Intro & Blog – Homepage

Categories

All (20)

Astrophysics (1)

Bayesian (1)

Beta Regression (4)

Causal Inferece (1)

Causal Inference (3)

Climate Change (1)

COVID-19 (2)

Data Science (7)

Democracy (1)

Elections (2)

Experiments (1)

Fixed Effects (2)

International Monetary Fund (1)

Limited Dependent Variables (1)

Linear Models (1)

Measurement (1)

Meta-analysis (1)

Ordered Beta Regression (1)

Panel Data (3)

Political Economy (2)

Probability Theory (1)

Proportions (3)

Public Health (1)

R (17)

Race (1)

Research Design (1)

Sampling (1)

Simulations (1)

Statistics (3)

Task Management (1)

Tenure (1)

Tunisia (2)

Robert Kubinec
Associate Professor
Texas A&M University

I am a political economist who studies the politics of economic development and constructing new methods for rigorous inference in the social sciences. This website hosts papers, code and data from my research projects and also my blog, Punk Rock Poli Sci, that mixes poli sci (the politics of economic development) with a splash of punk rock (rethinking methods for scientific research in the 21st century). You can find my CV here. If you want to contact me, send me an email at bob.kubinec@tamu.edu.

I am the author of two books:

Making Democracy Safe for Business: Corporate Politics During the Arab Uprisings

The Bayesian Hitman

You can subscribe to get email notifications about new blog posts here.

When Using OLS Hurts

Re-examining the Relationship Between Tenure Denial and Race in Masters-Waage et al. (2024)

R

Statistics

Ordered Beta Regression

Tenure

Race

People often use OLS for bounded continuous variables even though we know it isn’t the correct model. Ordered beta regression is a better model–but hard to predict when the results will change. For the article discussed in this blog post, OLS hid an important empirical relationship in the data.

All Your Data R Fixed

Transforming Your RHS Covariates Won’t Fix Your Outcome Distribution

R

Statistics

Linear Models

Probability Theory

In all conventional regression models, the data are fixed–they don’t have a probability distribution. This means transforming your data with logs or other functions won’t change the probability distribution of your outcome.

How to Use Notion for Academic Research

The only thing harder than task management is task management software.

Research Design

Task Management

I provide some advice for how to use Notion’s powerful management software for the unique challenges of academic research.

What Difference-in-Differences Has in Common with Dungeons & Dragons

They both rely on the willing suspension of disbelief.

R

Causal Inference

Panel Data

I discuss why DiD is not the panacea for causal inference that it is often claimed–like all assumptions, the assumptions of DiD can only be validated within a certain research design, and there is no guarantee that DiD is more credible than simple panel data models.

Modeling Uncertainty in NASA’s Asteroid Impact Predictions with the Beta Distribution

It’s Not Rocket Science… Or Is It?

R

Astrophysics

Proportions

Beta Regression

I use brms, Gaussian processes, and a time-varying beta distribution to model the uncertainty in how NASA’s impact probabilities for the asteroid 2024 YR4 are changing over time.

Lost in Transformation: The Horror and Wonder of Logit

Once You See What Logit Is, You’ll Love It Too

R

Statistics

Proportions

The logit model for binary outcomes is poorly understood because it involves thinking in non-linear spaces. In this post, I illustrate what logit is using maps–and why people who resist logit are a bit like flat earthers.

New Survey Research on Tunisia and the IMF Deal

Tunisian Public More Open to the IMF than Its Leaders

Political Economy

International Monetary Fund

Tunisia

Tunisia’s leaders remain adamantly opposed to negotiating or supporting a deal with the IMF, but Tunisian citizens have more nuanced views on the topic.

Robert Kubinec and Sujeong Shim

What Do Tunisians Really Think About President Kais Saied?

R

Tunisia

Democracy

I reproduce results from our recent survey experiment showing that people are opposed to President Kais Saied but prefer not to report their opposition directly.

Robert Kubinec and Amr Yakout

Fixing Fractional Logit?

R

Limited Dependent Variables

Beta Regression

The continuous Bernoulli model is a proposed alternative to the well-known fractional logit specification. I discuss in this blog post when that might or might not be the case.

What’s Logs Got to Do With It?

R

Elections

Beta Regression

Ordered beta regression can give you comparable, scale-free ATEs that can still be understood in the scale of the original data–all without using logs.

Simulating Turnout in Tunisia’s Constitutional Referendum

How much can we learn about true turnout from poll observing?

R

Elections

Simulations

Sampling

I examine how well examining turnout from polling stations is likely to work in estimating turnout for Tunisia’s constitution, especially if we have trouble trusting the numbers of voters reported by the eletion commission.

What To Do (And Not to Do) with Modeling Proportions/Fractional Outcomes

Presenting my new R package ordbetareg for bounded dependent variables.

R

Proportions

Beta Regression

I discuss the available approaches in R for handling proportional/fractional outcomes, and present my new R package ordbetareg as a synthesis of the best of what is currently available.

Energy Nationalism To the Rescue

Putting Trade Barriers on Foreign Oil Could Help the Climate

Political Economy

Climate Change

The Russian invasion of the Ukraine has raised a new specter of energy nationalism. Although only the United Kingdom and the United States have gone as far as outright…

The Causal Representation of Panel Data: A Comment On Xu (2022)

Why Intercepts (Fixed Effects) Aren’t Latent Confounding Variables

R

Fixed Effects

Panel Data

Data Science

I clarify the limitations of the current representation of panel data in causal frameworks and why these limitations lead to confusion about what causal methods can do with panel data.

Why People Are Doubting the AstraZeneca Vaccine Report

Modeling the Garden of Forking Paths with Simulations

R

Causal Inference

COVID-19

Data Science

In this blog post, I use Gelman and Loken’s garden of forking paths analysis to construct a simulation showing why skepticism of AstraZeneca’s vaccine results is warranted…

A More Realistic P-Value for the Pfizer Vaccine Report

Modeling the Vaccine Study Using the Beta-Binomial Distribution

R

Causal Inferece

COVID-19

Data Science

The clinical pre-registration for the Moderna vaccine defines clearly what statistic we should use to evaluate their trial, in contrast to their press release.

Why You Should Be Careful About the MaterniT 21 Test

Non-invasive screenings for Down’s syndrome have low accuracy for many women, but you won’t hear that from the test companies.

R

Public Health

Meta-analysis

Data Science

In this post I discuss the very popular MaterniT 21 test used to identify fetuses with Downe’s Syndrome and other genetic conditions, and show why it can’t really be trusted for women under 40.

What Panel Data Is Really All About

It’s not as confusing as it first appears, but it’s also frequently misused and poorly understood.

R

Panel Data

Fixed Effects

Data Science

In this post I talk about my recent publication with Jonathan Kropko about panel data. We de-mistify panel data, expose some serious weaknesses in common models like two-way fixed effects (2-way FE), and in general help people use panel data with (more) confidence.

A Proposed Model for Partial Identification of SARS-CoV2 Infection Rates Given Observed Tests and Cases

Helping us understand what the observed numbers of cases and tests mean for COVID-19 spread.

R

Measurement

Bayesian

Data Science

A model that shows how to estimate the effect of suppression measures on the SARS-CoV2 without knowing the true infection rate.

Simulating Conjoint Survey Experiments

Power Curves, Clustered Errors, Type S and Type M Error Rates

R

Experiments

Causal Inference

Data Science

R code that shows how to simulate a conjoint experiment using Monte Carlo techniques to produce power curves and test for the effect of clustering in standard errors.