I am thrilled to have found this channel. I haven't studied econometrics since my first degree and now I'm trying to model a dynamic panel data model and your videos are helping my confidence - thank you! I would appreciate more videos on Stata coding for xtdpdml, too.
You are welcome. I do not do much software demos because I teach multiple software and because, particularly with Stata, the documentation explains the software well. Understanding the conceptual side is more challenging, and once you know that, specifying the command is not very challenging.
you sir have just saved my dissertation. I do not normally comment on videos like this, but wanted to acknowledge your efforts and let you know that you are making an impact! Good work :)
@@mronkko @Mikko Rönkkö just a quick question - I have read Roodman (2009) and watched this video a few times. What are internal and external instruments? Are internal endogenous and external exogenous? My model tries to look at the impact of CSR on firm performance and takes the following structure: (FirmPerf)it = a + B1(FirmPerf)i(t-1) + B2(CSR)it + B3X'it + e, where X' is a vector of firm-level covariates that are used in similar studies. I wanted to use system GMM. I have used the lagged dependent variable as my endogenous internal instrument and used the predetermined covariates (X') as external instruments all lagged for two periods as I believe lagging them make them exogenous (as they will not be correlated with the error). Is this along the right track?
@@KiranSingh-vf8nc Internal instrument are a part of the model (p.100 in Roodman, 2009) and external instruments are additional variables that are not of specific interest but are used for causal identification of the model. If you look at any introductory econometrics book, the instrumental variables explained in these books are external instruments. One could say that internal instruments are endogenous to the system, if we consider all time periods. In an equation by equation consideration instruments are always exogenous. Unfortunately I do not have the capacity to focus on specific research questions in TH-cam comments. These always require a bit of thinking.
Hello, I normally do not comment on TH-cam, but your explanation is very clear and precise, extremely helpful for me as a non statistical student, thank you!
Dear Miko, It’s a privilege to hear your lecture on the Arrelano Bond approach to dynamic panel models. You’ve made a difficult concept much easier to understand and apply. Thank you for putting this video on youtube. However, I have a question on the Arellano-Bond test to check whether the errors in the first differencing model are correlated Arellano and Bond in their diagnostic testing perform the following: AR(1) test || where H0 states “No autocorrelation of order 1” AR(2) test || where H0 states “No autocorrelation of order 2” AR(3) test || where H0 states “No autocorrelation of order 3” From my understanding to ensure the first differencing model is correct and the coefficients can be interpreted appropriately (i.e. the model is not misspecified), you MUST be able to: Reject H0 for AR(1) (i.e., p0.10): i.e., where H0: There is no second-order autocorrelation of error terms Accept H0 (i.e. Fail to Reject H0) for AR(3) (i.e., p>0.10): i.e., where H0: There is no third-order autocorrelation of error terms MY QUESTION Why MUST WE ACCEPT H0 for AR(1)? i.e. Why do we want to conclude that there is first-order autocorrelation of the error terms in the first differencing model Δut & Δut-1 ? My answer (which I believe is correct - but perhaps it is not and I'm going down a deep rabbit hole) is: 1. It is expected that there will be some degree of first-order autocorrelation in the errors for the first differencing model. 2. When you include the first instrumental variable (yt-2) and sequentially other instrumental variables (yt-3, ….etc) to the first differencing model to help remove endogeneity, effectively we want to ensure that there is no second-order autocorrelation in the errors AR(2) or third order autocorrelation in the errors AR(3) . Why? If we were to identify autocorrelation in the error terms (i.e. by Accepting H0) in the second or third order, THEN it signals that we have omitted variables and this invalidates any correct interpretation of our estimated coefficients. The testing from my understanding For AR(1), I assume it tests the correlation between Δut & Δut-1 For AR(2), I assume it tests the correlation between Δut & Δut-1 + Δut-2 For AR(3), I assume it tests the correlation between Δut & Δut-1 , Δut-2 & Δut-3 An explanation that I found on the internet as to why we must Accept H0 for AR(1) Note: As a novice, I don’t seem to understand this explanation though “The aim of the Arellano-Bond tests is to check whether the idiosyncratic error term is serially correlated. The test is conducted for the first-differenced errors. If the error term in levels is serially uncorrelated, this implies that the error term in first differences has a negative first-order serial correlation (with a correlation coefficient of -0.5) but no second-order or higher-order serial correlation. Thus, we should reject the null hypothesis of no first-order serial correlation in first differences (AR(1) test) but should not reject the null hypothesis of no higher-order serial correlation in first differences (AR(2), AR(3), ...). If you do not reject the null hypothesis of the AR(1) test, this could indicate that your idiosyncratic error term in levels is highly serially correlated. In the extreme case, the error term in levels follows a random walk such that the first-differenced errors are serially uncorrelated. Such a situation would indeed invalidate the MSM.” Any help you can provide me in addressing my question would be super appreciated. Warm regards, Reuben
Well-structured question! The AB estimator assumes that the error term u (after first-differencing) is NOT serially correlated. The H0 in the autocorrelation tests is that there is no autocorrelation. Ideally, you would fail to reject H0 for all lags. However, often we find that u is serially correlated for short lags. When this happens, we need to increase the lags of the instruments to make sure that they satisfy the exclusion criterion. One way to understand this - and this was how I learned to understand it - is to look a the dynamic panel model in wide format data (see the last stlide in the talk) and then add first-order autocorrelation between the errors and think what happens to the estimates.
@@mronkko Hi Miko, Thanks for the super quick response. I really appreciate it! Your response reassured me that I am on the correct track of getting more understanding of this material. I understand that the STATA command: *xtabond2* is a bit like a “black box” and spits out results, but I don’t believe too many people that apply it fully understand the mechanics behind how the output is generated [though Roodman(2009) who programmed the xtabond2 code has provided a thorough paper on how it works with examples) I agree with you that the GMM estimator was cutting edge (30 years ago); and I do wonder (perhaps like yourself) why more researchers don’t use the maximum likelihood estimator (ML estimator) that can be applied with the STATA command: *xtdpdml*. The ML estimator has the benefit of handling missing values, which I believe the GMM estimator does not? I could be incorrect in my knowledge. I believe researchers use the STATA command: *xtabond2* because it’s (1) relatively easy to apply without deep understanding; (2) most reviewers of journal articles also might have a preliminary understanding of dynamic GMM, but may not have an understanding of the STATA command *xtdpdml* and most researchers/reviewers don’t know too much about cross-lagged panel models in my field (finance). That said, there is also a lot of criticism of cross-lagged panel models as well in the literature. That said, I wonder how well any of these STATA commands (i.e. *xtabond2* or *xtdpdml*) handle missing values in the panel dataset. Perhaps *xtdpdml* handles missing values because it uses a ML estimator, which I understand can do so. I suspect the STATA command *xtabond2* deletes observations through a listwise deletion; would the possible result that it can lead to biased results (especially if you have a lot of missing values in your panel dataset) For your information (though I’m sure you are aware), there is a new STATA command that suggests the *xtabond2* has errors, which is *xtdpdgmm* (see www.kripfganz.de/stata/xtdpdgmm.html) I shall keep on watching your videos, as they are super helpful. Thanks again and stay well, R p.s. sorry for this lengthy reply. I just got carried away
@@reuben_1973 No worries, I tend to get carried away too quite often. I agree with your interpretation that many researchers use AB estimator as implemented in Stata as a black box. I also think that this is because of modeling their studies based on past research. That being said, I do not think that AB is entirely outdated because GMM is a bit more flexible in its assumptions compared to ML. Whether this makes a difference is another matter. I do not think that neither of the commands that you mention deal with missing data. If you want to have an estimator that can use observations where just a part of the variables are measured, you can do that by specifying the model using wide format data and applying sem command in Stata.
Sir, thank you so much for this video. You are amazing. I'd like to ask a trivial question, if you (or anyone who sees this) could answer me: when differentiating, why doesn't beta_0 get canceled as well? I mean, when we subtract y_t-1 on both sides, wouldn't it imply in getting a negative beta_0 canceling the positive beta_0 on the right hand side? Thank you so much!
First differencing does eliminate beta_0, but there is an error in my slides. My video on first differencing has the correct equation th-cam.com/video/hQWSh_j3Oy0/w-d-xo.html
Thanks. I have tried to learn how to pronounce Poisson (French name) correctly. Now I need to add Arellano to the list. On my defence, most non-Finns pronounce my last name incorrectly ;)
You are welcome. If you or anyone else is seriously interested in studying with use, you can check out www.jyu.fi/jsbe/en/research/doctoral-school/admission/doctoral-programme-in-business-studies
Thanks, Mikko. This is timely. Please help. I have a panel in which T>N, and I need to run a dynamic panel ( lagged dependent is my explanatory variable), among others. Yes, GMM is not applicable. Can you suggest alternative model.
N=10, and T= 21, balanced. Am I right that basic IV regression with Y(t-2) to control for endogeneity. Dependent var being, Yt, and Y(t-1) is explanatory variable of interest.
Thank you very much--I have a quick question: I am trying to understand the GMM method vs 2SLS IV regression. 1. Does the GMM approach "solve" or help sidestep the issue of when there is no suitable instrument in a model? For example, fiscal decentralization as a regressor is suspect because there is no suitable time variant instrument to address endogeneity with 2SLS. 2. Is GMM a good method in this instance, and how does it help when there is no suitable instrumental variable (that is not a time-lag variable)? Thank you so much for any response!
1) No. GMM and 2SLS are both estimation approaches for the same kind of model. You could 2SLS to implement the Arellano-Bond approach, but the resulting estimator is inefficient. See doi.org/10.1177/1536867X0900900106 2) If our problem is the lack of suitable instruments, GMM is not a solution to that problem. GMM is an estimation technique that can be applied after you have a suitable model. The A-B approach is way of using lags of variables as instruments.
@@mronkko thank you very much , this is helpful! Just to clarify: I understand it is not a "solution", but in the absence of External instruments , is gmm then appropriate e.g. using system gmm when there is persistence of variables, N >t, etc. In other words, is gmm a way to avoid the endogeneity that is present when no external instruments are available?
@@francischoi7191 It really depends on the endogeneity problem. If you can justify the use of differences of past values as instruments, then using then use them. That is the magic ingredient, not GMM. GMM just makes the estimates more efficient than what you would get with 2SLS.
hi! Thanks for the informative video. it helped me a lot. one question : what do you mean by the term "exclusion " at 7:28 while explaining that exclusion assumption must be satisfied? Is the use of word similar to exogenous?
The exlusion restriction means that the lagged variables that you use as instruments should not correlate with the error term. For a general explanation of the exclusion restriction, I recommend the videos on the instrumental variables playlist: th-cam.com/video/mHjBXL2dgLs/w-d-xo.html
Hello teacher, First of all, I would like to thank you for your videos which help us a lot. Please I have a question for you: I would like to estimate a dynamic model by applying the GMM for a number of individuals which is equal to 12 and T=5, is this feasible? If not, is there another method that I can apply for this case? Thanks in advance.
@@mronkko Hello professor, I am currently working on a country with 12 regions. Therefore, I cannot increase the number of individuals. Knowing that I have obtained good results from the tests, with a very low autoregressive parameter of 1.0, and the results completely confirm the economic theory, do you think, Professor, that I should keep these results, or should I use another technique?
@@khalilelbachiri9977 It really depends on your data and research question. If I had 12 regions and T=5, I would probably go for a qualitative multiple case study instead of statistical analysis. With such small sample, it seems unlikely that you get very robust results. If you choose to go with a statistical approach, you should report confidence intervals for all parameters to make sure your readers understand how much uncertainty there is in the results.
An explanation of the difference between the two estimators would indeed be useful. I will add it to my list of things to do. I am currently running a course where I use the videos and this would fit well to what we go over in March with the students.
Hi. I do not have a video about Anderson-Hsiao estimator. But I do use it as an assignment when I teach Arellano-Bond. The A-H is a bit easier to understand and program than A-B.
@@mronkko Hi. Thanks. Can you recommend what estimator should I use when my panel data has N=3 and T=21? I will use lagged variables. I will appreciate your feedback.
Great content. I am replicating Acemoglu et al.2008 Democracyincome table 2, but the Arellano-Bond approach doesn't return the same values. Any clue why is that the case? I got precise results for the other columns library(pder) data(DemocracyIncome) ## Column 4 diff1
For some reason, I have always used Stata when working with dynamic panels. But the plm package seems to do the job www.rdocumentation.org/packages/plm/versions/2.4-3/topics/pgmm
A-B is a fixed effects approach that uses FD to form instruments. I am not sure what you mean with your question. Are you asking if you can estimate an FE model and not use instruments? No, unless your within sample size approaches infinity. Are you asking if there are other approaches for creating instruments? I guess that would be possible with some creativity, but I do not think that any general approaches exist but this would need to be done on a case by case basis.
@@mronkko Thanks a lot for your answer. Okay, I'll think about it. I was wandering, if instead of doing the first diff, we could keep the FE and simply instrument the lagged outcome with the second lagged (without first diff). so it's xtivreg y (l.y=l2.y) x i.t, fe
@@the_causal_mindset That would not work. You need to consider the sources of endogeneity: It is that the error term at t0 contributes to the estimate of a_i because y_i0 is used as a predictor of y_i1. So you need an instrument that is uncorrelated with a_i and the first difference qualifies for that, lagged levels do not. I will need to do another video about dynamic panel bias because understanding that is essential for understanding why the A-B or similar approaches are required.
Dear Prof Mikko,,, i have to admit that this is a great video,,many many thanks. Prof Mikko.i learn a lot from this video.. i am very new in this GMM dynamic panel. i have simple questions which need your enlightenment... question 1 in minute 4:17, you have two arrows directed to ut-1 and yt-1 and you says that it is an endogenous problem. i fully understand on this issue, since within yt-1 itself there is ut-1. since this has endogeneity problem you mention that you need instrumental variable (IV). at this stage, i thought that you want you replace yt-1 with IV so that the IV will no longer correlates with ut-1. but you have mentioned that the IV is actually yt-2 in which it deducts from yt-1 and yt-1 is still there. and i did not hear any info whether yt-1 will be replaced or not, hence yt-1 is still there and still correlates with ut-1. does this endogeneity still exist....?need your enlightenment from prof Mikko,,, question 2 actually question 2 is related to the previous question. at minute 4: 45, you mention that yt-2 can be the IV for the yt-1. my immediate question is that, how would modify the model of (yt-1 - yt-2) + (uit - uit-1). is it gonna be like this--> (yt-2 - yt-3) + (uit - uit-1) is it like this...?if yes, i would agree because uit-1 will no longer correlate with yt-2 ok these are my two stupid question i would really appreciate if prof mikko could enlightened me on this issue or other econometrician shall also help me on this...
I will post a new video about dynamic panel bias in the near future (a week or so) that can be helpful. Question 1: The equation is first-differenced (see the video about first-differncing if you have not done so already). Because of this, you need two time points to calculate the difference (t, t-1). Then you use a third timepoint (t-2) or even further lags as an instrument. Question 2: Instrumental variable estimation does not involve replacing one variable with another one. See the video where I explain 2sls. In that approach, we replace the endogenous regression with a fitted value from a regression of the endogenous variable on the instruments. With other estimation approaches (like GMM) it is a bit more complicated but accomplishes the same purpose.
Dear Mikko, are there are any reference updates you could suggest on these estimators (both Arellano-Bond - ML approach) accounting for other issues for large N large T estimators (e.g Cross-Sectional Dependence, structural breaks )?. I am reviewing updates to account for these issues in my research. Regards
I am not sure if I understand the question. If you are interested in understanding generally what an instrumental variable is, see th-cam.com/play/PL6tc6IBlZmOVIOhIKYNRAuPKUW4u_SrYq.html If you are asking how the instruments are formed in the AB-estimator, you use differences as instruments for the original values and original values as instruments for the differences if you apply the system-GMM technique. I am not sure what you mean by "groups" in this question. Do you mean number of time points in the panel? The number of instruments is often more than the number of time points, what about it?
Correct. That is an error in the slides. The first differencing presentation contains the correct differencing equation th-cam.com/video/hQWSh_j3Oy0/w-d-xo.html
You can use the xtabond2 command by David Roodman. The documentation and the article about the command (linked in the documentation) provide a number of examples.
Very Helpful Videos! Thank you very much. I have spent way too much time researching GMM and GMM estimation methods to now opt for the Maximum Likelihood Approach sadly. I have a question: All the applications of the Difference GMM and Sys-GMM that I have seen were for Panel Data Estimations. Is it possible to use such estimation techniques for Time Series Estimations ? What are the potential problems that I could run into when do so ? Again thank you very much for the videos !!
I am not an expert on time series analyses. My field is management (strategy and entrepreneurship) and we work with short time series (10-30 years at most) and multiple firms in the data. That being said, I do not think that the Arellano-Bond approach can be applied to time series data with just one observed unit (e.g. one company). The reason for this is that the within transformation does not really do anything if you have just one cluster in the data. I encourage you to try running the AB estimator on a time series data to see what happens.
@@mronkko Thank you for the Reply! My goal was that by applying the GMM estimation method to a Dynamic Time-Series Data Set consisting of around 80 Observations, was to control for Reverse Causality and Simultaneity rather than strictly fixed effects. So by creating artificial Instrumental Variables, I was hoping to control for that. (Hopefully that makes sense?). Thank you again for taking the time to answer my question.
@@raiden233 Right. Instrumental variables would be relevant for you then. However, you need to consider that AB estimator is not a general technique for dealing with endogeneity, but addresses the specific problem of dynamic panel bias caused by unobserved heterogeneity. If you have a single time series, this problem would not exist. I explain the problem briefly in the video, and plan to do a more comprehensive explanation in the future.
Hello teacher, First of all, I would like to thank you for your videos which help us a lot. Please I have a question for you: I would like to estimate a dynamic model by applying the GMM for a number of individuals which is equal to 16 and T=10, is this feasible? If not, is there another method that I can apply for this case? Thanks in advance.
The best explanation of this topic I've seen in the internet! Thank you!
Thanks. I have heard "best on TH-cam" before, but best on the Internet, that is something ;)
I am thrilled to have found this channel. I haven't studied econometrics since my first degree and now I'm trying to model a dynamic panel data model and your videos are helping my confidence - thank you! I would appreciate more videos on Stata coding for xtdpdml, too.
You are welcome. I do not do much software demos because I teach multiple software and because, particularly with Stata, the documentation explains the software well. Understanding the conceptual side is more challenging, and once you know that, specifying the command is not very challenging.
you sir have just saved my dissertation. I do not normally comment on videos like this, but wanted to acknowledge your efforts and let you know that you are making an impact! Good work :)
Glad I could help!
@@mronkko @Mikko Rönkkö just a quick question - I have read Roodman (2009) and watched this video a few times.
What are internal and external instruments? Are internal endogenous and external exogenous?
My model tries to look at the impact of CSR on firm performance and takes the following structure: (FirmPerf)it = a + B1(FirmPerf)i(t-1) + B2(CSR)it + B3X'it + e, where X' is a vector of firm-level covariates that are used in similar studies.
I wanted to use system GMM. I have used the lagged dependent variable as my endogenous internal instrument and used the predetermined covariates (X') as external instruments all lagged for two periods as I believe lagging them make them exogenous (as they will not be correlated with the error). Is this along the right track?
@@KiranSingh-vf8nc Internal instrument are a part of the model (p.100 in Roodman, 2009) and external instruments are additional variables that are not of specific interest but are used for causal identification of the model. If you look at any introductory econometrics book, the instrumental variables explained in these books are external instruments. One could say that internal instruments are endogenous to the system, if we consider all time periods. In an equation by equation consideration instruments are always exogenous.
Unfortunately I do not have the capacity to focus on specific research questions in TH-cam comments. These always require a bit of thinking.
Hello, I normally do not comment on TH-cam, but your explanation is very clear and precise, extremely helpful for me as a non statistical student, thank you!
You're very welcome!
Thanks mikko clear and concise. Everyone loves to make econometrics explanations complicated so thanks for breaking that trend!
You are welcome. I can relate to your pain with complicated explanations ;)
Dear Miko,
It’s a privilege to hear your lecture on the Arrelano Bond approach to dynamic panel models.
You’ve made a difficult concept much easier to understand and apply. Thank you for putting this video on youtube.
However, I have a question on the Arellano-Bond test to check whether the errors in the first differencing model are correlated
Arellano and Bond in their diagnostic testing perform the following:
AR(1) test || where H0 states “No autocorrelation of order 1”
AR(2) test || where H0 states “No autocorrelation of order 2”
AR(3) test || where H0 states “No autocorrelation of order 3”
From my understanding to ensure the first differencing model is correct and the coefficients can be interpreted appropriately (i.e. the model is not misspecified), you MUST be able to:
Reject H0 for AR(1) (i.e., p0.10): i.e., where H0: There is no second-order autocorrelation of error terms
Accept H0 (i.e. Fail to Reject H0) for AR(3) (i.e., p>0.10): i.e., where H0: There is no third-order autocorrelation of error terms
MY QUESTION
Why MUST WE ACCEPT H0 for AR(1)?
i.e. Why do we want to conclude that there is first-order autocorrelation of the error terms in the first differencing model Δut & Δut-1 ?
My answer (which I believe is correct - but perhaps it is not and I'm going down a deep rabbit hole) is:
1. It is expected that there will be some degree of first-order autocorrelation in the errors for the first differencing model.
2. When you include the first instrumental variable (yt-2) and sequentially other instrumental variables (yt-3, ….etc) to the first differencing model to help remove endogeneity, effectively we want to ensure that there is no second-order autocorrelation in the errors AR(2) or third order autocorrelation in the errors AR(3) .
Why?
If we were to identify autocorrelation in the error terms (i.e. by Accepting H0) in the second or third order, THEN it signals that we have omitted variables and this invalidates any correct interpretation of our estimated coefficients.
The testing from my understanding
For AR(1), I assume it tests the correlation between Δut & Δut-1
For AR(2), I assume it tests the correlation between Δut & Δut-1 + Δut-2
For AR(3), I assume it tests the correlation between Δut & Δut-1 , Δut-2 & Δut-3
An explanation that I found on the internet as to why we must Accept H0 for AR(1)
Note: As a novice, I don’t seem to understand this explanation though
“The aim of the Arellano-Bond tests is to check whether the idiosyncratic error term is serially correlated. The test is conducted for the first-differenced errors. If the error term in levels is serially uncorrelated, this implies that the error term in first differences has a negative first-order serial correlation (with a correlation coefficient of -0.5) but no second-order or higher-order serial correlation. Thus, we should reject the null hypothesis of no first-order serial correlation in first differences (AR(1) test) but should not reject the null hypothesis of no higher-order serial correlation in first differences (AR(2), AR(3), ...).
If you do not reject the null hypothesis of the AR(1) test, this could indicate that your idiosyncratic error term in levels is highly serially correlated. In the extreme case, the error term in levels follows a random walk such that the first-differenced errors are serially uncorrelated. Such a situation would indeed invalidate the MSM.”
Any help you can provide me in addressing my question would be super appreciated.
Warm regards,
Reuben
Well-structured question!
The AB estimator assumes that the error term u (after first-differencing) is NOT serially correlated. The H0 in the autocorrelation tests is that there is no autocorrelation. Ideally, you would fail to reject H0 for all lags. However, often we find that u is serially correlated for short lags. When this happens, we need to increase the lags of the instruments to make sure that they satisfy the exclusion criterion.
One way to understand this - and this was how I learned to understand it - is to look a the dynamic panel model in wide format data (see the last stlide in the talk) and then add first-order autocorrelation between the errors and think what happens to the estimates.
@@mronkko
Hi Miko,
Thanks for the super quick response. I really appreciate it!
Your response reassured me that I am on the correct track of getting more understanding of this material.
I understand that the STATA command: *xtabond2* is a bit like a “black box” and spits out results, but I don’t believe too many people that apply it fully understand the mechanics behind how the output is generated [though Roodman(2009) who programmed the xtabond2 code has provided a thorough paper on how it works with examples)
I agree with you that the GMM estimator was cutting edge (30 years ago); and I do wonder (perhaps like yourself) why more researchers don’t use the maximum likelihood estimator (ML estimator) that can be applied with the STATA command: *xtdpdml*. The ML estimator has the benefit of handling missing values, which I believe the GMM estimator does not? I could be incorrect in my knowledge.
I believe researchers use the STATA command: *xtabond2* because it’s (1) relatively easy to apply without deep understanding; (2) most reviewers of journal articles also might have a preliminary understanding of dynamic GMM, but may not have an understanding of the STATA command *xtdpdml* and most researchers/reviewers don’t know too much about cross-lagged panel models in my field (finance). That said, there is also a lot of criticism of cross-lagged panel models as well in the literature.
That said, I wonder how well any of these STATA commands (i.e. *xtabond2* or *xtdpdml*) handle missing values in the panel dataset. Perhaps *xtdpdml* handles missing values because it uses a ML estimator, which I understand can do so. I suspect the STATA command *xtabond2* deletes observations through a listwise deletion; would the possible result that it can lead to biased results (especially if you have a lot of missing values in your panel dataset)
For your information (though I’m sure you are aware), there is a new STATA command that suggests the *xtabond2* has errors, which is *xtdpdgmm*
(see www.kripfganz.de/stata/xtdpdgmm.html)
I shall keep on watching your videos, as they are super helpful.
Thanks again and stay well,
R
p.s. sorry for this lengthy reply. I just got carried away
@@reuben_1973 No worries, I tend to get carried away too quite often. I agree with your interpretation that many researchers use AB estimator as implemented in Stata as a black box. I also think that this is because of modeling their studies based on past research. That being said, I do not think that AB is entirely outdated because GMM is a bit more flexible in its assumptions compared to ML. Whether this makes a difference is another matter.
I do not think that neither of the commands that you mention deal with missing data. If you want to have an estimator that can use observations where just a part of the variables are measured, you can do that by specifying the model using wide format data and applying sem command in Stata.
I simply love you, you saved my life and my grades. THANK YOU
You are welcome.
Found this at the exact moment, thanks Mikko! You helped me a lot
Thank you very much. Was on time for this explanation We hope to communicate with you 🇩🇿
You are welcome.
Thanks a lot Mikko, this was really helpful for my econometrics assignment!
You are welcome!
You are amazing Mikko
Thanks.
Bookmarking video, thanks
You are welcome!
Sir, thank you so much for this video. You are amazing.
I'd like to ask a trivial question, if you (or anyone who sees this) could answer me: when differentiating, why doesn't beta_0 get canceled as well? I mean, when we subtract y_t-1 on both sides, wouldn't it imply in getting a negative beta_0 canceling the positive beta_0 on the right hand side? Thank you so much!
First differencing does eliminate beta_0, but there is an error in my slides. My video on first differencing has the correct equation th-cam.com/video/hQWSh_j3Oy0/w-d-xo.html
Excellent video, just as a side note. Arellano is an Spanish name, so it's pronounced like "Areyano". Anyway good information was useful for me
Thanks. I have tried to learn how to pronounce Poisson (French name) correctly. Now I need to add Arellano to the list.
On my defence, most non-Finns pronounce my last name incorrectly ;)
I wish to be one of your students. very clear, thank you
You are welcome. If you or anyone else is seriously interested in studying with use, you can check out www.jyu.fi/jsbe/en/research/doctoral-school/admission/doctoral-programme-in-business-studies
Very clear, thank you.
You are welcome.
Thank you very much prof
This is a god send
Good that you liked It. I should do another one on the diagnostics of the A-B approach.
Thank you Prof
Welcome!!
Hi. What do you think about stationarity when dealing with GMM. Does it matter if you use system gmm even if data is not stationary?
System GMM requires mean stationarity. See doi.org/10.1080/00036846.2018.1540854
Thanks, Mikko. This is timely. Please help. I have a panel in which T>N, and I need to run a dynamic panel ( lagged dependent is my explanatory variable), among others. Yes, GMM is not applicable. Can you suggest alternative model.
N=10, and T= 21, balanced. Am I right that basic IV regression with Y(t-2) to control for endogeneity. Dependent var being, Yt, and Y(t-1) is explanatory variable of interest.
@@mronkko Thanks, much for the wonderful feedback. Yes, about the data you are right. Will see what I can do.
may I know the feedback? I also face same problem T>N,
T=20
N=13
Thank you very much--I have a quick question: I am trying to understand the GMM method vs 2SLS IV regression.
1. Does the GMM approach "solve" or help sidestep the issue of when there is no suitable instrument in a model? For example, fiscal decentralization as a regressor is suspect because there is no suitable time variant instrument to address endogeneity with 2SLS.
2. Is GMM a good method in this instance, and how does it help when there is no suitable instrumental variable (that is not a time-lag variable)?
Thank you so much for any response!
1) No. GMM and 2SLS are both estimation approaches for the same kind of model. You could 2SLS to implement the Arellano-Bond approach, but the resulting estimator is inefficient. See doi.org/10.1177/1536867X0900900106 2) If our problem is the lack of suitable instruments, GMM is not a solution to that problem. GMM is an estimation technique that can be applied after you have a suitable model. The A-B approach is way of using lags of variables as instruments.
@@mronkko thank you very much , this is helpful! Just to clarify: I understand it is not a "solution", but in the absence of External instruments , is gmm then appropriate e.g. using system gmm when there is persistence of variables, N >t, etc. In other words, is gmm a way to avoid the endogeneity that is present when no external instruments are available?
@@francischoi7191 It really depends on the endogeneity problem. If you can justify the use of differences of past values as instruments, then using then use them. That is the magic ingredient, not GMM. GMM just makes the estimates more efficient than what you would get with 2SLS.
@@mronkko thank you kindly
What estimation technique should I use under the presence of cross sectional dependency and unit root?
hi! Thanks for the informative video. it helped me a lot. one question : what do you mean by the term "exclusion " at 7:28 while explaining that exclusion assumption must be satisfied? Is the use of word similar to exogenous?
The exlusion restriction means that the lagged variables that you use as instruments should not correlate with the error term.
For a general explanation of the exclusion restriction, I recommend the videos on the instrumental variables playlist:
th-cam.com/video/mHjBXL2dgLs/w-d-xo.html
Hello teacher,
First of all, I would like to thank you for your videos which help us a lot. Please I have a question for you: I would like to estimate a dynamic model by applying the GMM for a number of individuals which is equal to 12 and T=5, is this feasible? If not, is there another method that I can apply for this case? Thanks in advance.
Your sample size seem inadequate. I would recommend collecting data from more individuals.
@@mronkko
Hello professor,
I am currently working on a country with 12 regions. Therefore, I cannot increase the number of individuals. Knowing that I have obtained good results from the tests, with a very low autoregressive parameter of 1.0, and the results completely confirm the economic theory, do you think, Professor, that I should keep these results, or should I use another technique?
@@khalilelbachiri9977 It really depends on your data and research question. If I had 12 regions and T=5, I would probably go for a qualitative multiple case study instead of statistical analysis. With such small sample, it seems unlikely that you get very robust results. If you choose to go with a statistical approach, you should report confidence intervals for all parameters to make sure your readers understand how much uncertainty there is in the results.
Tx Sir
You are welcome
Thank you very much for the video! Could you explain maybe explain the difference between one step and two step GMM?
An explanation of the difference between the two estimators would indeed be useful. I will add it to my list of things to do. I am currently running a course where I use the videos and this would fit well to what we go over in March with the students.
Thank you for explanation. Do you have a video about Anderson-Hsiao estimator?
Hi. I do not have a video about Anderson-Hsiao estimator. But I do use it as an assignment when I teach Arellano-Bond. The A-H is a bit easier to understand and program than A-B.
@@mronkko Hi. Thanks. Can you recommend what estimator should I use when my panel data has N=3 and T=21? I will use lagged variables. I will appreciate your feedback.
Great content. I am replicating Acemoglu et al.2008 Democracyincome table 2, but the Arellano-Bond approach doesn't return the same values.
Any clue why is that the case? I got precise results for the other columns
library(pder)
data(DemocracyIncome)
## Column 4
diff1
I would suggest contacting the authors to see why your replication fails. These are always a bit tricky to troubleshoot.
Thanks for this informative video. By any chance, is there a reliable way to estimate this in the R program?
For some reason, I have always used Stata when working with dynamic panels. But the plm package seems to do the job www.rdocumentation.org/packages/plm/versions/2.4-3/topics/pgmm
Awesome vid! What if we have a fixed effect model to capture a_i? Is it necessary to use FD?
A-B is a fixed effects approach that uses FD to form instruments. I am not sure what you mean with your question. Are you asking if you can estimate an FE model and not use instruments? No, unless your within sample size approaches infinity. Are you asking if there are other approaches for creating instruments? I guess that would be possible with some creativity, but I do not think that any general approaches exist but this would need to be done on a case by case basis.
@@mronkko Thanks a lot for your answer. Okay, I'll think about it. I was wandering, if instead of doing the first diff, we could keep the FE and simply instrument the lagged outcome with the second lagged (without first diff). so it's xtivreg y (l.y=l2.y) x i.t, fe
@@the_causal_mindset That would not work. You need to consider the sources of endogeneity: It is that the error term at t0 contributes to the estimate of a_i because y_i0 is used as a predictor of y_i1. So you need an instrument that is uncorrelated with a_i and the first difference qualifies for that, lagged levels do not.
I will need to do another video about dynamic panel bias because understanding that is essential for understanding why the A-B or similar approaches are required.
Dear Prof Mikko,,,
i have to admit that this is a great video,,many many thanks. Prof Mikko.i learn a lot from this video.. i am very new in this GMM dynamic panel. i have simple questions which need your enlightenment...
question 1
in minute 4:17, you have two arrows directed to ut-1 and yt-1 and you says that it is an endogenous problem. i fully understand on this issue, since within yt-1 itself there is ut-1. since this has endogeneity problem you mention that you need instrumental variable (IV).
at this stage, i thought that you want you replace yt-1 with IV so that the IV will no longer correlates with ut-1. but you have mentioned that the IV is actually yt-2 in which it deducts from yt-1 and yt-1 is still there.
and i did not hear any info whether yt-1 will be replaced or not, hence yt-1 is still there and still correlates with ut-1. does this endogeneity still exist....?need your enlightenment from prof Mikko,,,
question 2
actually question 2 is related to the previous question.
at minute 4: 45, you mention that yt-2 can be the IV for the yt-1. my immediate question is that, how would modify the model of
(yt-1 - yt-2) + (uit - uit-1).
is it gonna be like this-->
(yt-2 - yt-3) + (uit - uit-1)
is it like this...?if yes, i would agree because uit-1 will no longer correlate with yt-2
ok these are my two stupid question i would really appreciate if prof mikko could enlightened me on this issue or other econometrician shall also help me on this...
I will post a new video about dynamic panel bias in the near future (a week or so) that can be helpful.
Question 1: The equation is first-differenced (see the video about first-differncing if you have not done so already). Because of this, you need two time points to calculate the difference (t, t-1). Then you use a third timepoint (t-2) or even further lags as an instrument.
Question 2: Instrumental variable estimation does not involve replacing one variable with another one. See the video where I explain 2sls. In that approach, we replace the endogenous regression with a fitted value from a regression of the endogenous variable on the instruments. With other estimation approaches (like GMM) it is a bit more complicated but accomplishes the same purpose.
@@mronkko thank you prof,,i will look into your video...anyway i have subscribed your channel
Dear Mikko, are there are any reference updates you could suggest on these estimators (both Arellano-Bond - ML approach) accounting for other issues for large N large T estimators (e.g Cross-Sectional Dependence, structural breaks )?. I am reviewing updates to account for these issues in my research. Regards
I plan to do a couple of things in the near future, 1-2 weeks.
@@mronkko Thanks for your concern,looking forward to that ,cheers!
WHAT ARE INSTRUMENTS? AND WHAT IF NUMBER OF INSTRUMENTS > NUMBER OF GROUP IN TWO STEP DYNAMIC GMM MODEL
I am not sure if I understand the question.
If you are interested in understanding generally what an instrumental variable is, see th-cam.com/play/PL6tc6IBlZmOVIOhIKYNRAuPKUW4u_SrYq.html
If you are asking how the instruments are formed in the AB-estimator, you use differences as instruments for the original values and original values as instruments for the differences if you apply the system-GMM technique.
I am not sure what you mean by "groups" in this question. Do you mean number of time points in the panel? The number of instruments is often more than the number of time points, what about it?
3:49 there shouln't be B0 in the difference, right?
Correct. That is an error in the slides. The first differencing presentation contains the correct differencing equation th-cam.com/video/hQWSh_j3Oy0/w-d-xo.html
Kindly show System GMM with stata commands, so that we the researchers can able to perform this test.
You can use the xtabond2 command by David Roodman. The documentation and the article about the command (linked in the documentation) provide a number of examples.
Very Helpful Videos! Thank you very much.
I have spent way too much time researching GMM and GMM estimation methods to now opt for the Maximum Likelihood Approach sadly.
I have a question: All the applications of the Difference GMM and Sys-GMM that I have seen were for Panel Data Estimations. Is it possible to use such estimation techniques for Time Series Estimations ? What are the potential problems that I could run into when do so ?
Again thank you very much for the videos !!
I am not an expert on time series analyses. My field is management (strategy and entrepreneurship) and we work with short time series (10-30 years at most) and multiple firms in the data.
That being said, I do not think that the Arellano-Bond approach can be applied to time series data with just one observed unit (e.g. one company). The reason for this is that the within transformation does not really do anything if you have just one cluster in the data. I encourage you to try running the AB estimator on a time series data to see what happens.
@@mronkko Thank you for the Reply!
My goal was that by applying the GMM estimation method to a Dynamic Time-Series Data Set consisting of around 80 Observations, was to control for Reverse Causality and Simultaneity rather than strictly fixed effects. So by creating artificial Instrumental Variables, I was hoping to control for that. (Hopefully that makes sense?).
Thank you again for taking the time to answer my question.
@@raiden233 Right. Instrumental variables would be relevant for you then. However, you need to consider that AB estimator is not a general technique for dealing with endogeneity, but addresses the specific problem of dynamic panel bias caused by unobserved heterogeneity. If you have a single time series, this problem would not exist. I explain the problem briefly in the video, and plan to do a more comprehensive explanation in the future.
Hello teacher,
First of all, I would like to thank you for your videos which help us a lot. Please I have a question for you: I would like to estimate a dynamic model by applying the GMM for a number of individuals which is equal to 16 and T=10, is this feasible? If not, is there another method that I can apply for this case? Thanks in advance.
I would go for GMM or ML (see the citation to the Allison paper toward the end of the presentation).