Thank you so much for sharing your knowledge with the world, Prof. Burkey. You helped me to get a good overview of this topic and will allow me to go forward in a more organized manner with finishing my master's thesis, in which I am going to implement a spatial econometric model to analyze the behavior of regional poverty. Thank you again and greetings from Colombia.
Thank you! I'm doing a research about the determinants of local public expenditures on education in Peru with a territorial approach; your content is very useful!
I really like this course! I hope it helps me I'll level up my Spatial analysis skills. Thank you very much! After I watched this video I have got a lot of intuitional understanding of complex issues.
Best line of the video - "There are some people who have a hatred of spatial econometrics" ahaha. Thank you very much Dr. Burkey for getting me pumped up for this course with this awesome introductory video.
I have been trying to learn Spatial Analysis to learn the relationship between police station locations and specific Points of interest. Do you think this matter is a part of Spatial Econometrics or something else? Thank you for making this series. I have been drowning in tons of materials and your videos just help me a really great direction!
Dear Sir, Thank you very much for the helpful and concise explanation. Could you please advise me what is the best software for spatial econometrics in economics? I am familiar with Stata and Python only.
I prefer R (with spdep package), but the other best option is Matlab (with Jim LeSage's spatial econometrics toolbox). I am working on an R guide right now, and will release a video soon!
Thanks for the video, the thing I didnt understand is in SEM model if you if we leave out same pWXbs, remains y= Xb+u; u=pWe+e; in which you said u=lambdaWu+e; can u please explain that?
No, I haven't gotten into that side of things, which is more of an approach used by Spatial Statistics than Spatial Econometrics. I have Chun and Griffith's "Spatial Statistics and Geostatistics", but haven't taken the time to work through it. But my overall impression (might not be a great one) is that while Spatial Stats people see spatial interactions as a "nuisance" they would like to control for or filter out, in Spatial Econometrics we see the spatial relationships as a very interesting part of the problem we'd like to understand, not just get rid of. I'll move the book from the bookshelf to my desk, and see if that helps motivate me to open it! :)
Actually what the eigenvector spatial filtering approach does is remove the spatial patterns in the residuals and treat them as additional independent variables, so they are not gotten rid off. Some of the papers I have read from Chun and Griffith show that this approach produces relatively more efficient and consistent regression estimates. And are also unbiased. Your video explanations really make spatial econometrics look easy and I love them so much, so I'll be looking forward to it. :)
Right- these patterns are not "thrown away", and of course filter out the noise to get better estimates than you would with OLS or ML ignoring the spatial interaction. But (as I understand it, which isn't very well) the spatial variables thus created can't be interpreted in a useful way, as you can with the direct/indirect/total effects framework- where you can get a good feel for the average magnitudes of the spatial interactions with regard to explanatory variables. An analogy in my mind would be the difference between correcting for heteroskedasticity in your model because you work to understand the structure of the heteroskedasticity versus using a White/Eicker/Huber sandwich correction. They both "work", but the former is more satisfying, model-driven, and informative, whereas the latter "filters" out the effects of the heteroskedasticity. Perhaps a strained analogy. But again, at the moment I am speaking forma position of extreme ignorance.
I see what you are saying Dr. Burkey. Kind of like choosing a dynamic panel estimation over a robust standard errors estimation but at the same time your lagged DV is not a relevant predictor variable. I'm now wondering whether if you interact the selected eigenvectors with the relevant explanatory variables you can make meaning out of them but I'll have to read more about it. Thank you.
It looks like one should be mathematically sophisticated in order to make good use of these models e.g., the hierarchical model you mentioned towards the end. Do you mind me asking how far into mathematics did you have to for your Economics Ph.D? I know that real analysis is typically required, but what else did you take beyond that..? How would you feel about someone implementing one of the more "basic" models, say the spatial Durbin models for exploratory/investigative purposes at a health department, if they had some statistics training before and after considering diagnostics/remedials and other things? You might recognize me from another video comment :) Trying to avoid p-hacking by doing good statistics... but statistics is not really my job :( Appreciate your work!
The more math and statistics you know, the better! Statistics is truly a lifetime full of constant learning, always with more to learn. After getting my undergraduate degree in economics, I spent another 1.5 years in school studying more math-- I was 1 class short of getting a math degree, but that one course was French 4 ☺. I really wish that I had gone even further, perhaps getting a master's degree in math before starting my Ph.D. I have been taking some classes toward a MA in the past couple of years. At least in my Ph.D. program at Duke, you are constantly struggling though very advanced mathematical concepts, using systems of differential equations, proofs, very advanced calculus, matrix calculus, and very advanced topics in statistics. Most people are underprepared, and get overwhelmed by the math. It would have been better if I could have focused more time on the ECONOMICS, rather than struggling with math as we all did in my Ph.D. program, because the goal in a Ph.D. program is not only to be able to understand exon and stats, but to develop NEW econ and stats theroies and tools. However, many people are able to do competent applied statistics without advanced math degrees-- the most important requirement is to TRY and CARE and do your best. In other words, think about what you are doing, try to do it correctly, and keep learning!
when creating a weighted spatial matrix to determine neighbors; I was wondering if it's possible to have difference measures in determining neighbors among different variables. Such as, I have a lot of zip code and county data, so will probably use a Hierarchical model to nest them and would consider queen rules for neighbors. However, for certain businesses, I want to find the closest 3 competitors and consider them neighbors. Some maybe be just down the block and others maybe 3 counties over. Is it possible to do this? thanks!
To show you (a simplified version of) how this equation simplifies to a SEM. Typically we think of Y=XB+e as being the standard linear model, so I was making a substitution for Y. As I said in the video, while the manipulations I am doing here are not technically correct, they can help give us some intuition about how these extra spatial terms behave, and especially about why the restriction theta=-pB can show that one can simplify a Durbin into a .spatial error model. This is important because showing that the Durbin and the Error models are NESTED allows us to do an LR test to see which model is more statistically appropriate.
What do you mean by "tree-based models"? I am only familiar with this term from machine learning classification algorithms. I don't see how this relates, unless I am misunderstanding the point.