I Have Results From PROC PSMATCH. Now What? Q&A, Slides, and On-Demand Recording
Watch this Ask the Expert session to explore the possibilities of no-code real-world evidence analytics in SAS® Viya®.
You will learn:
- How custom steps and SAS Visual Analytics can automate no-code analytics.
- To use analytics visualization for real-time model decisioning.
- How to choose a correct propensity score matching model.
- What methods to use for causal inference after matching.
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
Does the match differ whether or not we use exact feature?
Yes, it does. If you use exact feature, you are forcing that variable to be same for that pair or that ratio. You are potentially excluding some subjects from being considered as a match and therefore excluded from the control group. Your conclusions might be the same, but the groups that you move forward with for causal effect treatment are going to be different if you choose to use exact matching. So, exact matching narrows the number of subjects that could be a part of the final causal treatment because it changes who is allowed and who is not allowed to be a match.
In your psm code, you had a weight, is that something that is part of the dataset or standard part of psmatch?
When you run a propensity score model, you can choose the weighting and it's added to the output dataset that Proc PSMatch provides. There are four options listed below. It is not part of my input dataset but calculated during the matching.
Is the "magic" you displayed (no code analysis and results display) a function of SAS Viya or SAS Studio, or a combination of the two?
It is a combination of the two. First. I created that custom step in SAS Studio. Then I applied it in the Visual Analytics module as part of the dashboard. I like the visual interface because you get to adjust the analytic parameters and can see real time results.
Can PS Match allow for a pre and post match graph to show a density plot?
Yes. This the dashboard that I created here, those were things that I chose. You can create plots that show those density plots. Density plots are not an automatic output from Proc PSMatch, but you can create any plots or charts needed for your analytic process in Visual Analytics.
I want to highlight the matching method for each model shown on the results page. Can I do that?
Absolutely. When you are creating your dashboard, you can add whatever is an input or output in the model. This process saves each value to be pulled out and displayed on the dashboard at any time.
Can I use this method for other real-world evidence analytics?
This is a great method to use for when you have an analytic procedure or practice where you're exploring and testing. Anytime you reiterate with small changes, additional variables, or different methods, this is excellent. If you have a lot of stakeholders who need to make real-time decisions, this is a great way for everyone to see the results and discuss nuances in different methods without having to e-mail tons of PDFs. I created this tool for propensity score matching because that's what my background was in, but you can do this for any analytic method that you want.
What types of analytics make a good candidate for implementing this solution?
There are a few scenarios that make this analytic process extremely useful. First, any time you have to reiterate modeling, when you have input from other researchers, or when they want to test other methods without having coding skills. An interactive dashboard is an excellent way to bring in other voices and subject matter experts. Anybody can have access to this, and anybody can run those analytics. If you need more voices and more expertise around the results, this is a perfect method for you. Keep in mind that anytime we have big data and/or are running computationally intensive models, we want to be very cognizant of the resources that would be required for the methods that we're using.
I have used both PSMatch as well as a manual matching of cases and controls based on exact age, gender. My assumption is PSMATCH can't/should not replace the other matching, as they have different intentions. Am I correct?
PSMATCH can replace manual matching. There are options within the PROC to allow for exact matching.
Does this work with SAS 9.4 and Enterprise Guide?
PROC PSMATCH is available in SAS 9.4. Custom steps are specific to Viya.
Would you please provide a link to the book? Thanks!
Can we use PSM for small population also? Any population size constraint?
There isn't a strict sample size limit for using SAS PROC PSMATCH. However, the larger your sample size, the more reliable and robust the results will be. Specifically, using a large sample size (generally >100) is recommended when using approximations like the standard normal distribution. Propensity score matching is generally suitable for datasets with a large number of observations.
Is it possible to do exact match on variables only (e.g., age, sex, etc.) and not matching on propensity score with PROC PSMATCH?
In the match statement, you can assign method=exact, k=x.
Do you have an example with entropy balance?
Yes, there is an example in the documentation: https://go.documentation.sas.com/doc/en/pgmsascdc/v_062/statug/statug_psmatch_examples10.htm
Can it handle the matching for 1:15 for 500,000 cases with ~200,000,00 controls? I used a HASH object to get this done. I want to know if PSMATCH can do a better job?
For the optimal matching methods, no. This exceeds the size of problem that can be represented and solved by the underlying network optimization problem used by the optimal matching methods. For a greedy matching, it depends on hardware and storage. For reproducibility reasons, there will be a lot sorting of the data going on behind the scenes via some utility files, and for that size of data depending on the system you may encounter issues with that, and it will also be time consuming if feasible. Generally, for that size of data weighting scales is much better.
Does PS Match have an option for density plots instead of box plots?
As shown in various doc examples the CDFPlots, and mostly all the other balance diagnostics, already provide a comparison between various sets of data i.e. all observations, observations in a support region, matched observations, and matched observations with matching weights incorporated. You can further restrict that to be only two sets of data, all and a set dependent on the method you're using by via the NODETAILS plots option as discussed in the doc. The procedure by default includes balance diagnostic assessments for the input data.
Recommended Resources
SAS Visual Analytics Learning Subscription
SAS Visual Analytics 1 for SAS Viya: Basics
SAS Visual Analytics 2 for SAS Viya: Advanced
Developing Custom Steps with SAS Studio Analyst
Quantifying the Impacts of Clinical Intervention Programs Using Propensity Score Matching in SAS
SAS Studio: Working with Custom Steps
Real World Health Care Data Analysis
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.
