[21]{.chapter-number}  [LongBet: Heterogeneous Treatment Effects in Panel Data]{.chapter-title}

21 LongBet: Heterogeneous Treatment Effects in Panel Data

21.1 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 Wang, Martinez, and Hahn (2024), 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.

21.2 Key Features of LongBet

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

21.3 Comparison with Traditional Panel Methods

LongBet offers several advantages over traditional panel data methods:

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

21.4 Example Application

TODO

21.5 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.

--- 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.