To think of the money people pay for terrible teachers in college ”teaching” these concepts at students in such terrible ways; the service you do to the community is really something, great explanation. Thank you!
I want to thank you for being able to explain so well such a difficult concept in only 8 minutes. I was able to understand even though I have no idea about PSM. Thanks a lot!
Bro, your videos are so clear. You have huge talent in explaining statistics. You help me a lot in my studies. I wish professors would try to explain better. Not sure if on Masters level, they just want you to understand stuff yourself...
Hi Ben. Why shouldn’t one use a multivariate regression model with all the observable factors that can potentially cause selection bias as additional explanatory variables? Moreover, a regression model can be used to eliminate bias caused by time invariant unobservable factors using the fixed effects estimator. I can’t see what’s the point in using matching techniques.
Interesting! I think a benefit of matching is that it doesn't need the assumption that the effects of treatment are constant across individuals. The OLS method requires that. With matching, you don't need a functional form to be correctly identified as in a regression model.
Hello Ben, can you make a video talking about "The minimum-biased (MB) estimator" and "The bias-corrected (BC) estimator" used to address SB? Is there an email address I can contact you to discuss PSM with specific questions?
I love this video ur the best. Quick question, for the weighted average that we take in the end how do we define the weight for each stara exactly? Do we just take the weight based on the amount of people in each starta? If yes then what if in one strata there’s 10 users in control and 200 in test taking the total would assume that the buckets are balanced right?
Can you do propensity score matching when your y-variable is a dummy. Does it make sense, since you're calculating the average in Y both before and after treatment? Thanks a lot!
Yes, if your Y var is a dummy, that just means anything you do beforehand with your explanatory variables will produce an outcome between 0 and 1, which kind of represents the percent likelihood of that individual having the outcome characteristic or not. Does that make sense?
Hi please, can you tell me what software you are using to prepare this video. I kind of like your black board. I would like to use it for some of the presentations I prepare for my work. If possible to know the brand of the touch pen you use.
Nice video! Quick question though: From my understanding you need a classifier (logit, boosted tree, etc) to estimate the propensity score. What training data do you use? Here is the problem I see: you try to find observations that have the same propensity score but with a different output variable (matching), so how can you build an accurate model with observations that have the same explanatory variables and yet a different output variable?
The training data is the set of covariates that has been measured, while using whether the subject received the treatment or not as a response variable. Then, you plug the same covariates you trained with into your predict function, and output probabilities of receiving treatment instead of a classification of receiving treatment, and use those probabilities to match similar subjects.
Like your video since it is clearly structured and technical at the same time! Good job and thank you! Probably you did purposely (for simplicity and didactic reasons) not mention probit estimation, the common support assumption and matching processes like nearest neighbour, caliper, kernel and Mahalanobis. If you had specific reasons why not doing so, I would be very happy to hear them - or discussing the topic anyhow! :)
Ben - there is a recent "negative research" publication by Gary King: th-cam.com/video/rBv39pK1iEs/w-d-xo.html where the general advice is to stop using propensity score matching. What are your thoughts?
To think of the money people pay for terrible teachers in college ”teaching” these concepts at students in such terrible ways; the service you do to the community is really something, great explanation. Thank you!
This is an amazing explanation to what PSM is and how to achieve it. Thank you for taking the time to share your knowledge.
I want to thank you for being able to explain so well such a difficult concept in only 8 minutes. I was able to understand even though I have no idea about PSM. Thanks a lot!
Bro, your videos are so clear. You have huge talent in explaining statistics. You help me a lot in my studies. I wish professors would try to explain better. Not sure if on Masters level, they just want you to understand stuff yourself...
Fascinating, thorough explanations. Thank you!
Hi, glad to hear it was helpful! Best, Ben
Does anyone know with SB where the error is coming from? I need resources and any help would be greatly appreciated.
Do you have a data file for practicing the method?
Thank you for your explanation, very intuitive and practical. Even for stat dummies like me !
Good explanation. The MUST watch video for the beginers
Hi Ben. Why shouldn’t one use a multivariate regression model with all the observable factors that can potentially cause selection bias as additional explanatory variables? Moreover, a regression model can be used to eliminate bias caused by time invariant unobservable factors using the fixed effects estimator. I can’t see what’s the point in using matching techniques.
Interesting! I think a benefit of matching is that it doesn't need the assumption that the effects of treatment are constant across individuals. The OLS method requires that. With matching, you don't need a functional form to be correctly identified as in a regression model.
Thank you for making it so simple to comprehend!
Hello Ben, can you make a video talking about "The minimum-biased (MB) estimator" and "The bias-corrected (BC) estimator" used to address SB?
Is there an email address I can contact you to discuss PSM with specific questions?
I'm so so grateful to you
I love this video ur the best. Quick question, for the weighted average that we take in the end how do we define the weight for each stara exactly?
Do we just take the weight based on the amount of people in each starta?
If yes then what if in one strata there’s 10 users in control and 200 in test taking the total would assume that the buckets are balanced right?
Better explained than in Angrist and Pischke!!!!!!!
Great video! How can I do greedy matching in SPSS ( with no R plagins and other)? I would be very thankful for any advice
amazing clarity in explanation !!
thanks..
Can you do propensity score matching when your y-variable is a dummy. Does it make sense, since you're calculating the average in Y both before and after treatment? Thanks a lot!
Yes, if your Y var is a dummy, that just means anything you do beforehand with your explanatory variables will produce an outcome between 0 and 1, which kind of represents the percent likelihood of that individual having the outcome characteristic or not. Does that make sense?
Hi please, can you tell me what software you are using to prepare this video. I kind of like your black board. I would like to use it for some of the presentations I prepare for my work.
If possible to know the brand of the touch pen you use.
Nice video! Quick question though: From my understanding you need a classifier (logit, boosted tree, etc) to estimate the propensity score. What training data do you use? Here is the problem I see: you try to find observations that have the same propensity score but with a different output variable (matching), so how can you build an accurate model with observations that have the same explanatory variables and yet a different output variable?
The training data is the set of covariates that has been measured, while using whether the subject received the treatment or not as a response variable. Then, you plug the same covariates you trained with into your predict function, and output probabilities of receiving treatment instead of a classification of receiving treatment, and use those probabilities to match similar subjects.
thanks for your excellent tutor! very intuitive and helpful!
That was really helpful😊thanks
Thanks ben
Like your video since it is clearly structured and technical at the same time! Good job and thank you!
Probably you did purposely (for simplicity and didactic reasons) not mention probit estimation, the common support assumption and matching processes like nearest neighbour, caliper, kernel and Mahalanobis. If you had specific reasons why not doing so, I would be very happy to hear them - or discussing the topic anyhow! :)
This is very good, thank you!
What is Average Cause Effect?
thank you
Excellent. Thank you!
the audio is mad quiet but good video
Ben - there is a recent "negative research" publication by Gary King: th-cam.com/video/rBv39pK1iEs/w-d-xo.html where the general advice is to stop using propensity score matching. What are your thoughts?