Open jmausolf opened 4 years ago
Thank you for your work! Have you tried exploring more complex MCMC methods such as Reimannian Manifold HMC (https://arxiv.org/abs/0907.1100)? It works especially well when dealing with high dimensional data, which I'd imagine you'd come across in comparative politics.
Thank you in advance for the presentation! I'm wondering about if there's any limitation of the model you discussed in the paper?
Thanks for your representation! Like many questions proposed by my classmates, I am wondering what is the specific advantage of hdpGLM method compared with other well-known Bayesian methods such as MCMC that deal with drawbacks of multivariate linear models as well?
Thanks for the presentation. Although it touches down some modelling areas that might sound difficult to understand for me, can you kindly elaborate more about its applications in political science (or political behavior especially)? I am wondering how to apply these highly technical model to the real social science? How this model can help us better understand certain kind of political behavior (i.e. collective action)? Also, I am very looking forward to your Political Behavior course next quarter. Thanks.
Thanks for your presentation in advance and I am curious about the methods: I am wondering if your model is conducting kind of feature engineering to cluster the observations based on 'that' variable. It seems that it can wonderfully solve the Simpson's Paradox once we distinguish the latent variables that varies so largely across subgroups. Once we distinguish the context-based latent variables, we can make better interpretation about the phenomenon happening in different subgroups. My question is, if there is a more direct interpretation of how this model is able to use statistics to tell there is a context-base latent effect. And how can a social scientist easily pick up this exciting method on his/her researches?
It's always so exciting to have you as our speaker. I see most of my classmates have posted their questions, with mine included in Li Liu's comment, and I look forward to your presentation. Sorry if this is unrelated, but I'm recently usually at a loss at whether I should use Python or R to solve a specific problem, and I'd like to hear about your insights or experiences regarding this.
Thank you for your presentation in advance! I am wondering whether one of the advantages of the hdpGLM model compared to traditional models is that the hdpGLM could identify clusters with heterogeneity even if the division of clusters is unobservable.
I am very impressed by the detailed exposition in your paper, and learned so much about empirical analysis concerning latent variable. Here are two questions: 1. Is the number of clusters a hyperparameter just like in EM algorithms? 2. With the clustering obtained, how can we conclude what features this grouping is based on? For example, the data of voters in the US can be grouped into three clusters, but how do we extract the "division line" for each cluster? Thanks!
Thank you so much for your presentation! I am wondering how can you identify the covariates and decide on what factors should be considered in the model? Could you elaborate on the strengths of the hdpGLM model versus the ordinary GLM model?
Thanks for your presentation! The latent heterogeneity within political science research is worth discussion. You have introduced several different models in the paper, so I was wondering if you could talk more about the model selection, say, are there certain research questions that hdpGLM would have better performance over other models? Thank you!
Thank you for your presentation. I am curious how this model could be extended to other characteristics, especially some individual personalities.
Thank you for presenting your innovative way of using GLMs stacked on this additional layer of modeling. Exploring the assumptions of the hdpGLM led me down the rabbithole of trying to understand the mathematical motivation behind using this to model heterogeneity.
Can you comment on the intuition behind the method?
Thank you for this presentation. The method introduced in this paper captures the latent heterogeneity of covariates in GLM models. As the algorithm turns out to be somewhat complicated to code up, is there any tool provided for researchers to implement this identification strategy into research in practice?
It is quite exciting to see some research about the theories of computational methods in social science. My interest is why we need to use a Dirichlet process to build up the new framework and also is it possible to compare/combine this model with other econometric models like FEM(fixed-effect model) or REM(random effect model) also deal with the heterogeneity in panel data, to make a better explanation of the variance crossing regions or time periods?
Thanks for sharing your research! One question I had was about policy implementation. Once the technical work of constructing new methods is done, what advice do you have for facilitating its adoption in policy? More broadly, how can we, as computational social scientists, best translate our work in interdisciplinary fields.
Thanks for sharing your research! One question I had was about policy implementation. Once the technical work of constructing new methods is done, what advice do you have for facilitating its adoption in policy? More broadly, how can we, as computational social scientists, best translate our work in interdisciplinary fields?
Thanks for sharing your research! One question I had was about policy implementation. Once the technical work of constructing new methods is done, what advice do you have for facilitating its adoption in policy? More broadly, how can we, as computational social scientists, best translate our work in interdisciplinary fields?
Thanks in advance for your presentation. Could you please share your thoughts on the shortcomings of techniques to handle latent heterogeneity that have been developed in other fields? What are the specific demands of political science that make hdpGLM the best approach for political science applications?
Thank you for presenting your interesting research! I am wondering whether the hierarchical Dirichlet process of generalized linear models that you proposed can also use in other social science fields? Would it also be able to solve the Simpson's paradox problems?
Thank you for presenting! Is there any way to incorporate unsupervised learning to better estimate the number of clusters?
Thank you for presenting! This research is very interesting! I’m wondering what obstacles are there if it is applied in real life?
Thanks a lot for your presentation! Honestly speaking, I am stuck with a lot of professional and abstract terms in your paper, I hope you can give us some examples to explain it clearly! Thanks! Also, I am really excited to see some applications.
Comment below with questions or thoughts about the reading for this week's workshop.
Please make your comments by Wednesday 11:59 PM, and upvote at least five of your peers' comments on Thursday prior to the workshop. You need to use 'thumbs-up' for your reactions to count towards 'top comments,' but you can use other emojis on top of the thumbs up.