longbet.qmd

---
title: "LongBet: Heterogeneous Treatment Effects in Panel Data"
share:
  permalink: "https://book.martinez.fyi/longbet.html"
  description: "Business Data Science: What Does it Mean to Be Data-Driven?"
  linkedin: true
  email: true
  mastodon: true
---

## Introduction to LongBet

As businesses and policymakers increasingly rely on longitudinal data to make
causal inferences, there's a growing need for methods that can handle the
complexities of panel data while estimating heterogeneous treatment effects.
LongBet, introduced by @wang2024longbet, fills this gap by extending the
Bayesian Causal Forest (BCF) model to panel data settings.

LongBet is particularly suited for short panel data with large cross-sectional
samples and observed confounders. Unlike traditional difference-in-differences
methods that often rely on the parallel trends assumption, LongBet leverages
observed confounders to impute potential outcomes and identify treatment
effects. LongBet models the data generating process as follows:

$$
Y_{it} = \alpha\mu(X_i, T=t) + \beta_S\nu(X_i, S_{it}, T=t) + \epsilon_{it}
$$

where:

  - $Y_{it}$ is the outcome for individual $i$ at time $t$
  - $X_i$ are time-invariant covariates
  - $S_{it}$ is the time elapsed since treatment adoption
  - $T$ is the time period
  - $\mu(\cdot)$ is the prognostic function
  - $\nu(\cdot)$ is the treatment effect function
  - $\beta_S$ is a Gaussian process factor capturing the general trend of
    treatment effects
  - $\epsilon_{it}$ is an independent Gaussian error term

Both $\mu(\cdot)$ and $\nu(\cdot)$ are modeled using XBART and considering
splits on the time dimension.

## Key Features of LongBet

1.  Flexibility in Trend Modeling: LongBet doesn't require the parallel trends
    assumption, making it suitable for scenarios where this assumption may not
    hold.
2.  Handling of Staggered Adoption: The model can accommodate treatments adopted
    at different times across units.
3.  Separation of Prognostic and Treatment Effects: Like BCF, LongBet separates
    the modeling of prognostic and treatment effects, allowing for different
    regularization strategies.
4.  Time-Varying Treatment Effects: The model captures how treatment effects may
    change over time since adoption.
5.  Bayesian Uncertainty Quantification: As a Bayesian method, LongBet provides
    credible intervals for conditional treatment effects.

## Comparison with Traditional Panel Methods

LongBet offers several advantages over traditional panel data methods:

1.  Relaxed Assumptions: Unlike difference-in-differences, LongBet doesn't rely
    on the parallel trends assumption.
2.  Heterogeneous Effects: LongBet can capture complex, nonlinear heterogeneity
    in treatment effects.
3.  Flexible Functional Form: The use of tree-based models allows for flexible
    modeling of the relationship between outcomes, covariates, and time.
4.  Uncertainty Quantification: The Bayesian framework provides natural
    uncertainty estimates for all quantities of interest.
    
## Example Application

TODO    
    
## Conclusion

LongBet represents a significant advancement in causal inference for panel data.
By combining the flexibility of tree-based methods with the ability to handle
longitudinal data, it offers a powerful tool for estimating heterogeneous
treatment effects in complex, real-world scenarios.

However, like all methods, LongBet has its limitations. It assumes that all
relevant confounders are observed and time-invariant. It may also struggle in
scenarios with limited overlap between treated and control units across time. As
always in causal inference, careful consideration of the problem at hand and the
assumptions of the method is crucial.