bootstrapping difference between means

[bgreen]

Hello,

I have data derived for 63 participants. The data is not normally
distributed and when I compare the medians between groups, the value
is usually 0.  I gather data sets where there are frequent 0's is not
an uncommon statistical problems. It was suggested to me that
bootstrapping may be a solution to this problem. There are three
groups of interest (n=10, 23, 30) which I want to compare on
variables such as prior hospitalisation and imprisonment.

I found a macro by Marta Gracia-Granero which estimates a sample
mean.
http://spssx-discussion.1045642.n5.nabble.com/Bootstrapping-in-SPSS-td1071984.html

Is it possible to use bootstrapping to test the difference between
three sample means or medians, and if so how can I do this in SPSS?
My crude solution had been to omit values of 0 and just compare
individuals with values above 0, but appreciate this is problematic.

Any assistance is appreciated.

Bob

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Bruce Weaver]

I don't have access to SPSS right now, but I think something like the following will work (assuming your license includes the BOOTSTRAP module).

SET RNG=MT.
BOOTSTRAP.
ONEWAY Y BY group.

bgreen wrote

Hello,

I have data derived for 63 participants. The data is not normally
distributed and when I compare the medians between groups, the value
is usually 0.  I gather data sets where there are frequent 0's is not
an uncommon statistical problems. It was suggested to me that
bootstrapping may be a solution to this problem. There are three
groups of interest (n=10, 23, 30) which I want to compare on
variables such as prior hospitalisation and imprisonment.

I found a macro by Marta Gracia-Granero which estimates a sample
mean.
http://spssx-discussion.1045642.n5.nabble.com/Bootstrapping-in-SPSS-td1071984.html

Is it possible to use bootstrapping to test the difference between
three sample means or medians, and if so how can I do this in SPSS?
My crude solution had been to omit values of 0 and just compare
individuals with values above 0, but appreciate this is problematic.

Any assistance is appreciated.

Bob

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Rich Ulrich]

Bootstrapping certainly will not solve the problems you face:
1) unclear hypotheses;
2) tiny Ns for making any sort of meaningful comparisons when you
do not expect huge differences among three groups.

One way to make sense about data with too many 0's is to make
two statements:  Compare the fractions of zeros; and compare the
means among the non-zero subjects.  Comparing the number of zeros
is usually meaningful.  When your small-N data may have an extreme
number or two, comparing the means might not seem to give much
sense even after you strip out the zeros.

You have to consider the narrative that you can produce.  What makes
sense?  - If they are similar (or different), what sort of graphic or presentation
of numbers show what is going on?  With the size of the Ns, you should start
out *assuming* that important differences will not necessarily show up on
testing; anything you get from a test will be a bonus. 

 - When you have data that the means do not look particularly meaningful,
which seems to be the case here, I do not think that bootstrapping the
means will give a good statement, but it is not something that I have tried.

--
Rich Ulrich

> Date: Wed, 22 Oct 2014 07:10:39 +1000

> From: [hidden email]
> Subject: bootstrapping difference between means
> To: [hidden email]
>
> Hello,
>
> I have data derived for 63 participants. The data is not normally
> distributed and when I compare the medians between groups, the value
> is usually 0. I gather data sets where there are frequent 0's is not
> an uncommon statistical problems. It was suggested to me that
> bootstrapping may be a solution to this problem. There are three
> groups of interest (n=10, 23, 30) which I want to compare on
> variables such as prior hospitalisation and imprisonment.
>
> I found a macro by Marta Gracia-Granero which estimates a sample
> mean.
> http://spssx-discussion.1045642.n5.nabble.com/Bootstrapping-in-SPSS-td1071984.html
>
> Is it possible to use bootstrapping to test the difference between
> three sample means or medians, and if so how can I do this in SPSS?
> My crude solution had been to omit values of 0 and just compare
> individuals with values above 0, but appreciate this is problematic.
>
> Any assistance is appreciated.
>

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

[bgreen]

Hello Rich,

Thanks. The overall hypothesis was that the group who received the
most support would have a lower re-incarceration rate and the purpose
of the tests were simply to examine whether the groups differed in
any significant way (e.g in terms of age, sex, previous
hospitalisation days and frequency & previous imprisonment days and
number etc).

After removing the zeros I ran non-parametric analyses due to the
non-normal distribution of the data. I hadn't considered examining
fractions of zeroes, though did dichotomize some variables.

Zero-inflated Negative Binomial Regression was another option that
had been suggested.

Thanks again

Bob



At 02:56 PM 22/10/2014, Rich Ulrich wrote:

>Bootstrapping certainly will not solve the problems you face:
>1) unclear hypotheses;
>2) tiny Ns for making any sort of meaningful comparisons when you
>do not expect huge differences among three groups.
>
>One way to make sense about data with too many 0's is to make
>two statements:  Compare the fractions of zeros; and compare the
>means among the non-zero subjects.  Comparing the number of zeros
>is usually meaningful.  When your small-N data may have an extreme
>number or two, comparing the means might not seem to give much
>sense even after you strip out the zeros.
>
>You have to consider the narrative that you can produce.  What makes
>sense?  - If they are similar (or different), what sort of graphic
>or presentation
>of numbers show what is going on?  With the size of the Ns, you should start
>out *assuming* that important differences will not necessarily show up on
>testing; anything you get from a test will be a bonus.
>
>  - When you have data that the means do not look particularly meaningful,
>which seems to be the case here, I do not think that bootstrapping the
>means will give a good statement, but it is not something that I have tried.
>
>--
>Rich Ulrich
>
> > Date: Wed, 22 Oct 2014 07:10:39 +1000
> > From: [hidden email]
> > Subject: bootstrapping difference between means
> > To: [hidden email]
> >
> > Hello,
> >
> > I have data derived for 63 participants. The data is not normally
> > distributed and when I compare the medians between groups, the value
> > is usually 0. I gather data sets where there are frequent 0's is not
> > an uncommon statistical problems. It was suggested to me that
> > bootstrapping may be a solution to this problem. There are three
> > groups of interest (n=10, 23, 30) which I want to compare on
> > variables such as prior hospitalisation and imprisonment.
> >
> > I found a macro by Marta Gracia-Granero which estimates a sample
> > mean.
> >
> http://spssx-discussion.1045642.n5.nabble.com/Bootstrapping-in-SPSS-td1071984.html
> >
> > Is it possible to use bootstrapping to test the difference between
> > three sample means or medians, and if so how can I do this in SPSS?
> > My crude solution had been to omit values of 0 and just compare
> > individuals with values above 0, but appreciate this is problematic.
> >
> > Any assistance is appreciated.
> >

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD