Logistic Regression and Risk Ratios

[Cocoababy]

CONTENTS DELETED

The author has deleted this message.

[Maguin, Eugene]

Responses are embedded

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Cocoababy
Sent: Thursday, September 27, 2012 4:51 PM
To: [hidden email]
Subject: Logistic Regression and Risk Ratios

Dear Members,

>> I have a literature review in which
I am looking at two time points of data (cross-sectional data years
2000-2005 and 2007-2012). I am comparing the time points in terms of risk ratios to see if there is an improvement in results between the two time points. Is the risk ratio (confidence intervals) the right statistic or a right choice?

Too little information! Pick one: you have 1) a dichotomous variable measured twice, 2) a continuous variable measured twice, 3) a multicategory ordinal variable measured twice, 4) a multicategory nominal variable measured twice, 5) two different variables measured twice.


>>In another analysis with the same data, I am looking at several possible predictors (one continuous and the other three binary) and one binary outcome variable. I chose to do a logistic regression and I'm not entirely sure if that was the right choice. I will convert the one continous predictor of course so the model will run. But overall, I am concerned that maybe there was a simplier statistics I could have used such as a t-test.

Yes, it is the right choice to do a logistic regression. Why and how are you going to convert the continuous predictor?



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Logistic-Regression-and-Risk-Ratios-tp5715330.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

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

[Cocoababy]

CONTENTS DELETED

The author has deleted this message.

[Rich Ulrich]

I can't tell what you are trying to do.

So far, from what is stated concretely, it looks like you
have a very simple question in this post.  You have a
23-item sum, possibly called Risk, and you want to see if it differs
between two samples.  That could be a t-test.

"Relative risk" is a term that is used for an alternative to
the Odds Ratio.  It is difficult to work with, and it is usually not
a good idea to try.  If you have an Odds Ratio, which is
relevant to logistic regression, what are you predicting
from this computed score?

 - If you have two prediction situations, one for an early sample
and one for a later sample, possibly you to see if the same
prediction equation fits them both. 

If you have a continuous variable, then you use an ordinary
multiple regression; you do not dichotomize it to create a
criterion for logistic regression.

--
Rich Ulrich

> Date: Thu, 27 Sep 2012 16:15:10 -0700

> From: [hidden email]
> Subject: Re: Logistic Regression and Risk Ratios
> To: [hidden email]
>
> Hello,
>
>
> I will probably dichotmize the continuous variable (I have to anyway) for
> the logistic.
>
> For the relative risk ratio I have several binary variables (predictors)
> measured for both time points but different populations. And I am seeing if
> the later population improved in their response. So another words 23
> criteria (all binary 1=Yes 2=No). And I am hoping to get more yes answers in
> the second (later population). I should probably do an asbolute difference
> risk ratio as well (my thinking). What do you think?
>
> ...

[Cocoababy]

CONTENTS DELETED

The author has deleted this message.

[Rich Ulrich]

Sorry.  I still don't get it.

If you have a score which is the count for each person
of Yes-response, where each Yes is some aspect of Risk,
then you have a frequency distribution of Risk scores. 
"Relative Risk" might seem to you to be a nice phrase for
it, but RR is a technical term to statisticians -- It is the ratio
of percent-risk in one group, divided by percent-risk in another
group.  For small percentages, it is approximately equal to the
Odds Ratio.  Sometimes the name is loosely applied to the OR.

You have not yet described the basis for any (technical) RR.

2nd paragraph, which ends in a sentence fragment -- Yes,
you *could* make a dichotomy to represent all-Yes versus Other.
If there are a whole lot of them, that could be sensible.  In that
case, I would probably reverse the scoring to put the data in a more
familiar (to data analysts) form -- mostly zeroes, with higher scores
distributed in what *might*  look something like Poisson.

If there are many non-zero scores in that form, it would then be
worthwhile (probably) to also look at correlations and predictions
for the count of (non)Risk among those for whom that reversed
score was not zero.

--
Rich Ulrich

> Date: Thu, 27 Sep 2012 18:41:12 -0700

> From: [hidden email]
> Subject: Re: Logistic Regression and Risk Ratios
> To: [hidden email]
>
> Hello!
>
> Rich I have 23 attributes or variables that are binary (yes=criteria met
> no=criteria not met). The 23 attributes were collected over two time points.
> I calculated the RR, relative risk ratio, and confidence intervals. And the
> counts are the number of 'yes' values for the two time points. That's what
> the risk is based on. Hope this is clear enough.
>
> The second thing is I am trying to show the odds of getting a particular
> outcome response (1=all 23 are met and 0 = not met) using logistic
> regression. My predictors will be binary a they should.
>
> ...

[Cocoababy]

CONTENTS DELETED

The author has deleted this message.

[Bruce Weaver]

IMO, this *is* a very helpful forum compared to many others.  

The point Rich made quite clearly, I thought, was that you appeared to be using "relative risk" in a very non-standard fashion.  See the Wikipedia page on the RR for examples of how it is typically used.

   http://en.wikipedia.org/wiki/Relative_risk

As you can see there, the RR is typically used in the context of a 2x2 table, not in the context of 23 (or however many) dichotomous variables.  

So, IMO the reason you were not getting the answers you were hoping for was that people were still trying to understand exactly what kind of data you had, and what question(s) you were trying to answer.  If you want clear on-target answers, you have to provide a very clear description of the problem--and *examples* are often very helpful.

HTH.

Cocoababy wrote

Dear Rich,


Never mind. I figured it out on my own. I hope this is the right forum for exchanging ideas in a helpful thoughtful manner. It doesn't sound like it.

[David Marso]

As he hurls his crystal ball across the room, scaring the sh!# out of his dog , destroying the brand new 90" plasma screen with holographic sensurround. ;-)) and alienating the cat who struts away as if nothing unusual had happened.

Bruce Weaver wrote

IMO, this *is* a very helpful forum compared to many others.  

The point Rich made quite clearly, I thought, was that you appeared to be using "relative risk" in a very non-standard fashion.  See the Wikipedia page on the RR for examples of how it is typically used.

   http://en.wikipedia.org/wiki/Relative_risk

As you can see there, the RR is typically used in the context of a 2x2 table, not in the context of 23 (or however many) dichotomous variables.  

So, IMO the reason you were not getting the answers you were hoping for was that people were still trying to understand exactly what kind of data you had, and what question(s) you were trying to answer.  If you want clear on-target answers, you have to provide a very clear description of the problem--and *examples* are often very helpful.

HTH.

Cocoababy wrote

Dear Rich,


Never mind. I figured it out on my own. I hope this is the right forum for exchanging ideas in a helpful thoughtful manner. It doesn't sound like it.

[Rich Ulrich]

I left a strong though implicit question, "What is it that you are
calling Relative Risk?"  -- That was still there, even if the concrete
recommendations were off-target.

Perhaps the advice I offered did not fit the problem; but it did
potentially provide a basis that you could have corrected.

My interpretation here:  You took unwarranted offense at my
first sentence, and therefore you did not pay useful attention to
the rest of my post.

--
Rich Ulrich

> Date: Fri, 28 Sep 2012 13:44:39 -0700

> From: [hidden email]
> Subject: Re: Logistic Regression and Risk Ratios
> To: [hidden email]
>
> Dear Bruce,
>
> This is not a pissing contest. I'm simply posted a preliminary discussion in
> hopes of exploring it further. And I think a helpful post would include
> questions. Not condescending remarks. We are all statisticians on here.
> There is no need to imply otherwise. So I will ask again. Is this the tone
> of the forum? I will be more than happy to spread the word.
>

...

[Swank, Paul R]

I expect that most of us are "teachers". I am. And I tell my graduate students that the only stupid question is the one you have but don't ask. However, we should not treat the questioner as a blithering idiot because they asked a question. It's far worse if they don't ask. I have noted a number of replies on the list that I thought were more curt and condescending than they needed to be. Let's try to be helpful and not just try to show everyone how smart we are. With that, I'll duck and run.

Paul

Paul R. Swank, Ph.D., Professor
Health Promotions and Behavioral Sciences
School of Public Health
University of Texas Health Science Center Houston
________________________________________
From: SPSSX(r) Discussion [[hidden email]] On Behalf Of Bruce Weaver [[hidden email]]
Sent: Friday, September 28, 2012 3:23 PM
To: [hidden email]
Subject: Re: Logistic Regression and Risk Ratios

IMO, this *is* a very helpful forum compared to many others.

The point Rich made quite clearly, I thought, was that you appeared to be
using "relative risk" in a very non-standard fashion.  See the Wikipedia
page on the RR for examples of how it is typically used.

   http://en.wikipedia.org/wiki/Relative_risk

As you can see there, the RR is typically used in the context of a 2x2
table, not in the context of 23 (or however many) dichotomous variables.

So, IMO the reason you were not getting the answers you were hoping for was
that people were still trying to understand exactly what kind of data you
had, and what question(s) you were trying to answer.  If you want clear
on-target answers, you have to provide a very clear description of the
problem--and *examples* are often very helpful.

HTH.



Cocoababy wrote
> Dear Rich,
>
>
> Never mind. I figured it out on my own. I hope this is the right forum for
> exchanging ideas in a helpful thoughtful manner. It doesn't sound like it.





-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/

"When all else fails, RTFM."

NOTE: My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Logistic-Regression-and-Risk-Ratios-tp5715330p5715368.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

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