statistical test for 2x8 table with small expected values

[J McClure]

Hi,
 From a survey I have three clinical characteristics  (depression, PTSD,
and alcohol misuse) which are either present (yes) or not present (no).
I also have the answers to multiple questions about suicidal thoughts
and behaviors which I have collapsed into a dichotomous variable of yes
or no.
I want to see if the proportion of respondents answering positively to
each of the questions differs by whether they have one or more of the
clinical characteristics.
So, I created a variable of clinical characteristics (Clin_Char) with 8
categories which represent all possible combinations including none and
all three characteristics.
The problem is that with an n of 482, the numbers are small for three of
the categories and when I do a chi-squared test there are cells with an
inadequate expected value.
Collapsing categories or dropping some categories would loose
information but it seems since the contingency table is bigger than 2x2
I can't use Yates correction Fisher's exact test.
Suggestions?
Thanks,
Jan

=====================
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

[Art Kendall]

It appears that the METHOD=EXACT specification in CROSSTABS might be
what you need.
The simulation in the syntax below ran in a few seconds.
It assumes depression occurs in 10% of cases, PTSD in 4 %, depression in
10%. You can change the RV arguments to be more like your actual situation.

Open a new instance of SPSS. Paste the syntax in a syntax window. Run it.
Is this what you need?

Art Kendall
Social Research Consultants

new file.
set seed 20110110.
input program.
loop id = 1 to 482.
compute depression  = RV.BERNOULLI(.10).
compute PTSD  = RV.BERNOULLI(.04).
compute alcohol  = RV.BERNOULLI(.10).
end case.
end loop.
end file.
end input program.
formats id (f3) depression to alcohol (f1).
value labels  depression to alcohol 0 'no' 1 'yes' .
CROSSTABS tables = depression by PTSD by alcohol
  /cells = all
  /method =exact.


On 1/10/2011 2:06 PM, J McClure wrote:

> Hi,
> From a survey I have three clinical characteristics  (depression, PTSD,
> and alcohol misuse) which are either present (yes) or not present (no).
> I also have the answers to multiple questions about suicidal thoughts
> and behaviors which I have collapsed into a dichotomous variable of yes
> or no.
> I want to see if the proportion of respondents answering positively to
> each of the questions differs by whether they have one or more of the
> clinical characteristics.
> So, I created a variable of clinical characteristics (Clin_Char) with 8
> categories which represent all possible combinations including none and
> all three characteristics.
> The problem is that with an n of 482, the numbers are small for three of
> the categories and when I do a chi-squared test there are cells with an
> inadequate expected value.
> Collapsing categories or dropping some categories would loose
> information but it seems since the contingency table is bigger than 2x2
> I can't use Yates correction Fisher's exact test.
> Suggestions?
> Thanks,
> Jan
>
> =====================
> 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
>

=====================
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

[Art Kendall]

oops it would be 4 way like this.

new file.
set seed 20110110.
input program.
loop id = 1 to 482.
compute depression  = RV.BERNOULLI(.10).
compute PTSD  = RV.BERNOULLI(.04).
compute alcohol  = RV.BERNOULLI(.10).
compute suicidal =  RV.BERNOULLI(.10).
end case.
end loop.
end file.
end input program.
formats id (f3) depression to alcohol (f1).
value labels  depression to suicidal 0 'no' 1 'yes' .
CROSSTABS tables = depression by suicidal  by PTSD by alcohol
  /cells = all
  /method =exact.


Art

On 1/10/2011 3:59 PM, Art Kendall wrote:

> It appears that the METHOD=EXACT specification in CROSSTABS might be
> what you need.
> The simulation in the syntax below ran in a few seconds.
> It assumes depression occurs in 10% of cases, PTSD in 4 %, depression
> in 10%. You can change the RV arguments to be more like your actual
> situation.
>
> Open a new instance of SPSS. Paste the syntax in a syntax window. Run it.
> Is this what you need?
>
> Art Kendall
> Social Research Consultants
>
> new file.
> set seed 20110110.
> input program.
> loop id = 1 to 482.
> compute depression  = RV.BERNOULLI(.10).
> compute PTSD  = RV.BERNOULLI(.04).
> compute alcohol  = RV.BERNOULLI(.10).
> end case.
> end loop.
> end file.
> end input program.
> formats id (f3) depression to alcohol (f1).
> value labels  depression to alcohol 0 'no' 1 'yes' .
> CROSSTABS tables = depression by PTSD by alcohol
>  /cells = all
>  /method =exact.
>
>
> On 1/10/2011 2:06 PM, J McClure wrote:
>> Hi,
>> From a survey I have three clinical characteristics  (depression, PTSD,
>> and alcohol misuse) which are either present (yes) or not present (no).
>> I also have the answers to multiple questions about suicidal thoughts
>> and behaviors which I have collapsed into a dichotomous variable of yes
>> or no.
>> I want to see if the proportion of respondents answering positively to
>> each of the questions differs by whether they have one or more of the
>> clinical characteristics.
>> So, I created a variable of clinical characteristics (Clin_Char) with 8
>> categories which represent all possible combinations including none and
>> all three characteristics.
>> The problem is that with an n of 482, the numbers are small for three of
>> the categories and when I do a chi-squared test there are cells with an
>> inadequate expected value.
>> Collapsing categories or dropping some categories would loose
>> information but it seems since the contingency table is bigger than 2x2
>> I can't use Yates correction Fisher's exact test.
>> Suggestions?
>> Thanks,
>> Jan
>>
>> =====================
>> 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
>>

=====================
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

[J McClure]

Hi Art,
Thank you very much for your input!
I don't have the module for exact tests but am happy to purchase it.
Before doing so, a question.
I have eight categories of depression, ptsd, and alcohol misuse (i.e.
none, depression only, ptsd only, alcohol misuse only, depression and
PTSD, depression and alcohol misuse....all three).
So, would I have an eight way table?
Thanks,
Jan



On 1/10/2011 1:07 PM, Art Kendall wrote:

> oops it would be 4 way like this.
>
> new file.
> set seed 20110110.
> input program.
> loop id = 1 to 482.
> compute depression  = RV.BERNOULLI(.10).
> compute PTSD  = RV.BERNOULLI(.04).
> compute alcohol  = RV.BERNOULLI(.10).
> compute suicidal =  RV.BERNOULLI(.10).
> end case.
> end loop.
> end file.
> end input program.
> formats id (f3) depression to alcohol (f1).
> value labels  depression to suicidal 0 'no' 1 'yes' .
> CROSSTABS tables = depression by suicidal  by PTSD by alcohol
>  /cells = all
>  /method =exact.
>
>
> Art
>
> On 1/10/2011 3:59 PM, Art Kendall wrote:
>> It appears that the METHOD=EXACT specification in CROSSTABS might be
>> what you need.
>> The simulation in the syntax below ran in a few seconds.
>> It assumes depression occurs in 10% of cases, PTSD in 4 %, depression
>> in 10%. You can change the RV arguments to be more like your actual
>> situation.
>>
>> Open a new instance of SPSS. Paste the syntax in a syntax window. Run
>> it.
>> Is this what you need?
>>
>> Art Kendall
>> Social Research Consultants
>>
>> new file.
>> set seed 20110110.
>> input program.
>> loop id = 1 to 482.
>> compute depression  = RV.BERNOULLI(.10).
>> compute PTSD  = RV.BERNOULLI(.04).
>> compute alcohol  = RV.BERNOULLI(.10).
>> end case.
>> end loop.
>> end file.
>> end input program.
>> formats id (f3) depression to alcohol (f1).
>> value labels  depression to alcohol 0 'no' 1 'yes' .
>> CROSSTABS tables = depression by PTSD by alcohol
>>  /cells = all
>>  /method =exact.
>>
>>
>> On 1/10/2011 2:06 PM, J McClure wrote:
>>> Hi,
>>> From a survey I have three clinical characteristics  (depression, PTSD,
>>> and alcohol misuse) which are either present (yes) or not present (no).
>>> I also have the answers to multiple questions about suicidal thoughts
>>> and behaviors which I have collapsed into a dichotomous variable of yes
>>> or no.
>>> I want to see if the proportion of respondents answering positively to
>>> each of the questions differs by whether they have one or more of the
>>> clinical characteristics.
>>> So, I created a variable of clinical characteristics (Clin_Char) with 8
>>> categories which represent all possible combinations including none and
>>> all three characteristics.
>>> The problem is that with an n of 482, the numbers are small for
>>> three of
>>> the categories and when I do a chi-squared test there are cells with an
>>> inadequate expected value.
>>> Collapsing categories or dropping some categories would loose
>>> information but it seems since the contingency table is bigger than 2x2
>>> I can't use Yates correction Fisher's exact test.
>>> Suggestions?
>>> Thanks,
>>> Jan
>>>
>>> =====================
>>> 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
>>>
>

=====================
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

[J McClure]

Yes, thanks Art. I guess I had a brain fart (again).
Thanks to you and the others on the listserv for providing such a valuable service.
Jan

If I understand correctly what you have, the cells and margins of the 4 way crosstab should have all that you need.
Here is the output from the simulation.  Does it have what you need?

On 1/10/2011 1:07 PM, Art Kendall wrote:

oops it would be 4 way like this.

new file.
set seed 20110110.
input program.
loop id = 1 to 482.
compute depression  = RV.BERNOULLI(.10).
compute PTSD  = RV.BERNOULLI(.04).
compute alcohol  = RV.BERNOULLI(.10).
compute suicidal =  RV.BERNOULLI(.10).
end case.
end loop.
end file.
end input program.
formats id (f3) depression to alcohol (f1).
value labels  depression to suicidal 0 'no' 1 'yes' .
CROSSTABS tables = depression by suicidal  by PTSD by alcohol
 /cells = all
 /method =exact.


Art

On 1/10/2011 3:59 PM, Art Kendall wrote:

It appears that the METHOD=EXACT specification in CROSSTABS might be what you need.
The simulation in the syntax below ran in a few seconds.
It assumes depression occurs in 10% of cases, PTSD in 4 %, depression in 10%. You can change the RV arguments to be more like your actual situation.

Open a new instance of SPSS. Paste the syntax in a syntax window. Run it.
Is this what you need?

Art Kendall
Social Research Consultants

new file.
set seed 20110110.
input program.
loop id = 1 to 482.
compute depression  = RV.BERNOULLI(.10).
compute PTSD  = RV.BERNOULLI(.04).
compute alcohol  = RV.BERNOULLI(.10).
end case.
end loop.
end file.
end input program.
formats id (f3) depression to alcohol (f1).
value labels  depression to alcohol 0 'no' 1 'yes' .
CROSSTABS tables = depression by PTSD by alcohol
 /cells = all
 /method =exact.


On 1/10/2011 2:06 PM, J McClure wrote:

Hi,
From a survey I have three clinical characteristics  (depression, PTSD,
and alcohol misuse) which are either present (yes) or not present (no).
I also have the answers to multiple questions about suicidal thoughts
and behaviors which I have collapsed into a dichotomous variable of yes
or no.
I want to see if the proportion of respondents answering positively to
each of the questions differs by whether they have one or more of the
clinical characteristics.
So, I created a variable of clinical characteristics (Clin_Char) with 8
categories which represent all possible combinations including none and
all three characteristics.
The problem is that with an n of 482, the numbers are small for three of
the categories and when I do a chi-squared test there are cells with an
inadequate expected value.
Collapsing categories or dropping some categories would loose
information but it seems since the contingency table is bigger than 2x2
I can't use Yates correction Fisher's exact test.
Suggestions?
Thanks,
Jan

=====================
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

===================== 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