Vladislav Morozov
@vladislavmorozov.bsky.social
Assistant Professor of Econometrics at Uni Bonn
Interested in econometrics and statistics for a heterogeneous world
https://vladislav-morozov.github.io/
Interested in econometrics and statistics for a heterogeneous world
https://vladislav-morozov.github.io/
Link to notes (new sections 11-14):
vladislav-morozov.github.io/econometrics...
Source repo: github.com/vladislav-mo...
vladislav-morozov.github.io/econometrics...
Source repo: github.com/vladislav-mo...
Econometrics with Unobserved Heterogeneity
Learn unobserved heterogeneity-robust causal inference methods — heterogeneous coefficients, nonparametric models, and quantile and distribution regression
vladislav-morozov.github.io
June 10, 2025 at 12:19 PM
Link to notes (new sections 11-14):
vladislav-morozov.github.io/econometrics...
Source repo: github.com/vladislav-mo...
vladislav-morozov.github.io/econometrics...
Source repo: github.com/vladislav-mo...
My favorite result in this block:
Even with just 2 periods of data, you can identify average causal effects, even if people differ in infinitely many unobserved ways and the outcome function is completely unrestricted.
That's the power of panel data.
Even with just 2 periods of data, you can identify average causal effects, even if people differ in infinitely many unobserved ways and the outcome function is completely unrestricted.
That's the power of panel data.
June 10, 2025 at 12:19 PM
My favorite result in this block:
Even with just 2 periods of data, you can identify average causal effects, even if people differ in infinitely many unobserved ways and the outcome function is completely unrestricted.
That's the power of panel data.
Even with just 2 periods of data, you can identify average causal effects, even if people differ in infinitely many unobserved ways and the outcome function is completely unrestricted.
That's the power of panel data.
Beyond linearity:
What can we still learn when we don’t restrict functional form and allow arbitrarily rich unobserved heterogeneity?
This new section covers:
• A gentle intro
• Heterogeneity bias
• Average effects via panel data
• Stayers and why they matter
• Local polynomial regression
What can we still learn when we don’t restrict functional form and allow arbitrarily rich unobserved heterogeneity?
This new section covers:
• A gentle intro
• Heterogeneity bias
• Average effects via panel data
• Stayers and why they matter
• Local polynomial regression
June 10, 2025 at 12:19 PM
Beyond linearity:
What can we still learn when we don’t restrict functional form and allow arbitrarily rich unobserved heterogeneity?
This new section covers:
• A gentle intro
• Heterogeneity bias
• Average effects via panel data
• Stayers and why they matter
• Local polynomial regression
What can we still learn when we don’t restrict functional form and allow arbitrarily rich unobserved heterogeneity?
This new section covers:
• A gentle intro
• Heterogeneity bias
• Average effects via panel data
• Stayers and why they matter
• Local polynomial regression
Wrote a short post with details about why it happened and what I like.
vladislav-morozov.github.io/blog/web/qua...
vladislav-morozov.github.io/blog/web/qua...
Why I Switched from Beamer to Quarto Reveal.js for My Presentations
Why I switched from Beamer to Quarto Reveal.js for reproducible, maintainable, and portable slides in teaching, research, and data science.
vladislav-morozov.github.io
June 4, 2025 at 11:45 AM
Wrote a short post with details about why it happened and what I like.
vladislav-morozov.github.io/blog/web/qua...
vladislav-morozov.github.io/blog/web/qua...
I version-control everything with Git, sync and deploy via GitHub, and present directly from a browser.
It’s reproducible, portable, and just works.
It’s reproducible, portable, and just works.
June 4, 2025 at 11:45 AM
I version-control everything with Git, sync and deploy via GitHub, and present directly from a browser.
It’s reproducible, portable, and just works.
It’s reproducible, portable, and just works.
Executable slides: code runs during render, outputs (plots, tables) are embedded automatically.
Simple syntax, responsive HTML, and interactive options too.
Simple syntax, responsive HTML, and interactive options too.
June 4, 2025 at 11:45 AM
Executable slides: code runs during render, outputs (plots, tables) are embedded automatically.
Simple syntax, responsive HTML, and interactive options too.
Simple syntax, responsive HTML, and interactive options too.
Lecture notes here: (new sections are 7-10)
vladislav-morozov.github.io/econometrics...
Or
github.com/vladislav-mo...
vladislav-morozov.github.io/econometrics...
Or
github.com/vladislav-mo...
7 Variance of Heterogeneous Coefficients – Econometrics with Unobserved Heterogeneity
Learn how to identify and estimate the variance of heterogeneous coefficients in linear panel models under minimal assumptions (Lecture Notes)
vladislav-morozov.github.io
April 29, 2025 at 1:31 PM
Lecture notes here: (new sections are 7-10)
vladislav-morozov.github.io/econometrics...
Or
github.com/vladislav-mo...
vladislav-morozov.github.io/econometrics...
Or
github.com/vladislav-mo...
The main idea: you can identify the full distribution of effects almost as easily as the average!
But these results aren’t widely used — maybe because the original treatment is pretty dense. I tried to make them more accessible via a clean special case.
But these results aren’t widely used — maybe because the original treatment is pretty dense. I tried to make them more accessible via a clean special case.
April 29, 2025 at 1:31 PM
The main idea: you can identify the full distribution of effects almost as easily as the average!
But these results aren’t widely used — maybe because the original treatment is pretty dense. I tried to make them more accessible via a clean special case.
But these results aren’t widely used — maybe because the original treatment is pretty dense. I tried to make them more accessible via a clean special case.
Sorry, missed it! Maybe for some very tractable models?
Otherwise, only the usual characterization for misspecified likelihood: that you are estimating the parameter that minimizes the KL-divergence between the true model and the specified one
I usually find it hard to interpret those...
Otherwise, only the usual characterization for misspecified likelihood: that you are estimating the parameter that minimizes the KL-divergence between the true model and the specified one
I usually find it hard to interpret those...
April 21, 2025 at 6:41 PM
Sorry, missed it! Maybe for some very tractable models?
Otherwise, only the usual characterization for misspecified likelihood: that you are estimating the parameter that minimizes the KL-divergence between the true model and the specified one
I usually find it hard to interpret those...
Otherwise, only the usual characterization for misspecified likelihood: that you are estimating the parameter that minimizes the KL-divergence between the true model and the specified one
I usually find it hard to interpret those...
Turns out, the answer is only mostly right:
1. Yes, adjusted multiple testing can lead to a huge loss of power.
2. Surprisingly, in some cases, simultaneous testing actually performs worse (though only slightly).
1. Yes, adjusted multiple testing can lead to a huge loss of power.
2. Surprisingly, in some cases, simultaneous testing actually performs worse (though only slightly).
April 2, 2025 at 7:56 AM
Turns out, the answer is only mostly right:
1. Yes, adjusted multiple testing can lead to a huge loss of power.
2. Surprisingly, in some cases, simultaneous testing actually performs worse (though only slightly).
1. Yes, adjusted multiple testing can lead to a huge loss of power.
2. Surprisingly, in some cases, simultaneous testing actually performs worse (though only slightly).
Very true! Comes down to what you care about.
As an aside, if you drop linearity of the model, OLS — fixed effects models in this case — can give you "bad" weighted averages with potentially negative weights.
Then you really don't have a nice estimand.
www.aeaweb.org/articles?id=...
As an aside, if you drop linearity of the model, OLS — fixed effects models in this case — can give you "bad" weighted averages with potentially negative weights.
Then you really don't have a nice estimand.
www.aeaweb.org/articles?id=...
Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects
(September 2020) - Linear regressions with period and group fixed effects are widely used to estimate treatment effects. We show that they estimate weighted sums of the average treatment effects (ATE)...
www.aeaweb.org
March 21, 2025 at 3:15 PM
Very true! Comes down to what you care about.
As an aside, if you drop linearity of the model, OLS — fixed effects models in this case — can give you "bad" weighted averages with potentially negative weights.
Then you really don't have a nice estimand.
www.aeaweb.org/articles?id=...
As an aside, if you drop linearity of the model, OLS — fixed effects models in this case — can give you "bad" weighted averages with potentially negative weights.
Then you really don't have a nice estimand.
www.aeaweb.org/articles?id=...
If a researcher
1. Knows that the effect is non-negative
2. Thinks that the within regression is targeting the ATE,
they will conclude that that there is no effect.
Even if M is very large and there are many people with β_i = M, so you would have a strong effect from intervening on x.
1. Knows that the effect is non-negative
2. Thinks that the within regression is targeting the ATE,
they will conclude that that there is no effect.
Even if M is very large and there are many people with β_i = M, so you would have a strong effect from intervening on x.
March 18, 2025 at 3:13 PM
If a researcher
1. Knows that the effect is non-negative
2. Thinks that the within regression is targeting the ATE,
they will conclude that that there is no effect.
Even if M is very large and there are many people with β_i = M, so you would have a strong effect from intervening on x.
1. Knows that the effect is non-negative
2. Thinks that the within regression is targeting the ATE,
they will conclude that that there is no effect.
Even if M is very large and there are many people with β_i = M, so you would have a strong effect from intervening on x.
A simple example: suppose that you have two periods, one covariate x_{it}, and two types for β: some crazy big number M and 0.
1. Units with positive β do not change x.
2. Units with β=0 change x.
The estimand of the within regression is 0, regardless of the proportions of the types and M.
1. Units with positive β do not change x.
2. Units with β=0 change x.
The estimand of the within regression is 0, regardless of the proportions of the types and M.
March 18, 2025 at 3:13 PM
A simple example: suppose that you have two periods, one covariate x_{it}, and two types for β: some crazy big number M and 0.
1. Units with positive β do not change x.
2. Units with β=0 change x.
The estimand of the within regression is 0, regardless of the proportions of the types and M.
1. Units with positive β do not change x.
2. Units with β=0 change x.
The estimand of the within regression is 0, regardless of the proportions of the types and M.
Good point!
It is a perfectly fine estimand under a linear model — a convex average of individual effects.
The problem is in (economic) practice: people often interpret that as the genuine ATE. Then one may draw wrong conclusions — this effect can have the opposite sign from the ATE.
It is a perfectly fine estimand under a linear model — a convex average of individual effects.
The problem is in (economic) practice: people often interpret that as the genuine ATE. Then one may draw wrong conclusions — this effect can have the opposite sign from the ATE.
March 18, 2025 at 3:13 PM
Good point!
It is a perfectly fine estimand under a linear model — a convex average of individual effects.
The problem is in (economic) practice: people often interpret that as the genuine ATE. Then one may draw wrong conclusions — this effect can have the opposite sign from the ATE.
It is a perfectly fine estimand under a linear model — a convex average of individual effects.
The problem is in (economic) practice: people often interpret that as the genuine ATE. Then one may draw wrong conclusions — this effect can have the opposite sign from the ATE.
I have learned a lot much from others openly sharing their specialized materials.
It's only fair to offer my epsilon as well and I hope these materials can serve someone.
It's only fair to offer my epsilon as well and I hope these materials can serve someone.
March 18, 2025 at 8:50 AM
I have learned a lot much from others openly sharing their specialized materials.
It's only fair to offer my epsilon as well and I hope these materials can serve someone.
It's only fair to offer my epsilon as well and I hope these materials can serve someone.
Example with productivity: onlinelibrary.wiley.com/doi/abs/10.3...
Example with worker skills:
academic.oup.com/restud/artic...
The Jochmans and Weidner paper above cites some more examples.
Example with worker skills:
academic.oup.com/restud/artic...
The Jochmans and Weidner paper above cites some more examples.
Econometrica | Econometric Society Journal | Wiley Online Library
Firms are more productive, on average, in larger cities. Two main explanations have been offered: firm selection (larger cities toughen competition, allowing only the most productive to survive) and ....
onlinelibrary.wiley.com
February 18, 2025 at 11:39 AM
Example with productivity: onlinelibrary.wiley.com/doi/abs/10.3...
Example with worker skills:
academic.oup.com/restud/artic...
The Jochmans and Weidner paper above cites some more examples.
Example with worker skills:
academic.oup.com/restud/artic...
The Jochmans and Weidner paper above cites some more examples.
I dunno if it's what you mean, but some other examples are:
1. Firm-level productivity (TFP)
2. Worker skills
3. Teacher value added.
You may care about their distribution, but you have to estimate all these (with noise).
A paper on working with such estimates:
arxiv.org/abs/1803.049...
1. Firm-level productivity (TFP)
2. Worker skills
3. Teacher value added.
You may care about their distribution, but you have to estimate all these (with noise).
A paper on working with such estimates:
arxiv.org/abs/1803.049...
Inference on a Distribution from Noisy Draws
We consider a situation where the distribution of a random variable is being estimated by the empirical distribution of noisy measurements of that variable. This is common practice in, for example, te...
arxiv.org
February 18, 2025 at 11:39 AM
I dunno if it's what you mean, but some other examples are:
1. Firm-level productivity (TFP)
2. Worker skills
3. Teacher value added.
You may care about their distribution, but you have to estimate all these (with noise).
A paper on working with such estimates:
arxiv.org/abs/1803.049...
1. Firm-level productivity (TFP)
2. Worker skills
3. Teacher value added.
You may care about their distribution, but you have to estimate all these (with noise).
A paper on working with such estimates:
arxiv.org/abs/1803.049...
Honestly? Everything just works. It’s fast, integrates with Zotero, and fits my workflow way better.
Still figuring out the best setup, but I'll document it when I find a winning approach.
Still figuring out the best setup, but I'll document it when I find a winning approach.
February 12, 2025 at 8:59 AM
Honestly? Everything just works. It’s fast, integrates with Zotero, and fits my workflow way better.
Still figuring out the best setup, but I'll document it when I find a winning approach.
Still figuring out the best setup, but I'll document it when I find a winning approach.