will work"/
>>
>> Depending on your sample size (and other things) there's a chance that
>> the models won't run at all. Your fallback, if that happens, is either
>> complex samples or generalized estimating equations in SPSS. But
>> let's talk about that if it happens. :)
>>
>> J
|