Need advice on sensitivity and specificity analysis

[stace swayne]

Dear List,

I am interested in running a sensitivity and specificity analysis and need some advice on setting up my data:

I have a sample of 1043 older adults, a total test score on a 15 item Geriatric Depression Scale, and various demographic variables (this is all raw data).

The depression scale is defined as follows: 0-4=no depression, 5-15=depression.

Can someone lend some advice on how I should proceed in order to run this analysis.

All suggestions are welcomed.

S

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

[jmdpulido]

I think the best approach will be some form of logistic regression.

Depending how your data distributes, collapsing the scale into 2 categories (No depression = 0; Depression=1) and running an ordinary logistic regression might be a good option. Of course, there you loose information about how severe the depression is, and your model will only predicts wheter someone is in positions 1-4 or 5-15.

Multinomial and ordinal logistic regression will allow you to predict the probability of one person to get a particular point in the depression scale. However, they are more complex.

In order to choose between ordinal and multinomial you should run the paralell lines test. The hypothesis of ordinal logistic regression is the "proportional hazards". It means that any of your independent variables have the same effect in the probability of passing from point 1 to point 2 on the scale, that of passing from point 2 to point 3, etc... SPSS gives you the results of the test, so it is easy to choose. Of course, if you could assume proportional hazards, ordinal will be much better, as you loose less degrees of freedom (as beta coefficients are restricted to be equal for all predicted categories of your dependent variable). Multinomial logistic will need 14 different equations to estimate, which can be a bit too much.

There is always the option to collapse your 15 point scale in 3 or 4 categories. Of course, you will need to read the theory behind this scale and also to inspect the distribution of your data.

On the other hand, as a preliminary step, just to get a "flavour" of what is happening, you could first run a linear regression by OLS, so you get to inspect the coefficients, and then go to the logistic regression.

You could ask SPSS to save the predicted probabilities and predicted groups, so you can easily calculate the sensitivity, the specificity, the positive predictive value and all other measures that compare the predictions of your model with the observed data.

If you need any further help, please do not hesitate to ask.

J. Pulido.
PhD Student.

[Rolf Pfister-2]

Guten Tag

ich bin bis am 15. August 2011 nicht im Büro und lese die Emails erst, wenn ich zurück bin. In dringenden Fällen wenden Sie sich bitte an [hidden email] oder an den Support ([hidden email]).

Vielen Dank für Ihr Verständnis

 

Hello,

I'm not in the office until 15 August 2011 and I will read my emails only when I'm back. For urgent cases, please contact [hidden email] or our support ([hidden email]).

Thank you for your comprehension.

 

[John F Hall]

If you're new to SPSS go to the tutorials on my website.

I take it the items are scored 0,1 and summed to generate a GDS 0 - 15?

How are your data stored?  Are they in SPSS already, in Excel or do you
still have to enter them?

You'll need one line of syntax:

recode GDS ( 0 thru 4 =1) (5 thru 15 = 2) (else = sysmis) into gdsgroup .

You can do quite a lot with simple tabulation, but without seeing your
actual instruments and data I can only suggest a schema:

crosstabs <demographic varlist> by gdsgroup /cel row.


John F Hall

[hidden email]
www.surveyresearch.weebly.com




-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
stace swayne
Sent: 25 July 2011 05:56
To: [hidden email]
Subject: Need advice on sensitivity and specificity analysis

Dear List,

I am interested in running a sensitivity and specificity analysis and need
some advice on setting up my data:

I have a sample of 1043 older adults, a total test score on a 15 item
Geriatric Depression Scale, and various demographic variables (this is all
raw data).

The depression scale is defined as follows: 0-4=no depression,
5-15=depression.

Can someone lend some advice on how I should proceed in order to run this
analysis.

All suggestions are welcomed.

S

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

[jmdpulido]

Thank you so much for your response,

to follow-up, I have a geriatric depression instrument which is made up of
15 with a yes/no answer format. I have created a total score variable, and a
variable that reflects people without depression who scored (0-4) and people
with depression who scored 5-15.

I also have various demographic variables (gender, age etc.)

I am interested in finding out how well the instrument is as classifying
people as depressed vs. non-depressed. I thought that a sensitivity &
specificity analysis would be appropriate for what I was trying to do.

My logic, was that I could run a cross-tab between gender and the
depression/non-depression variable and based on those results I would
calculate the sensitivity and specificity by hand.

Do you think this is a reasonable approach?

All suggestions are welcomed,

S

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Need-advice-on-sensitivity-and-specificity-analysis-tp4629536p4631679.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

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

[jmdpulido]

Dear Stace,

Contingency tables are always a reasonable approach. I am an economist. So I don't know anything about your scale. Nevertheless, it sounds fine to me to compute the score and divide people in two groups (depressed, not depressed).

Try first contingency tables and use basic statistics (chi-square, phi, lambda, etc) to find if there is a relationship.

Of course, you can calculate sensitivity and specifity by hand from a contingency table.

Then, when you have finished this analysis you might want to run a multivariate analysis, so you can check the effect of one variable (e.g. age) letting the other variables constant (e.g. gender). That's what we call "ceteris paribus". For that you need a logistic regression.

Never forget to try interaction terms in your logistic model (e.g. age*gender), so you can find out if the effect of one variable interacts with other (e.g. if the effect of age is the same for men and women or if it is different).

SPSS command for contingency tables is crosstab:

CROSSTABS
  /TABLES=Depend  BY Gender
  /FORMAT= AVALUE TABLES
  /STATISTIC=CHISQ CC PHI LAMBDA
  /CELLS= COUNT
  /COUNT ROUND CELL .

SPSS command for logistic regression is logistic:

LOGISTIC REGRESSION VARIABLES  Depend
  /METHOD = ENTERGender Sex
  /CONTRAST (Sexo)=Indicator(1)  
  /SAVE = PRED PGROUP RESID
  /CLASSPLOT
  /PRINT = CI(95)
  /CRITERIA = PIN(.05) POUT(.10) ITERATE(20) CUT(.5) .

SPSS will save the predicted group as a new variable PGroup, so you can run another crosstable with your observed depression variable and your predicted depression variable to calculate specificity and sensitivity.

Kind Regards

J. Pulido
PhD Student

[Rich Ulrich]

A crosstab between gender and depression will show you whether
gender is associated with depression.  It does not seems very
likely to me that you are really seeking to know the Sensitivity
and Specificity of Gender -- How good is Gender at predicting
(scaled) Depression?  You do say that you are interested in how
well the *depression* scale classifies depression.

What you need for the ideal measure of Sensitivity and Specificity
is a "gold standard" criterion of depression.  You cross-tabulate
your scaled score against the gold standard, and you obtain
measures of sensitivity and specificity directly.  

For s-and-s,  it is absolutely true that you need *some*  second
measure of depression, so you can see how much they disagree.
To the extent that your second measure shares "error" with the
scale, you can only place a lower limit on the un-reliability.

If you don't have a second measure of depression, you surely have
to aim at some other demonstration of reliability or validity.

For instance, Cronbach's alpha is a measure of the internal
consistency/reliability of a scale. 

"Predictive validity"  is how well a scale predicts something else
that the dimension is supposed to predict.


--
Rich Ulrich

> Date: Mon, 25 Jul 2011 10:35:38 -0700

> From: [hidden email]
> Subject: Re: Need advice on sensitivity and specificity analysis
> To: [hidden email]
>
> Thank you so much for your response,
>
> to follow-up, I have a geriatric depression instrument which is made up of
> 15 with a yes/no answer format. I have created a total score variable, and a
> variable that reflects people without depression who scored (0-4) and people
> with depression who scored 5-15.
>
> I also have various demographic variables (gender, age etc.)
>
> I am interested in finding out how well the instrument is as classifying
> people as depressed vs. non-depressed. I thought that a sensitivity &
> specificity analysis would be appropriate for what I was trying to do.
>
> My logic, was that I could run a cross-tab between gender and the
> depression/non-depression variable and based on those results I would
> calculate the sensitivity and specificity by hand.
>
> Do you think this is a reasonable approach?
>
> All suggestions are welcomed,
>
> S

[Beckstead, Jason]

I will be away on vaction July 26 through August 21.