My use of propensity-score methods recently suggests to me that you'll only be able to generate a propensity score for one treatment compared to all others (1 vs. 3), rather than (1 vs. 1 vs. 1 vs. 1). I'd be very interested to see if others can quote from any literature about ordinal approaches, because I'm not aware of any.

There are several approaches to deal with multivalued treatments, but matching is actually the least desirable approach. In Stata 13 a treatment effects suite was introduced which includes several propensity score weighting approaches (as well as regression adjustment). In addition, I wrote a program called mmws (downloadable from SSC), which allows you to perform treatment effect estimates for multivalued treatments using a couple different weighting schemes.

Thanks Ariel for your help. I will download the command . However, I am not sure which treatment the propensity score will be predicted  for when you have more than two treatment

Look at the examples I give in the help file. There are examples for binary, ordinal, and nominal treatments.

Many thanks

Ariel, I looked on you example  for the nominal treatment. Is it necessary to go through all the steps of weighting and stratification, is it enough to use the propensity scores that was generated from the  mlogit model to use them directly in a regression model. 

Moreover, the four treatments that I have differ in the sample size, so would the propensity score be biased regarding that.

No, you cannot use propensity scores directly in the regression for multivalued treatments. -mmws- does all the hard work for you, so I am not sure why you would not operationalize the model as intended? You can use the IPTW option in -mmws- to get the IPT weights for the model as well. No, sample sizes do not need to be identical in each treatment group.

See the link to the paper I wrote on the topic. This should answer your questions more thoroughly...

Ariel

Article Combining propensity score-based stratification and weightin...

Article Using Propensity Score-Based Weighting in the Evaluation of ...

Hi Ariel:

Many thanks for your papers.I read them and I started to use the commands that you built in STATA. However, I faced the following challenges during running the commands for the nominal treatments.

First,  In the first published paper that you attached, you mentioned that it is better to stratify the propensity scores into five quantiles. However, in the example you provided in the STATA help, you used different number of quantiles i.e. 5, 8,10. In my case, I have four treatments with the following numbers 195, 35, 8,and 5, respectively . How many quantiles do I have to stratify them for ?

Second, Why did you use IPTW and re stratify the strata using nominal strata. Is it enough to use only marginal weight  mmws ed, pscore(ps1 ps2 ps3 ps4) nominal strata( strataA strataB strataC strataD).

Third, can I use 'predict' instead of 'margins' after I regress my outcome.  I tried to use margins, However, It gave me an error  'factor is not found.

Fourth, Based in your experience ,do you think is it enough if I used regression method from beginning instead of  propensity score. I get confused with both terms because both approaches may give similar results. 

Sorry for this lengthily email, but I am a little bit worried that I may misunderstood the concept of conducting the propensity score 

You have a serious problem in that you have two treatment groups with extremely small samples. That will pose a problem for causal inference regardless of approach.

As for the number of strata, I typically start at 5 and then go up incrementally to ensure that each cell is still populated. The more the better. But as a general rule, 5 strata are sufficient. I gave several different strata sizes in the example for the sole purpose of demonstration.

The IPTW option does not require stratification, and it will calculate the weights regardless of what you do with the MMWS options.

I am not sure what you'd do with predict? That doesn't give you the contrasts that you need between treatments. As for the error message, you'll need to provide me with the syntax you used that gave you that error.

Whether you use regression or a propensity score model is up to you. If the regression model is correctly identified and there is no extrapolation occurring between groups on their covariates, then regression may be more efficient.

However, I am very concerned about the small sample sizes of your two last groups. This will be a problem regardless of adjustment approach.    

Hi Ariel:

Thanks for your reply 

Is it enough to do MMWS instead of IPTW. I get confused why do I have to do both approaches.

What I meant by predict is that I would like to predict the outcome considering the propensity score weight for each patients. Margins does not allow to do that  

Attached is the syntax that I used . Also I attached my do file with all the commands that used for propensity score. RX_cat are the types of treatment, CSIZE_mimp, AGE are the covariates that I would like to adjust for   

To be honest, I am also so concerned about the sample size .However, these treatments are part of the clinical practice so I cannot exclude them .

Attached is the do file 

What version of Stata are you using? You don't need the prefix "xi" if you are using Stata version 11 or newer. Thus your model should look like:

glm totalcost_alltreatment i.RX_cat i.First_event [pw=_mmws],family (Gaussian) link(log) robust

margins RX_cat

I reiterate that your outcome is cost and you have very small sample size. You will probably not find any statistical difference between the groups.

Rimal,

Yes, you can use the predict to estimate the predicted value for each patient, but that does not provide the counterfactual. margins gives the treatment effect estimates and contrasts between groups. That is the meaning of marginal treatment effects.

In MMWS, the IPTW is an optional weight, it is not intended to replace the MMWS weights. So you can run the MMWS as designed, and then use the IPTW weight in the outcome model.

However, I am going to caution you again. Your outcome is cost, which is always a variable with a high degree of variability, and you have extremely small sample sizes...

HI Ariel:

Many thanks for your help .I followed your recommendation and it worked. 

I tried to use the marginal weights generated from mmws commands in the Weibull model to predict the individuals' life expectancy . However, it did not allow me to include the marginal weight in the model. STATA gave me an error message weight should stset .It is not appropriate to stset weight because it is not considered as the time to event. Attached is the commands with the syntax.

STATA also did not generate the marginal weights for all individuals although they have complete dataset  of the variables included in the mlogit model .I am trying to calculate the weights manually for these individuals .However, I am not sure which propensity scores corresponding for which treatment i.e. are the propensity score generated as PS1, PS2, PS3, PS4 corresponding for individuals receiving treatments with codes 0,1,2,3, respectively?.

I am really appreciate your effort to help me. I will cite your commands and papers in my research. 

Hi Rimal,

First, I am not clear as to why you believe that you cannot specify the weights when using stset? Given that the weights are derived from preintervention characteristics, then it is entirely appropriate to include them for the purpose of causal inference.

Second, you need to investigate why you have some individuals for which a propensity score is not being estimated. I don't believe that you should be estimating the weights manually if there is something inherently wrong with the person's values across the variables.

Yes, the propensity scores line up with the treatments, so ps1 and ps2 are for the first and second treatments (where the first is what you are referring to as the control group).

Ariel

Hi Ariel:

Many thanks ,I tired to include weights without stset them and it works .

Unfortunately ,I could not figure out the reason behind the missing data. At the moment, I imputed them manually. 

It was such a kindness from you to help me through all these procedures.

Hello Ariel.my name is Felix. Is it possible to share documents on your mmws with me via email. My email is: [email protected]. I'm facing similar challenge here. Thank you

Hi Ariel, I'm currently doing a study on estimating the causal effects under the CIA assumption: Using data from the labour intensive public works in Ghana. The treatment variable I'm dealing with is a nominal multiple treatment variable. I will need to make significant use of your literature. Please help me.