Synthetic Difference-in-Differences Estimation

Role of income and policy in reducing water pollution: Evidence from the Mississippi River Basin Healthy Watersheds Initiative

“…We explore the extent to which our ATET estimates are robust to the inclusion of multiple MRBI watershed program implementation years, specifically 2015 and 2016. Due to potential econometric issues with staggered adoption and the interpretation of results, we use a synthetic difference-in-differences (SDID) estimator (Arkhangelsky et al, 2021;Clarke et al, 2023). Like synthetic control methods (Abadie et al, 2010(Abadie et al, , 2015Abadie & Gardeazabal, 2003), SDID forms a unique convex weighting of underlying control units to create a synthetic control that is closely matched to the treated unit in pre-treatment outcomes.…”

The heterogeneous impact of house purchase limits policy on housing prices: Comparison between elite and non‐elite school district houses

“…Second, we employ the synthetic differences‐in‐differences (SDID) (Clarke et al., 2023; Arkhangelsky et al., 2021) estimator to reestimate the baseline results. Compared with the canonical DID model, SDID calculates optimal weights that balance the pre‐trends between the ESDH and NSDH, and trends before and after the limits policy within the ESDH; thus, we can reduce the potential bias and improve the precision of the baseline estimations.…”

The influence of ChatGPT on artificial intelligence related crypto assets: Evidence from a synthetic control analysis

“…The inconclusive outcomes observed in the investigation of parallel trends necessitate the adoption of a more sophisticated analytical framework to ensure the veracity of model inferences and derived estimations. We continue by implementing the SDID methodology of Arkhangelsky et al (2021) and Clarke et al (2023) for determining the impact of a binary treatment variable (i.e., equal to one after the ChatGPT launch, 𝐺𝑃𝑇 𝑖𝑡 , and zero otherwise) on an outcome variable (i.e., crypto-asset returns, 𝑅 𝑖𝑡 ) for a panel of 𝑁 crypto assets, observed over 𝑇 time intervals. 10 The procedure applies a treatment (i.e., ChatGPT launch) to be received 9 Figure A1 provides some qualitative support for parallel pre-treatment trends using observed means and linear trend models.…”

“…This has the advantage of seeking to match treated and control units on pre-treatment trends, rather than pre-treatment trends and levels, allowing for a constant difference between treatment and control units. Overall, SDID provides a flexible and robust approach for estimating ATT, by accounting for both shared temporal aggregate factors and unit-specific factors (Clarke et al, 2023). To further account for exogenous time-varying covariates of market capitalization (𝐶𝑎𝑝 𝑖𝑡 ) and liquidity (𝑉𝑜𝑙 𝑖𝑡 ) of crypto assets, the SDID model is adjusted as:…”

“…Firstly, we utilize the synthetic difference-in-difference (SDID) approach of Arkhangelsky et al (2021) and Clarke et al (2023) to examine the relative evolutionary performance of a large set of AI-assets relative to controls. 5 Results demonstrate that AI-assets exhibited positive average treatment effects (ATTs) of 10.7% to 15.6% (35.5% to 41.3%) in the one-month (twomonth) period following the ChatGPT launch.…”