Hi all
You may or may not remember my query about using Chi square vs multinomial logistic regression to look at categorical predictors (all yes or no) and categorical 'outcomes'.
I have opted to change the outcome to 2 categories, rather than 3, so am I right in thinking that a standard logistic regression would be appropriate?
Many thanks for your help,
Fleur-Michelle
_________________
Fleur-Michelle Coiffait
Trainee Clinical Psychologist
University of Edinburgh & NHS Lothian
On 12 Oct 2010, at 18:58, Jeremy Miles wrote:
> One thing to do is to try it. :)
>
> The other thing to do is to think about the number of 'cells' in the
> analysis. You need to ensure that every combination of predictor and
> outcome has a few (say 5-10) people in it (that is you want at least 5
> people with risk factor 1 in each of your three outcome groups, and
> five people without risk factor 1 in each of your three outcome
> groups). I think you're going to be pushing that. I would do them
> one risk factor at a time, using either chi-square tests or
> multinomial logistic regression - they're the same.
>
>
> Jeremy
>
>
>
> On 12 October 2010 10:01, Fleur-Michelle Coiffait
> <[log in to unmask]> wrote:
>> Thanks to all who have responded to my query.
>>
>> Just wondering if you had any idea how large the sample would need to be to use multiple predictors in a multinomial logistic regression? My sample is 90.
>>
>> Many thanks,
>> _________________
>> Fleur-Michelle Coiffait
>> Trainee Clinical Psychologist
>> University of Edinburgh & NHS Lothian
>>
>> On 12 Oct 2010, at 17:50, Jeremy Miles wrote:
>>
>>> Hi Fleur-Michelle,
>>>
>>> If you just wanted to run one risk-factor predictor at a time, then
>>> the chi-square test will be exactly the same as a multinomial logistic
>>> regression. (Note, that's the chi-square test you find buried under
>>> crosstabs in the SPSS menus, not the chi-square test that you find in
>>> non-parametric tests, which is different).
>>>
>>> If you have multiple predictors, then you need to use multinomial
>>> logistic regression. But you'll need a large sample, and you'll get a
>>> LOT of parameter estimates out, which will take some interpreting.
>>>
>>> If you have a large enough sample, I'd consider doing a principal
>>> components analysis first, to reduce the number of predictors.
>>>
>>> Jeremy
>>>
>>> On 12 October 2010 06:58, Fleur-Michelle Coiffait
>>> <[log in to unmask]> wrote:
>>>> Hi all
>>>>
>>>> I am analysing some data derived from routine information collected via a screening form, which has twelve 'risk factors' on it that get ticked if present (e.g. substance misuse, domestic abuse, etc). I am trying to look at whether the presence of any of these risk factors is related to three 'outcome' categories (no further service involvement, service X involvement, or social work involvement). I had initially thought about doing a Chi square analysis, but came across multinomial logistic regression and wondered if this might be more appropriate? Or are neither of these suitable analyses for this kind of data?
>>>>
>>>> I realise this may be a really silly question, but any advice/clarification would be much appreciated.
>>>>
>>>> Many thanks,
>>>> Fleur-Michelle
>>>> _________________
>>>> Fleur-Michelle Coiffait
>>>> Trainee Clinical Psychologist
>>>> University of Edinburgh & NHS Lothian
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> The University of Edinburgh is a charitable body, registered in
>>>> Scotland, with registration number SC005336.
>>>>
>>>
>>>
>>>
>>> --
>>> Jeremy Miles
>>> Psychology Research Methods Wiki: www.researchmethodsinpsychology.com
>>>
>>
>>
>>
>>
>>
>>
>>
>> --
>> The University of Edinburgh is a charitable body, registered in
>> Scotland, with registration number SC005336.
>>
>>
>
>
>
> --
> Jeremy Miles
> Psychology Research Methods Wiki: www.researchmethodsinpsychology.com
>
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
|