While estimating a model for gender wage gap using Oaxaca-Blinder and addressing the selectivity bias through Heckman, do we have to include the observations of women who are not a part of the workforce?
Hi Mikko, I have a question. At 7:42, you mentioned that well-being doesn't influence job performance, but I think well-being may influence job performance right?
Great video.........Am trying analysis factors influencing the adoption of a technology, I used the Unified Theory of Acceptance and Use of technology model to construct my questionnaire..............................What steps do i take in other to use the Heckman two stage for the analysis
@@mronkko thanks for the information about the Books but going through am not seeing the link between my responses and the Heckman tool.............Simply put i am using a total of 125 response that has no missing values but looking for a way to use the tool since the model i used for the construction of my questionnaire is the UTAUT model
Thanks for this helpful video! One question is, can we use the same principle when analyzing the panel data? Meaning, can we analyze the heckman selection model on the panel data by running xtprobit in the first stage, calculating IMR, and running the xtreg with IMR? Or do we need to consider anything else? Thanks for your time in advance!
My gut feeling is that if you do the within transformation to all data, you can just apply Hekman sample model, but I would need to check it. So yes, but verify it from a good econometrics book.If you are using Stata, extended regression models (erm) tools will handle the kind of analysis that you are suggesting.
I thought about this a bit today and the answer is not straightforward. It depends on the missing data process. If the missing data is purely between person and you are estimating within person effects, then I do not think selection effect matters. (I might be wrong about this.) So you need to think about the levels of effects when you make any decisions. See journals.sagepub.com/doi/10.1177/1094428119877457 for discussion on the different effects.
zeta is the error term of the selection model. See the previous slide. Now that you asked it, it is very confusing that I use u for error term in this slide but zeta in the previous slide. Why, you may ask? Because I did this slide using my slide on probit models and that slide happened to use u for error term.
Thanks for the feedback. Can you be more specific: what exactly did you find confusing about it? I will run a course next academic year where I will use this video and I might record it if I know what the pain points are.
Hey Mikko, great video! Thanks so much for detailed explanation. I would like to ask a question about using Heckman selection correction to control for self-selection bias. I have been doing a survey study with participants from social media sites (e.g., Twitter/ Facebook). I reached out to 500 people on social media and only received 200 responses. Because of the low response rate, I would need to correct the selection bias introduced by the low response rate. I tried to use SPSS to do the test. In the first stage of the analysis, a probit regression needs to be used to estimate the likelihood of users not participating in the survey. However, I do not have data for non-participants (as they did not respond to my invitation). Thus, the selection model is of no use because the selection dependent variable is participation status, a binary variable that accounts for participation vs. non-participation. My questions are as follows: 1) Can I use the Heckman selection correction to control for self-selection bias? If so, how do I run the selection model when the data for non-participants are unavailable? 2) Are there any other ways to address self-selection bias in an online survey study such as mine? 3) Does Heckman's correction apply to logistic regression or ordinal logistic regression? I realised its second step model is a linear regression but my dependent variable is a binary one (hence logistic regression). Many thanks for answering my questions. Much appreciated!
1) You cannot. Hekman sample selection model addresses missing data in the dependent variable assuming that the missigness depends on the missing value itself. 2) If you do not have any data on the non-respondent, then I do not think that you can say anything about sample selection empirically. I have a slideset on survey response, but I just realized that I have forgotten to record it on video. I will do it in a couple of days and will post the video on the channel. 3) Heckman uses profit regression, not logistic.
@@mronkko Thanks very much Mikko! Regarding your comment on 3), can I understand that in my case, where the logistic regression is used (having a binary outcome variable, NOT continuous), the assumption of joint normality for the error terms does not hold. Hence, the Heckman correction is not applicable? What are other approaches that I can use to address the possible self-selection bias in my case?
@@maggiewang1336 I am not aware of any valid uses of logistic regression in selection models. Based on your description, I do not think that modern missing data techniques are applicable. When you cannot use the techniques, you are often left out with just reporting your assessment on whether a problem exits or not. For example comparing early and late responses, but note that this is not applicable to all scenarios. (I will post a video about this in the future.)
I do not use SPSS so I cannot help you with that. With Stata, I think the best way to do sample selection models is ERM: ua-cam.com/video/bPhNq6RYd-I/v-deo.html
Thanks for sharing, this is a very helpful video for explaining the IMR!
Thanks, buddy. Keep up the good job you're doing!
Your videos are great. I’m glad I found this channel. Will you be doing a video on Tobit models in the future?
Great job buddy and keep the ball rolling
Thanks! Will do!
Amazing video, thanks from Norway :)
You are welcome!
While estimating a model for gender wage gap using Oaxaca-Blinder and addressing the selectivity bias through Heckman, do we have to include the observations of women who are not a part of the workforce?
Well-done, as usual! I truly appreciate your time and effort. 0:54 subtitles should be "missing not random." :)
Thanks. I fixed the caption. The captions are done by non-experts (students or generic captioning service) so there can be errors like this.
thank you for your hard work this is super well explained
You are welcome!
Hi Mikko, I have a question. At 7:42, you mentioned that well-being doesn't influence job performance, but I think well-being may influence job performance right?
It might, but in the example that I got from the Enders book it was assumed not to ;)
Great video.........Am trying analysis factors influencing the adoption of a technology, I used the Unified Theory of Acceptance and Use of technology model to construct my questionnaire..............................What steps do i take in other to use the Heckman two stage for the analysis
@@mronkko thanks for the information about the Books but going through am not seeing the link between my responses and the Heckman tool.............Simply put i am using a total of 125 response that has no missing values but looking for a way to use the tool since the model i used for the construction of my questionnaire is the UTAUT model
Thanks for sharing, this is very helpful.
Glad it was helpful!
Thanks a lot from Korea.
Welcome!
Thanks for this helpful video! One question is, can we use the same principle when analyzing the panel data? Meaning, can we analyze the heckman selection model on the panel data by running xtprobit in the first stage, calculating IMR, and running the xtreg with IMR? Or do we need to consider anything else? Thanks for your time in advance!
My gut feeling is that if you do the within transformation to all data, you can just apply Hekman sample model, but I would need to check it. So yes, but verify it from a good econometrics book.If you are using Stata, extended regression models (erm) tools will handle the kind of analysis that you are suggesting.
I thought about this a bit today and the answer is not straightforward. It depends on the missing data process. If the missing data is purely between person and you are estimating within person effects, then I do not think selection effect matters. (I might be wrong about this.) So you need to think about the levels of effects when you make any decisions. See journals.sagepub.com/doi/10.1177/1094428119877457 for discussion on the different effects.
thanks!
You are welcome
great video. I have a question please. In the slide of minute 9, what is zeta? thank you!
zeta is the error term of the selection model. See the previous slide.
Now that you asked it, it is very confusing that I use u for error term in this slide but zeta in the previous slide. Why, you may ask? Because I did this slide using my slide on probit models and that slide happened to use u for error term.
@@mronkko Thank you very much for the quick reply!
You always have very good videos, but this one was confusing.
Thanks for the feedback. Can you be more specific: what exactly did you find confusing about it? I will run a course next academic year where I will use this video and I might record it if I know what the pain points are.
I'm amazed to watch all of your videos. Can you give me some recommended books for learning statistics?
@@mronkko thank you so much. 🙏🙏
Hey Mikko, great video! Thanks so much for detailed explanation. I would like to ask a question about using Heckman selection correction to control for self-selection bias.
I have been doing a survey study with participants from social media sites (e.g., Twitter/ Facebook). I reached out to 500 people on social media and only received 200 responses. Because of the low response rate, I would need to correct the selection bias introduced by the low response rate.
I tried to use SPSS to do the test. In the first stage of the analysis, a probit regression needs to be used to estimate the likelihood of users not participating in the survey. However, I do not have data for non-participants (as they did not respond to my invitation). Thus, the selection model is of no use because the selection dependent variable is participation status, a binary variable that accounts for participation vs. non-participation.
My questions are as follows:
1) Can I use the Heckman selection correction to control for self-selection bias? If so, how do I run the selection model when the data for non-participants are unavailable?
2) Are there any other ways to address self-selection bias in an online survey study such as mine?
3) Does Heckman's correction apply to logistic regression or ordinal logistic regression? I realised its second step model is a linear regression but my dependent variable is a binary one (hence logistic regression).
Many thanks for answering my questions. Much appreciated!
1) You cannot. Hekman sample selection model addresses missing data in the dependent variable assuming that the missigness depends on the missing value itself.
2) If you do not have any data on the non-respondent, then I do not think that you can say anything about sample selection empirically. I have a slideset on survey response, but I just realized that I have forgotten to record it on video. I will do it in a couple of days and will post the video on the channel.
3) Heckman uses profit regression, not logistic.
@@mronkko Thanks very much Mikko! Regarding your comment on 3), can I understand that in my case, where the logistic regression is used (having a binary outcome variable, NOT continuous), the assumption of joint normality for the error terms does not hold. Hence, the Heckman correction is not applicable? What are other approaches that I can use to address the possible self-selection bias in my case?
@@maggiewang1336 I am not aware of any valid uses of logistic regression in selection models. Based on your description, I do not think that modern missing data techniques are applicable. When you cannot use the techniques, you are often left out with just reporting your assessment on whether a problem exits or not. For example comparing early and late responses, but note that this is not applicable to all scenarios. (I will post a video about this in the future.)
@@mronkko Noted & thanks!!! It helps a lot :)
thanks bro
You are welcome
Sir.. If its possible for u to make step by step video on running the Heckman sample selection model..using SPSS or Stata ....plisss
I do not use SPSS so I cannot help you with that. With Stata, I think the best way to do sample selection models is ERM: ua-cam.com/video/bPhNq6RYd-I/v-deo.html
@@mronkko Respected Assoc. Prof. Thank you, kindly.