Quantifying the Impacts of Clinical Intervention Programs Using Propensity Score Matching in SAS®
- Article History
- RSS Feed
- Mark as New
- Mark as Read
- Bookmark
- Subscribe
- Printer Friendly Page
- Report Inappropriate Content
Watch this Ask the Expert session to learn how propensity score matching methods can be applied within SAS to control for confounder biases often present in historical clinical/claims data.
You will learn:
- The definition of propensity score matching.
- How SAS tools can be used to help detect potential biases in cohort-based studies.
- How PROC PSMATCH can be used to implement a propensity score matching algorithm.
- How to determine the quality of your propensity score matching algorithm.
- Beyond clinical intervention programs, what other applications in health care exist.
The questions from the Q&A segment held at the end of the webinar are listed below and the slides from the webinar are attached.
Q&A
Do you consider matched data an independent sample?
If the observations in the data set are independent before matching, then after matching they remain independent.
How to choose covariates? Should we choose as many as possible if there are many factors available?
I want to briefly talk about how we can discover potential confounders. One major way that I like to look at just data exploration with respect to considering confounders is looking at data visualization approaches. Covariance matrix and core plots are helpful. You can have basically a correlation heat map where you have all your variables on one axis and repeated on the other axis, and you can have a heat map specifying the degree of correlation between those variables in that lattice approach. Then also you can use variable importance rankings from various tree-based ensemble methods. For example, you could create a random forest model trained on the studied outcome of interest and then look at the variables that are most important. And then also, there's specialized knowledge in the field. Clinical analytics should not be done very far removed from actual clinicians.
Just to restate, propensity score reduces selection bias and possible confounders that affect outcome?
It can help. You can compute the propensity scores, use them to do matching, and then do your analysis based on the matched data set. It can usually reduce the effect of confounders.
What are the risks of using PSM when the propensity score model is not a good fit?
It's an iterative process. You calculate propensity score and then do matching. After, you assess your matched data set, so you can assess each covariate to see if the distribution for that covariate is close enough in both cohorts. If you are not satisfied with your result, you could go back, maybe change to a different matching method. There are lots of criteria you could adjust during matching and do this again and see if you get a better result. Proc PSMATCH fits a logistic model to compute propensity scores which is commonly used. There are lots of discussions in literature on choosing the propensity score model.
So can you just regress the outcome on treatment, controlling for all potential confounders?
It’s a different approach. The regression adjustment requires sufficient overlap between the covariate distributions for the treatment and control groups. You can ensure a sufficient covariate overlap by performing a propensity score analysis.
In theory if you included all confounders, is propensity score matching equivalent to an RCT?
Remember, going into this, you don't know what a confounder is. All you really can know is correlations between factors. But, in theory, if you include all the confounders, if you catch them all, is propensity score matching equivalent to a randomized control trial? I would be very careful to use the word equivalent because the matched data set is a good approximate to data from a random controlled study. It has very similar features. When you run your analysis on the matched data set, if you got the perfect matched data set, it shouldn't affect your results.
Is Proc psmatch only available in SAS Viya?
It's available in SAS 9. If you get SAS Viya, you will have access to it.
Does FULL option use the entire sample?
It does. It tries to keep all the observations when you do full matching. However, users can specify the maximum number of control units to be matched with each treated unit. If the specified total number of control units to be matched is less than the number of available control units, then constrained full matching is performed—that is, not all observations are matched.
In the model for your outcome, do you use the same covariates you used in the PS model?
You can perform outcome analysis as usual on the matched data set.
I know PSMATCH currently cannot handle more than 2 group matching. Is there any other available procs or macros in SAS to match exposure of more than 2 categories?
No.
How do you decide between adjusting for the PS, versus PS matching?
It depends on the study and analysis.
How is the propensity score matching method better than running a regression model with all the covariates?
It’s a different approach. The regression adjustment requires sufficient overlap between the covariate distributions for the treatment and control groups. You can ensure a sufficient covariate overlap by performing a propensity score analysis.
I'm a little confused. It doesn't seem like matched observations are independent observations if they are selected based on confounding factor(s).
Please refer to literature on propensity score matching methods.
Is Proc PSMATCh available in general SAS products? Or how do I know if PROC PSMATCH is available in my current SAS?
It's available in SAS 9. If you get SAS Viya, you will have access to it.
When doing regression models, we usually add control variables to limit the influence of confounding and other extraneous variables. Then, what are the differences between adopting a regression approach and using PSM? Is PSM always better than regression?
It’s a different approach. The regression adjustment requires sufficient overlap between the covariate distributions for the treatment and control groups. You can ensure a sufficient covariate overlap by performing a propensity score analysis.
Once we matched the 2 groups, using the new matched data, do we have to still adjust for those covariates? Also, if I do a KM curve for the treatment and control groups with the matched data, can I say that this KM is adjusted for all covariates?
You can perform outcome analysis as usual on the matched data set.
Can we specify multiple variables for EXACT option?
Yes. Only classification variables.
How do you match on propensity score distance within a caliper and match the covariate age within 3 years of age?
It depends on the data. One way might be taking a multiple step approach, performing propensity score matching and evaluating covariate age.
Do you need to have IML to be able to use MAH option?
No.
It is possible one patient shows in the matched sample multiple times. Should we make some considerations for that repeated patient after matching?
It depends on the study. There are several matching methods that match observations without replacement. Please refer to the method= option on the MATCH statement of PROC PSMATCH the SAS documentation.
How do you decide which matching method to pick? And when to use matching or weighting?
There are many discussions in literature on propensity score methods.
In propensity score matching, the caveat is that we could not match the most expensive or sick people from the SG to CG . How can we enhance the matching to include these population?
The MATCH statement in PROC PSMATCH matches observations in the control group to observations in the treatment group. The treated units are kept in the matched data set. How good a matching is depends on the matching method as well as the data.
Can you specify maximum distance between of matches on propensity scores, so that if the maximum distance is exceeded no match is made?
Yes. The caliper= option on the MATCH statement of PROC PSMATCH.
PS matching require a large sample size, how large is appropriate?
Different matching methods have different requirements. You can use the ASSESS statement to evaluate the matched data set.
Once we get the propensity score, how do we use it to examine the effect of the treatment on the outcomes?
You can perform your analysis as usual with the matched data set.
If I have used a parameter in the propensity score matching, can that variable then also be used in the final analyses or is it then only a part of the composite propensity score?
You can perform your analysis as usual with the matched data set.
If we do PS matching and the assessment shows that there's not much difference pre-post, should we exclude that covariate from the PS model?
You could and then repeat the process until you are satisfied with the matching results.
What would be the option for more than two groups matching?
PROC PSMATCH currently only covers two groups.
What are the key considerations when evaluating when to match with replacement (i.e., a control can be matched to more than one treated patient)?
It depends on the study and analysis.
Can we do an exact match for more than one variable?
Yes. Only classification variables.
Is PS matching only for outcomes with two levels?
PROC PSMATCH currently only covers treatment and control groups.
Recommended Resources
SAS Documentation for PROC PSMATCH
Vignette for “MatchIt” Package in R
Moving from SAS®9 to SAS® Viya®
Please see additional resources in the attached slide deck.
Want more tips? Be sure to subscribe to the Ask the Expert board to receive follow up Q&A, slides and recordings from other SAS Ask the Expert webinars.
