propensity score for 3 treatments
Locked
propensity score for 3 treatments
 
|
Hi all To control for confounding bias from non-random treatment assignment with 3 different treatments, I would like to calculate a propensity score I later can use in a cox regression model. I know the procedures for a two-treatment approach (logistic regression). But how would I calculate such a score when I have three treatments? Thanks in advance Christian ********************************** |
Re: propensity score for 3 treatments
Administrator
|
Multinomial logistic regression (NOMREG)?
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
Re: propensity score for 3 treatments
|
|
In reply to this post by la volta statistics
I believe you would use a multinomial logistic regression. However, I think it’s important to consider what it is you are trying to accomplish in order to ensure that your model is setup correctly. In a typical
propensity score case of one treatment condition and one control condition, you are trying to ensure that all the cases in the control condition are the same as those in the treatment condition (The probability is equal across groups). The Control condition
cases are thus assigned a likelihood score for being in the treatment, instead of control. Of course, to ensure equality, we do this for all cases, those in treatment as well, as we may have treatment cases that are more similar to or more likely to be in
control, but we frame it around the control units being in treatment. In your case, there isn’t one treatment, but multiple treatment groups. Presumably there is one control condition, and two treatment conditions. In that case, what are you referencing
it too? I would argue that we weren’t really referencing it to the treatment condition in the first place, that was just an easy way to think about the problem. In fact, we were simply creating a probability that individuals were in the group they were supposed
to be in (Realistically we don’t want to be able to predict the group they are in, we want the probabilities to be equal across groups). Since our actual goal is to create probabilities that indicate that everyone was placed as they should (with the goal
being we can’t actually tell who should have been in what group), the referent group doesn’t matter. In your situation, I’d actually pick the control group as my referent group.
I’ll also put this out there. A propensity score is only as good as the variables used to predict the probability. I worked on a problem where the project PI was very excited that his propensity score model
was non-significant with poor ability to place participants. He correctly understood that this indicated an inability to differentiate the participants by these variables, what he failed to realize was that he lacked appropriate instrumental variables for
modeling this. In other words, you really need to be sure the model is solid as well. It’s highly unlikely that the model will be non-significant if done right, you just want a generally poor fit with a solid set of instrumental variables. Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois 510 Devonshire Dr. Champaign, IL 61820 Phone: 217-265-4576 email: [hidden email] From: SPSSX(r) Discussion [mailto:[hidden email]]
On Behalf Of la volta statistics Hi all To control for confounding bias from non-random treatment assignment with 3 different treatments, I would like to calculate a propensity score I later can use in a cox regression
model. I know the procedures for a two-treatment approach (logistic regression). But how would I calculate such a score when I have three treatments? Thanks in advance Christian
**********************************
|
Re: propensity score for 3 treatments
|
|
In reply to this post by Bruce Weaver
I first thought of an ordinal regression but then realized that there is no assurance that the parallel lines assumption would be met and then switched to multinomial regression. Maybe this is not a problem but the usual propensity setup has matching on the computed logit or log odds value from a logistic regression with a dichotomous DV. But with a three category DV you get two computed logits or log odds values irrespective of whether you use ordinal or multinomial. So how is the matching to be done.
You've got to figure that this problem has come up before and an accepted method has been developed; however, I don't know what it is. My suggestion is to average the logits and match on that (but check the resulting match to see how well the match is on each logit value). My rationale is that if you formed matched triplets and assigned them to condition then did the multinomial regression analysis, each member of the triplet would have the same average logit and logit components. Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver Sent: Friday, November 30, 2012 9:39 AM To: [hidden email] Subject: Re: propensity score for 3 treatments Multinomial logistic regression (NOMREG)? la volta statistics wrote > Hi all > > To control for confounding bias from non-random treatment assignment > with > 3 > different treatments, I would like to calculate a propensity score I > later can use in a cox regression model. I know the procedures for a > two-treatment approach (logistic regression). But how would I > calculate such a score when I have three treatments? > > > > Thanks in advance > > Christian > > > > > > ********************************** > la volta statistics > Christian Schmidhauser, Dr.phil.II > Weinbergstrasse 108 > CH-8006 Zürich > Tel: +41 (043) 233 98 01 > Fax: +41 (043) 233 98 02 > email: <mailto: > schmidhauser@ > > mailto: > schmidhauser@ > > Web: <http://www.lavolta.ch> www.lavolta.ch ----- -- 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/propensity-score-for-3-treatments-tp5716540p5716544.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 |
Re: propensity score for 3 treatments
|
|
Thanks Matthew, Eugene, and Bruce
I thought of NOMREG as well but then you get a probability for each category. So how would that be used as a score? Christian On Fri, 30 Nov 2012 10:11:28 -0500 "Maguin, Eugene" <[hidden email]> wrote: > I first thought of an ordinal regression but then realized that >there is no assurance that the parallel lines assumption would be met >and then switched to multinomial regression. Maybe this is not a >problem but the usual propensity setup has matching on the computed >logit or log odds value from a logistic regression with a dichotomous >DV. But with a three category DV you get two computed logits or log >odds values irrespective of whether you use ordinal or multinomial. >So how is the matching to be done. > > You've got to figure that this problem has come up before and an >accepted method has been developed; however, I don't know what it is. >My suggestion is to average the logits and match on that (but check >the resulting match to see how well the match is on each logit >value). My rationale is that if you formed matched triplets and >assigned them to condition then did the multinomial regression >analysis, each member of the triplet would have the same average >logit and logit components. > > Gene Maguin > > -----Original Message----- >From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf >Of Bruce Weaver > Sent: Friday, November 30, 2012 9:39 AM > To: [hidden email] > Subject: Re: propensity score for 3 treatments > > Multinomial logistic regression (NOMREG)? > > > > la volta statistics wrote >> Hi all >> >> To control for confounding bias from non-random treatment assignment >> with >> 3 >> different treatments, I would like to calculate a propensity score I >> later can use in a cox regression model. I know the procedures for a >> two-treatment approach (logistic regression). But how would I >> calculate such a score when I have three treatments? >> >> >> >> Thanks in advance >> >> Christian >> >> >> >> >> >> ********************************** >> la volta statistics >> Christian Schmidhauser, Dr.phil.II >> Weinbergstrasse 108 >> CH-8006 Zürich >> Tel: +41 (043) 233 98 01 >> Fax: +41 (043) 233 98 02 >> email: <mailto: > >> schmidhauser@ > >> > mailto: > >> schmidhauser@ > >> >> Web: <http://www.lavolta.ch> www.lavolta.ch > > > > > > ----- > -- > 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/propensity-score-for-3-treatments-tp5716540p5716544.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 ===================== 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 |
|
|
In reply to this post by Maguin, Eugene
Donald Rubin at Columbia reviews propsensity scoring for two and three
groups in this article: http://www.stat.columbia.edu/~gelman/stuff_for_blog/propensity.html Quoting from the article: |With more than two treatment conditions, the propensity score usually |differs for each pair of treatment groups being compared (that is, with |three treatment groups labelled A, B, and C, there are three propensity |scores: A compared with B, A compared with C, and B compared with C). |At first, this may seem to be a limitation of propensity score technology |relative to a model-based analysis, but in fact it is an important strength |and points to further weaknesses in a model-based approach. We show |this by exploring a range of hypothetical modifications to Cochran's [2] |smoking example. Rubin, D. B. (1997). Estimating causal effects from large data sets using propensity scores. Annals of internal medicine, 127, 757-763. Given that this is a 1997 publication, it is reasonable to assumed that there are more recent articles that might be relevant, such as the following: http://europepmc.org/articles/PMC3105164/pdf/nihms276566.pdf -Mike Palij New York University [hidden email] On Fri, Nov 30, 2012 at 10:11 AM, Maguin, Eugene <[hidden email]> wrote: > I first thought of an ordinal regression but then realized that there is no assurance that the parallel lines assumption would be met and then switched to multinomial regression. Maybe this is not a problem but the usual propensity setup has matching on the computed logit or log odds value from a logistic regression with a dichotomous DV. But with a three category DV you get two computed logits or log odds values irrespective of whether you use ordinal or multinomial. So how is the matching to be done. > > You've got to figure that this problem has come up before and an accepted method has been developed; however, I don't know what it is. My suggestion is to average the logits and match on that (but check the resulting match to see how well the match is on each logit value). My rationale is that if you formed matched triplets and assigned them to condition then did the multinomial regression analysis, each member of the triplet would have the same average logit and logit components. > > Gene Maguin > > -----Original Message----- > From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver > Sent: Friday, November 30, 2012 9:39 AM > To: [hidden email] > Subject: Re: propensity score for 3 treatments > > Multinomial logistic regression (NOMREG)? > > > > la volta statistics wrote >> Hi all >> >> To control for confounding bias from non-random treatment assignment >> with >> 3 >> different treatments, I would like to calculate a propensity score I >> later can use in a cox regression model. I know the procedures for a >> two-treatment approach (logistic regression). But how would I >> calculate such a score when I have three treatments? >> >> >> >> Thanks in advance >> >> Christian >> >> >> >> >> >> ********************************** >> la volta statistics >> Christian Schmidhauser, Dr.phil.II >> Weinbergstrasse 108 >> CH-8006 Zürich >> Tel: +41 (043) 233 98 01 >> Fax: +41 (043) 233 98 02 >> email: <mailto: > >> schmidhauser@ > >> > mailto: > >> schmidhauser@ > >> >> Web: <http://www.lavolta.ch> www.lavolta.ch > > > > > > ----- > -- > 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/propensity-score-for-3-treatments-tp5716540p5716544.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 ===================== 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 |
| Free forum by Nabble | Edit this page |