I am a masters student studying social work and even when the focus of the research is different the process of calculating these scores is more or less similar. This video was a great deal of help and extremely digestible, thank you so much!
Very nice video. Covered the big picture information on why we do it and how we do it in a very concise format. Would be great to get the details on how to Performa matching in a stats package (is it a manual process or can be automated). That would probably make it much longer than 5 mins, but I believe it would be very useful). Another thing: I wanted to watch with subtitle but somehow the closed caption is Vietnamese. Thank you for the video!
Thank for the concise videos that explain the overall concept very well. It was very easy to understand. There are 3 implementation details during the matching phase (at 5:00) hoping you could clarify. 1. After we get the score, if there are two rows with scores 0.5 in control group (where `SGLT2 = No`), should both of them be matched to row 3 in treatment group? 2. In the real world, the score would rarely be exactly the same if we allow infinite decimal. How do you consider a match? Would there be an arbitrary numeric difference allowed? 3. If there are two rows in treatment with score 9.15 and 9.18 (row A and B), and two rows in control with score 9.16 and 9.25 (row C and D). Would row A and B both be matched to C? Or would row A matched to C as they are have the smallest difference, then row B has to be matched to D?
Can you please explain the slide at 5:10? Why don't we have the same number of persons, means...? Is it just that we do not take into account people without a match?
Hi Michael, Appreciate this breakdown, despite most of it going over my head. When you say these statistical methods do not mimic RCTs, what exactly does that mean? How closely do we get to the results of a RCT using this method, and has this method been verified against the findings of RCTs? The reason I'm asking this is b/c I came across a cohort article comparing adolescents who took antipsychotic medication vs. those who didn't, and compared how they were doing 5 years later. The article suggests, after using IPTW, that antipsychotic medication in those w/ their first episode of psychosis, actually makes for a worse 5 year outcome. I'm skeptical though b/c it's a cohort study and no randomization was done. The authors acknowledge this weakness, and then state that the Stabilized IPTW is used to eliminate the possibility that those with worse psychosis were the ones who were given an antipsychotic medication. (I think this is the most probable explanation for why those who were given AP medication fared worse. . .simply b/c they already were experiencing more profound psychosis and so we would expect them to be doing worse at a 5 year follow-up). Seems super fishy to me. . .any thoughts here are appreciated, thx.
You are totally right to be skeptical. There’s no way to answer this Q “how close do we get to the results of a RCT”. It would be like asking how much does this glass of wine taste like a beer. They are two different things. To learn more about why RCTs are so powerful and what randomization achieves (that no cohort study can) here is link to my 6min crash course on RCTs: ua-cam.com/video/oQt8jR5RgVQ/v-deo.htmlsi=cpMn3c6YQnWYj3wv
if probability score matching improves the validity of overall impact of cofounders, why aren't they used mor often in entertainment, movies, ect? Sounds like an entertaining way to infer or narrow in on a hunch, so im confused why this isn't as generally mentioned? I had to go down so many A.I. rabbit holes to even ask the question - What are propensity scores.
Case control matching is unrelated. Here is great resource on everything you need to know about case control studies sphweb.bumc.bu.edu/otlt/MPH-Modules/EP/EP713_Case-Control/
thank you for this great explanation! Is there also a possibility when having 3 treatment groups to calculate propensity scores with the same strategy?
You’re most welcome. There is but that gets complicated fast! If there are 3 groups and one is the ref then I do two pair wise comparisons to the ref. Hope that helps.
Pro - you retain all participants (or almost all) Con - harder to explain to readers, you create a “synthetic” population (which is also hard to understand!) The risk of bias is similar with matching vs weighting. So I stick with matching
@@Fralickmike thanks. one other question please, how come some do both propensity score matching/weighting and then Cox adjustments for variables? if the population is similar after propensity score matching then what does the cox adjustments add?
that is a good question. you are right that it would be ok to pick at random. In some studies people match "many to one". so you could keep all 3. But the "ideal" approach is 1:1 based on prior studies.
Can one use the propensity score as the probability of death in 14 days (say, using an online calculator like CRASH for traumatic brain injury) and then match patients with similar probabilities of death to see if a particular treatment has lower mortality than another one? If the treatment had an effect, there would be a difference in the actual mortality rates? Is that a valid propensity matched analysis? How does one find the sample size for such an analysis?
Been reading articles about propensity scoring in multiple different articles but this is by far the most reasonable and down to earth description
thank you very much!
Wah, many thanks, Sir. This 5 mins is probably more useful than a paid stats course🙏👍😊
Michael, the man that you’re!! Thank you from the bottom of my heart.
I am a masters student studying social work and even when the focus of the research is different the process of calculating these scores is more or less similar. This video was a great deal of help and extremely digestible, thank you so much!
That’s very kind of you. And I’m glad you found it helpful !
Very nice video. Covered the big picture information on why we do it and how we do it in a very concise format.
Would be great to get the details on how to Performa matching in a stats package (is it a manual process or can be automated). That would probably make it much longer than 5 mins, but I believe it would be very useful). Another thing: I wanted to watch with subtitle but somehow the closed caption is Vietnamese.
Thank you for the video!
Excellently lucid explanation!! Great job ✅
Fantastic! Nicest explanation of PSM ever
Thx so much!
genius! deeply appreciate this video! helpful for med students drowned by statistical terms
Thank you !
Omg!! This was perfect thanks!
Thank you for the video!! saved me when I got totally lost in my intermediate pharmacoepidemiology class 😂
Wonderfull, thank you so much!
Excellent, clear and concise interpretation!
Thank you!!
Thank for the concise videos that explain the overall concept very well. It was very easy to understand.
There are 3 implementation details during the matching phase (at 5:00) hoping you could clarify.
1. After we get the score, if there are two rows with scores 0.5 in control group (where `SGLT2 = No`), should both of them be matched to row 3 in treatment group?
2. In the real world, the score would rarely be exactly the same if we allow infinite decimal. How do you consider a match? Would there be an arbitrary numeric difference allowed?
3. If there are two rows in treatment with score 9.15 and 9.18 (row A and B), and two rows in control with score 9.16 and 9.25 (row C and D). Would row A and B both be matched to C? Or would row A matched to C as they are have the smallest difference, then row B has to be matched to D?
The stats section of this article will help answer your Qs
www.acpjournals.org/doi/10.7326/M19-2610
Can you please explain the slide at 5:10?
Why don't we have the same number of persons, means...?
Is it just that we do not take into account people without a match?
Thank you so much for making this video!
My pleasure. Thank you for your kind words!!
thank you, I finally got it (2 days before my exam) :)
Great to hear!
thank you so much for this video helped me with my journal club!
wow! Thank you for such a concise and clear explanation! It was pure pleasure to watch!
thank you!!
Thanks for the clear explanation!
My pleasure. Thank you for watching
Many Thanks sir. Your explanation is clear and easy to understand.
You are most welcome
Excellent presentation !
very informative and concise.
thanks for sharing
Thank you!
Sir we need more and more videos like this. Survival analysis, what test to use in what situation etc.
Thank you! i agree and hope to add more soon
Hi Michael,
Appreciate this breakdown, despite most of it going over my head.
When you say these statistical methods do not mimic RCTs, what exactly does that mean? How closely do we get to the results of a RCT using this method, and has this method been verified against the findings of RCTs?
The reason I'm asking this is b/c I came across a cohort article comparing adolescents who took antipsychotic medication vs. those who didn't, and compared how they were doing 5 years later. The article suggests, after using IPTW, that antipsychotic medication in those w/ their first episode of psychosis, actually makes for a worse 5 year outcome.
I'm skeptical though b/c it's a cohort study and no randomization was done. The authors acknowledge this weakness, and then state that the Stabilized IPTW is used to eliminate the possibility that those with worse psychosis were the ones who were given an antipsychotic medication. (I think this is the most probable explanation for why those who were given AP medication fared worse. . .simply b/c they already were experiencing more profound psychosis and so we would expect them to be doing worse at a 5 year follow-up).
Seems super fishy to me. . .any thoughts here are appreciated, thx.
You are totally right to be skeptical. There’s no way to answer this Q “how close do we get to the results of a RCT”.
It would be like asking how much does this glass of wine taste like a beer. They are two different things.
To learn more about why RCTs are so powerful and what randomization achieves (that no cohort study can) here is link to my 6min crash course on RCTs:
ua-cam.com/video/oQt8jR5RgVQ/v-deo.htmlsi=cpMn3c6YQnWYj3wv
awesome, thank you for the insanely prompt and helpful response Michael!@@Fralickmike
if probability score matching improves the validity of overall impact of cofounders, why aren't they used mor often in entertainment, movies, ect? Sounds like an entertaining way to infer or narrow in on a hunch, so im confused why this isn't as generally mentioned? I had to go down so many A.I. rabbit holes to even ask the question - What are propensity scores.
Really helpful, I wanted to ask What is
the difference between Propensity score matching and Case control matching (SPSS)?
Case control matching is unrelated.
Here is great resource on everything you need to know about case control studies
sphweb.bumc.bu.edu/otlt/MPH-Modules/EP/EP713_Case-Control/
Very informative and concise introduction! Thank you very much!
thank you for this great explanation! Is there also a possibility when having 3 treatment groups to calculate propensity scores with the same strategy?
You’re most welcome. There is but that gets complicated fast!
If there are 3 groups and one is the ref then I do two pair wise comparisons to the ref. Hope that helps.
Thank you dear teacher. Excellent, in spite of it wasn't in Spanish or without subtitles
You are welcome!
Nice job. Thank you. 😊
You are most welcome
Thanks for the video
thanks, great video.
what are the pros and cons of weighting vs matching?
Pro - you retain all participants (or almost all)
Con - harder to explain to readers, you create a “synthetic” population (which is also hard to understand!)
The risk of bias is similar with matching vs weighting. So I stick with matching
@@Fralickmike thanks. one other question please, how come some do both propensity score matching/weighting and then Cox adjustments for variables? if the population is similar after propensity score matching then what does the cox adjustments add?
great video!
Thank you!!
There are 3 scores as 0.6, would it be fair/unbiased to match with random 2 only?
that is a good question. you are right that it would be ok to pick at random. In some studies people match "many to one". so you could keep all 3. But the "ideal" approach is 1:1 based on prior studies.
@@Fralickmike Thank you!!
Can one use the propensity score as the probability of death in 14 days (say, using an online calculator like CRASH for traumatic brain injury) and then match patients with similar probabilities of death to see if a particular treatment has lower mortality than another one? If the treatment had an effect, there would be a difference in the actual mortality rates? Is that a valid propensity matched analysis? How does one find the sample size for such an analysis?
No. Propensity scores are not used to match on outcomes. That’s a disease risk score
thank you for this
You are most welcome
Can you use mediators in addition to / in place of confounders as your variables when calculating propensity score?
Good question. I have never done this before and would probably suggest against it.
kindly guide how to analysis step by step in.spss
avoid SPSS. use R, it is way easier for propensity score methods like the matchit package
how do we get weights as 0.6 and 0.9 ? can you please elaborate
what do you mean?
You run the regression in your software using the age and sex as inouts
🤩
Didnt't understand anything.
Very good lesson except that you are extremely fast. I wish you could slow down your pace
Thank you for feedback. Note you can slow the speed to 0.75 or 0.5 to slow it down