Thanks everyone for your feedback on this. I had some very interesting answers (some of which are below) which really helped clarify the issue for me.
Best wishes,
John
________________________________________
From: Jeremy Miles [[log in to unmask]]
Sent: 31 July 2012 04:49
To: [log in to unmask]
Subject: Re: % & parametric tests?
Yes, bang on.
All parametric tests are a form of regression analysis. And regression
analysis takes the form:
Y = BX + C
Where X is the predictor of interest. If you have percentages, it's
feasible that there are values of X which will lead to impossible
values for Y.
But, this isn't really much of a problem. If it was, we'd never use
parametric statistics ever.
Jeremy
On 28 July 2012 03:03, Ben Haysom-Newport <[log in to unmask]> wrote:
> Hi Steve & All
>
> I think what Jeremy was saying was with true continuous variables you can
> have negative numbers and there isn't usually a cap - however with
> percentages you don't very often get less than 0% - that's the problem.
>
> I'm pretty sure you can use parametric tests on negative numbers as long as
> the data assumptions are met. If not I'm sure someone will correct me!
>
>
> Best Wishes
> Ben Haysom-Newport MSc, BSc (Hons), MBPsS
>
>
>
> On 28 July 2012 10:44, Steve Lukito <[log in to unmask]> wrote:
>>
>> Hi all!
>>
>> I am new to PSYPAG. There seems to be interesting discussion going on here
>> and there. Hope I correctly follow this thread by the hitting reply-all
>> option.
>>
>> I don't see any problem with conducting statistical analysis on data that
>> is turned into percentage. The issue here would be what tyoe of data they
>> were turned into percentage from. So long as they are at least of the
>> interval type then the percentage data will also be of the interval type, so
>> fulfilling the parametric criteria. Maybe I am missing something?
>>
>>
>> "The problem being that you can predict a value greater than 100% or less
>> than 0%, which is probably impossible."
>>
>> I am not sure what you meant here, Jeremy, because we can still conduct
>> parametric statistics on negative data (less than 0%). Can you explain more?
>>
>> Best,
>> Steve
>>
>>
>>
>> On Fri, Jul 27, 2012 at 10:31 PM, Jeremy Miles <[log in to unmask]>
>> wrote:
>>>
>>> Strictly, no.
>>>
>>> The problem being that you can predict a value greater than 100% or
>>> less than 0%, which is probably impossible. (If it's not, the answer
>>> is yes, and ignore the rest of this email).
>>>
>>> However, strictly you can never use a parametric test and we often do.
>>> So you're probably OK.
>>>
>>> If the distribution looks approximately normal, then go ahead and
>>> don't worry about it. If it looks non-normal - in particular if it
>>> appears to be uniformly distributed, then a transformation like the
>>> logistic will work. Turn your variable into a proportion, from zero to
>>> one.
>>>
>>> Then, If x is your variable:
>>> log(x/(1-x)) will convert to something that looks pretty normal. (You
>>> can use the inverse normal probability function, which is a *tiny* bit
>>> better, and 17 times harder to use.
>>>
>>> If you want to see this, paste this code into R:
>>> x <- runif(1000)
>>> y <- log(x/(1-x))
>>> hist(x)
>>> hist(y)
>>>
>>> If you want to get really excessive, then you use a beta regression.
>>> If you want to do that, you're on your own. (Well, you've got Google
>>> on your side).
>>>
>>> Jeremy
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On 27 July 2012 08:25, Barry, John <[log in to unmask]> wrote:
>>> > Dear All,
>>> >
>>> >
>>> >
>>> > I have got an interesting question for you: can you use parametric
>>> > tests on
>>> > data where the dependent variable is measured in percentages? I realise
>>> > that
>>> > the answer should be 'hell no! get out of my classroom!!' but I have a
>>> > nagging feeling that if the scores on the dependent variable are
>>> > normally
>>> > distributed then it shouldn't be considered a form of psychosis if you
>>> > apply
>>> > a t-test or other parametric test.
>>> >
>>> >
>>> >
>>> > Come on you clever people, help me overcome my irrational beliefs!
>>> >
>>> >
>>> >
>>> > Thanks,
>>> >
>>> >
>>> >
>>> > John
>>> >
>>> >
>>> >
>>> >
>>
>>
>>
>>
>> --
>> --------------------------------------------------------------------
>> 117 Melody Road
>> Wandsworth
>> London
>> SW18 2QQ
>> UK
>>
>> Mobile +44 7838397458
>> --------------------------------------------------------------------
>
>
|