I'm restarting open Econometrics Office Hours this week, now from
@BowdoinCollege
, where I am visiting this year.
Open to anyone anywhere needing econometrics help! Now Wednesdays 10-12 AM Eastern or by appointment.
Trying something new: starting next Wednesday, I will be hosting weekly Econometrics office hours free and open to anyone in the world with Econometrics questions.
Sign up at
Updated notes for my ongoing Causal Econometrics class are now online, so far covering Causal Frameworks, Experiments, Adjustment, Semiparametrics, and IV.
Please yell at me if you see anything you disagree with.
Starting tomorrow: teaching a new 1/2 semester PhD class: "Causal Econometrics" (CMU 47-873).
Somebody told me this "causality" thing might be useful.
Syllabus with prospective plans and readings: Comments welcome!
I'm making my Computational Methods for Economics class repo public this year.
It's mostly classical numerical analysis and closely follows the Judd textbook and notes, but with a few updates and examples, so some may find it useful.
Kernel methods provide a simple and robust solution to high dimensional (conditional) distribution estimation problems like those that show up in causal inference with continuous variables. This talk gave a compelling and comprehensible intro and how-to:
Course notes for my class Econometrics II, including causal inference (potential outcomes and causal graphs), panel data, nonlinear estimation, and time series, with
#rstats
code.
Updated notes for my ongoing Causal Econometrics class are now online, so far covering Causal Frameworks, Experiments, Adjustment, Semiparametrics, and IV.
Please yell at me if you see anything you disagree with.
This very long awaited work on the statistical properties of macroeconometric models is finally out.
I believe the content of the critique deserves to be taken seriously by those of us working in the area.
The Laplacian is ubiquitous in financial and economic mathematics, where it represents evolution under continuous uncertainty. Geometric perspectives can explain its origin and relate its features to those of the underlying domain, and facilitate computation.
Just uploaded a video on the Laplacian:
If you've heard of the Laplacian and always wondered what it was, or feel like you don't have great intuition for what it means, take a look—we explore how the Laplacian shows up in geometry, physics, & computation.
This project introduces new, easy-to-use tools for Bayesian estimation of nonlinear dynamic economic models.
Try them out with DifferentiableStateSpaceModels.jl! 1/🧵
Hamiltonian Monte Carlo is a great approach to estimating dynamic models in economics: open-source libraries to do so, from
@DonskerClass
, Jesús Fernández-Villaverde, Jesse Perla,
@ChrisRackauckas
, and Peifan Wu
Where frequentism really excels is at easy semiparametrics: the intersection of "I don't know" and "I don't care." You can focus on the things you need to get right like treatment effects and be robust to nuisances like error forms or serial correlation 🧵
But those estimators can be studied from a frequentist perspective, and no strong assumptions are needed.
My hesitation with Bayesian methods -- when they differ from classical ones -- is that they are not "robust" in the econometrics sense.
I closed out my Causal Econometrics class with a grab bag of lectures on Outcome Heterogeneity
Decision-making and Optimal Policy
and External Validity and Sample Selection
Updated notes for my ongoing Causal Econometrics class are now online, so far covering Causal Frameworks, Experiments, Adjustment, Semiparametrics, and IV.
Please yell at me if you see anything you disagree with.
Notes from my class, Forecasting for Economics and Business, covering statistical, Bayesian, online, and machine learning methods for forecasting and time series, with applications and code in
#rstats
@instrumenthull
This can also give big computational advantages for 2 step estimates: you can run the (expensive) step 1 once and plug in, then use reweighted draws of the same influence function, instead of doing both steps in every bootstrap resample. See Section 3 of
Had a lovely visit to
@Brown_Economics
, where
@instrumenthull
was the first person ever to ask about the profile pic. It is of course a realization of a Karhunen-Loeve decomposition of a Brownian bridge, the limit distribution implied by Donsker’s theorem.
Week 1 of Econometrics Office Hours went well. Thanks to everyone who signed up or asked questions!
Sessions continue next week, same time.
Sign up at or by form at
Trying something new: starting next Wednesday, I will be hosting weekly Econometrics office hours free and open to anyone in the world with Econometrics questions.
Sign up at
Amazing resource! Explains how to derive a nonparametric influence function estimator (eg for DoubleML, TMLE, etc) using basic calculus without assuming you know the jargon or skipping over the key finicky bits of algebra.
Modern statistical methods often use estimators based on efficient influence functions. See our latest tutorial paper on the subject with Oliver Dukes,
@karlado
@SVansteelandt
2021 updated notes from my class Forecasting for Economics and Business, covering statistical, Bayesian, online, and machine learning forecasting methods with code in
#RStats
Lots to say about this (fascinating!) work, but I'll start with: it's one of the only critiques of "how did economics get like that" to seriously engage with what most economists actually did in policy (at least up to ~1985), and not just cover a few unrepresentative celebrities.
This was a fun project combining the Generalized Method of Moments with online learning to choose sequentially from multiple data sources to estimate a parameter.
@shantanug7
and
@zacharylipton
taught me a lot about machine learning in the process.
Congrats
@shantanug7
(& co-advisor
@DonskerClass
) on
#neurips2021
spotlight! While data fusion problems are old hat (esp in econ/causal), the available data are always given. Here, we iteratively decide what source to query to converge on our parameter.
Important new paper (by Auclert, Bardóczy, Rognlie & Straub) on solution of linearized heterogeneous agent methods based on the sequence space (Wold or MA(∞)) representation (1)
@nickchk
Aside from recent DiD papers (on their way to wide use):
-Formal mediation analysis: claims about mechanisms are causal and a well identified main effect is not enough.
-For data exploration, quantile or distributional random forests. Conditional distributions tell you a lot!
@lewbel
@PhilHaile
@yudapearl
@steventberry
@VC31415
I have a paper defining DSGE structure in terms of DAGs: see Cdn 3. Auclert et al do likewise, and have pictures: eg in both cases, per "GE", this a computational graph of equilibrium conditions: the full equilibrium is not acyclic.
Economists should be aware of recent remarkable work by Alon et al showing equivalence between Private PAC learnability and finite Littlestone dimension. In words, this says precisely what functions will be consistently estimable from private census data.
After jazz and beignets, the highlight of my
#neurips22
was learning from Masa Uehara about the deep connections between causal inference and policy evaluation in reinforcement learning.
He now has a "Rosetta Stone" survey paper about it, w/ Shi & Kallus
@zacharylipton
Interdisciplinary reading group? Cosma teaches learning theory,
@KevinZollman
teaches epistemology, I teach... one handwavy lecture rushing through Hume, Bayes, Wald, and Wolpert before spending the rest of a semester on basic R. Others might be interested too.
My MCMC is running beautifully today, after some tweaking, but I love the idea of some journalist just deciding that some perspective on an issue should be ignored because otherwise the log Sobolev constant for their sampler would to too large to make the deadline.
Question about common pattern of persistent forecast revisions below (ht
@Claudia_Sahm
): how can its' ubiquity be rationalized, given that Bayes forecasts should be martingales (revisions mean 0)? 3 conjectures (please chime in with others):
Frequentist guarantees form a "contract" with the data analyst, providing that if you do certain things to your data, your analysis will have certain properties. But real data analysis inevitably involves other steps, which can bork traditional guarantees.
Retweeting to promote publicity for demonstrations of methods working poorly outside of their ideal domain of applicability and showcase examples in papers. Performance issues are often known to practitioners (and implicit in papers' guarantees), but passed on mainly as folklore.
Since my grad students told me they read my tweets, I'd like to mention that the Simons Institute videos on Learning & Games are a great use of time for anyone interested in computation and estimation in econ, especially the boot camp to get up to speed.
@VC31415
I use ggdag along with dagitty for DAG visualization in R. In Python, see dowhy as well as and (in beta, ping
@eliasbareinboim
for previews) for more estimation and identification tools.
Updated notes for my ongoing Causal Econometrics class are now online, so far covering Causal Frameworks, Experiments, Adjustment, Semiparametrics, and IV.
Please yell at me if you see anything you disagree with.
Trying something new: starting next Wednesday, I will be hosting weekly Econometrics office hours free and open to anyone in the world with Econometrics questions.
Sign up at
PSA that due to long-ago choice of username, I am morally obligated to engage with any tweet regarding Donsker's theorem. Please use this power wisely.
Reading Pearl's latest book and seeing economists debate variable selection have really emphasized how badly current instruction that ignores the backdoor criterion has prepared our profession to think clearly about causal inference from observational data.
As I’ve said before, one of the reasons to at least familiarize yourself with Pearls work is for the idea of a “collider”, as well as for reminding ourselves that theory dictates the selection of covariates and to ignore that opens us up to bias.
Nice review of advances in analysis of matched firm-worker "bipartite network" data, including issues with AKM-style two-way fixed effects. To start a fire: "limited mobility bias" is distinct from heterogeneity bias in DiD. Thoughts on how they interact?
Bravo! Isaiah has provided simple and practical methods for hard problems like interpretation, communication, and publication bias of estimates, using asymptotic theory to distill down to simple decision-theory problems using localization.
Some idle musings on why I, as a user of Bayesian statistical methods, don't find Bayesianism normatively compelling as a theory of inductive inference
Anh Nguyen uses variation in government health insurance plan offerings to estimate a nonlinear multilevel model of adverse selection & moral hazard using
@mcmc_stan
. She will be joining CMU econ this year.
To the philosopher of Quantum Mechanics who left
@bayesianboy
's with my shoes last night, I see how Everettianism raises problems for determinate identity, but I don't understand decoherence well enough to draw macroscopic implications.
Also, contact me, I have your shoes.
@causalinf
There's a long history of IV Phillips curve estimates, albeit beset by hetrogeneity, specification search, exclusion and validity issues, weak IV, etc, leading to a wide range of estimates.
FWIW the IV you describe is close to Jordà Schularick & Taylor's "trilemma instrument."
Important new paper by Kocherlakota provides new foundations for causal inference in macroeconomics, linking game theoretic equilibrium and modern causal inference assumptions
Translating for classical time series folks, their STU is a projection onto a 1-sided filter defined by Hankel eigenfunctions after prewhitening, and they give approximation results in a Local-to-Unity model. Reminds me of which is 2-sided.
or you need real Bayesian semiparametrics with ∞-dimensional priors expressing uncertainty about everything you left out then marginalizing (see for tutorial)
@jmwooldridge
@lauretig
@mcmc_stan
@pymc_devs
Yep can handle complicated forms for error without specification via some nonparametrics. Here is a recent example. Dirichlet Process mixture on the left. spline on right. All I did was specify dep. & ind. vars.
@twiecki
Mostly tradition, though my experience coding BVARs in Stan is that unless BLAS code has improved a lot Gibbs methods are in fact orders of magnitude faster and more reliable. And more complex models can't run at all without backend hacks (I have a paper on this forthcoming).
A useful survey of the burgeoning field of information design, which examines the outcomes that can be achieved by providing or withholding information from rational strategic agents, as via recommender systems, certifications, or aggregators.
In a recent interview,
@ben_moll
described the Kolmogorov Forward Equation as the most important equation in macroeconomics. His notes are a great introduction to these applications
@nickchk
R packages for random forests for distributions: (and also grf on CRAN) Seeing these always tells me a lot about my data I wouldn't know otherwise.
Intriguing new approach used for Hamilton-Jacobi-Bellman equations in up to 10,000 dimensions. (Uses tricks based on Feynman-Kac representations and neural networks.)
Love to debate the finer points of Bayesian epistemology when the actual purpose of Bayes is to see the words "LAPACKException" print on my screen 100 times a second for 4 days (and counting) on a run I am almost certain will fail.
@sp_monte_carlo
In my computation class, I introduce FBSDEs as the path space dual to HJB equations as a stochastic analogue to dynamic programming vs optimal control, with similar pros & cons. was a helpful survey of applications;
@momin_rayhan
collects more.
@PossebomVitor
@causalinf
@nickchk
@MatheusFacure
I teach a class centered on the Athanasopoulos and Hyndman Forecasting book that everyone else is (rightly) recommending, but I add on notes with added topics and extensive additional R code examples and exercises:
A foundational work of data visualization, inspirational for any data or social scientist in your life.
P.S.: Data CSVs and example code at thanks to
@ajstarks
MCMC does not seem to play well with non-identified parameters with unbounded support and nearly flat priors, huh.
(This tweet brought to you by the number 10^153.)
@jmwooldridge
You are correct: see
@isabelfulcher
et al for its characterization as one type of indirect effect. Under frontdoor assumptions, the indirect effect is the total effect, but their estimator (code at ) is valid without that.
@VC31415
Unfaithful distributions are measure 0 but dense, such that the Markov structure is not uniformly estimable: see S4 of Whether this can be resolved by stronger uniform versions of faithfulness is an interesting question I'm not certain about.
@sp_monte_carlo
Distinguishing the objective and the probability model has been a mainstay of econometrics since White 1980. Aronow & Miller's Agnostic Statistics is an intro textbook with that perspective.
Berk et al 2019 is a nice expository piece for statisticians.
In macroeconomics, by the Kolmogorov Forward Equation, it describes how a population evolves under persistent forces and impulses, and dually, generates the expectations that a rational agent would use to make plans under that evolution via the HJB equation.
@Andrew___Baker
@rlmcelreath
Assuming you're not interested in macro (in which case Schorfheide's books & chapters are the place), Peter Rossi's work is clear and builds up to real applicable methods. For modern PPL-focused approaches, I don't know of an econ-specific book, though
@abiylfoyp
was trying.
@xuanalogue
This is the point of SWIGs. Single World Intervention Graphs represent intervened nodes by splitting. The Hernán and Robins textbook uses them extensively.
@economeager
@btshapir
@c_r_walt
A friend of mine who was a survey enumerator for a well known development economist before grad school told me that on particularly hot days they would just make the numbers up and go smoke weed in the shade. Real world motivation for the Huber contamination model!
“I would like to see Yusuke continue a very promising career as a scholar,” Dr. Angrist said. “So my main concern in a case like his is that he’s being distracted by other things, and that’s kind of a shame.”
Can someone who speaks Japanese explain what's going on here?
Yusuke Narita, a Yale economist, has pushed Japan's hottest button his suggestions that the elderly commit "mass suicide". He says his comments were "taken out of context;" his critics accuse him of hate speech
@adamkapor
This blog post is a start.
The daggity users guide
@_MiguelHernan
's textbook (
@causalinf
's textbook is on my to read list)
Pearl's survey
Hal White's papers introduced me to the topic (1/2)
@ben_golub
Backtracking references, the answer seems to be yes there are genericity results for *local* Nash eq'a of 2 player games. Screenshot in , found in survey , referencing result in
Welcome to twitter my brilliant colleague
@edwardhkennedy
! If you are interested in causal inference, follow him, and use his
#rstats
package for doubly robust nonparametric treatment effect estimation using machine learning.
@jamesbrandecon
There's a nice paper by Serena Ng on the practice of sketching methods for regressions when X is too big to fit in memory that may have some useful advice
A notebook with a bonus example of what you can do with DifferentiableStateSpaceModels.jl: add heavy tails and stochastic volatility to your macro model.
Probabilistic programming makes it easy to add model features without having to redesign your code.
This project introduces new, easy-to-use tools for Bayesian estimation of nonlinear dynamic economic models.
Try them out with DifferentiableStateSpaceModels.jl! 1/🧵