Stata中含unit specific trends的Dif-in-Dif回归adjusted R²过高问题咨询

Hey there, let’s unpack why you’re seeing those near-perfect adjusted R² values (even 0.99!) in your DID model with unit-specific trends, and walk through how to verify if this is a problem or just a quirk of your setup.

First, why the sky-high R²? #

This is actually pretty common in panel models with unit-specific trends, especially given your dataset structure:

• Unit-specific trends are powerful fitters: You’re adding 201 separate linear trends (one per country) to your model. With 36 years of data per country, these trends will soak up almost all the long-term, country-specific variation in your dependent variable—think things like gradual GDP growth, population increases, or policy trajectories unique to each nation. If your DV has strong time-dependent patterns at the country level, these trends will explain most of its variation.
• Parameter count relative to observations: You’ve got ~450 parameters for 5000 observations. While that’s not an extreme ratio (roughly 11 observations per parameter), combining country fixed effects, year fixed effects, 5 controls, and 201 trend terms means your model is accounting for a huge portion of the data’s structure.
• DV type matters: If your dependent variable is a macroeconomic or slow-changing variable (e.g., GDP per capita, life expectancy), it naturally has strong linear trends at the country level. The unit-specific trends will essentially "fit" those trends almost perfectly, driving R² way up.

What to do next to validate your model #

Don’t panic—high R² doesn’t automatically mean your model is wrong, but you should check a few key things to ensure your core DID estimate is reliable:

• Test the model without unit-specific trends:
  Run a simpler DID specification first to compare results:

  xtset country year
  reg y treat i.year controls, robust fe
  

  Then compare it to your full model with trends:

  reg y treat i.year controls i.country#c.year, robust fe
  

  Look at how your treatment effect coefficient changes—does it stay significant? Does its magnitude make economic sense? If the treatment effect disappears or becomes implausible when adding trends, that’s a red flag. Also, note the drop in R² when removing trends—this will tell you how much variation the trends are actually explaining.

• Don’t fixate on adjusted R² for DID models:
  Adjusted R² measures overall fit, but in DID, we care far more about the causal effect of the treatment, not how well the model predicts the DV. Even if R² is 0.99, if your treatment effect is statistically significant and aligns with your theoretical expectations, the model might still be valid.

• Decompose your dependent variable’s variation:
  Use Stata’s xtsum command (after xtset-ing your data) to split your DV into between-country variation and within-country over-time variation:

  xtsum y
  

  Unit-specific trends explain within-country variation. If most of your DV’s variation is within-country (high within percentage), that’s why the trends are driving R² up—and that’s totally normal for many macro panel datasets.

• Check for overfitting (if you’re worried):
  Split your dataset into a training and test sample (e.g., use the first 30 years to estimate the model, then predict the last 6 years). If the predicted values are close to the actual values, the high R² is just a reflection of your data’s structure, not overfitting. If the predictions are way off, you might need to revisit your model specification.

• Verify no data anomalies:
  Quick checks to rule out silly mistakes:

  • Are there any extreme outliers in your DV? Use summarize y, detail or histogram y to check.
  • Is there accidental multicollinearity? While fixed effects models mitigate this, you can run estat vif on a model without fixed effects to see if controls are highly correlated (though take VIF results with a grain of salt for panel models).

Final takeaway #

In most cases, a high adjusted R² in a DID model with unit-specific trends is not an error—it’s just the model doing its job of accounting for country-specific time patterns. The real test is whether your treatment effect is robust, statistically significant, and makes sense for your research question.

内容的提问来源于stack exchange，提问作者Angelo Cozzubo