[Purpose] Matching is used to reduce selection bias by making treatment and control groups comparable in observable characteristics, thereby enabling more accurate estimation of causal effects.
[In General] There are many matching techniques (e.g., Propensity Score Matching, Coarsened Exact Matching, Nearest Neighbor Matching, Entropy Balancing, etc.).
[Actually] Including control variables and fixed effects (FEs) in a regression serves a similar purpose to matching by statistically adjusting for observed covariates and unobserved heterogeneity, thereby improving comparability between treatment and control groups. — Thanks to Prof. Vladimir Mukharlyamov
[Using controls and FEs could be good enough: then why matching?] Even after controlling for covariates and fixed effects in a regression, matching may further improve comparability between treatment and control groups by ensuring common support and reducing reliance on functional form assumptions.
[Purpose] It is not enough to find something (empirically or theoretically). We should also understand why and how it happens—that is, identify the underlying mechanisms and channels.
Mechanism: Why and through what underlying economic principles does the effect arise along that pathway? (How and why does it work?)
Channel: Through what pathway is an economic effect transmitted? (Through what pathway?)
[Important 1] The channel or mechanism you suggest must make sense to scholars in the relevant field and be theoretically sound. We cannot simply make any arbitrary claims.
[Important 2] A convincing mechanism test should identify empirical patterns that are difficult to explain through alternative channels but are more consistent with the proposed mechanism. — Thanks to Prof. Laura Veldkamp
Frisch–Waugh–Lovell (FWL) Theorem: [Frisch and Waugh (1933)], [Lovell (1963)]
[Purpose] A fixed-effects residual plot visualizes the relationship between two variables after controlling for fixed effects. It shows the relationship after removing the effects of unit-specific characteristics and common shocks captured by the fixed effects (FEs). — Thanks to Prof. Marco Sammon
[How?]
Step 1. Regress the dependent variable (Y) on the specified FE(s), and extract its residuals (Y_resid or Y_tilde) by subtracting Y_hat (the fitted value) from the actual Y value.
Step 2. Regress the variable of interest (X) on the specified FE(s), and extract its residuals (X_resid or X_tilde) by subtracting X_hat (the fitted value) from the actual X value.
Step 3. Create a scatter plot with X_resid on the x-axis and Y_resid on the y-axis.
Step 4. Add a fitted regression line. Its slope is identical to beta_hat (the estimated coefficient on X) from the original regression that includes the same fixed effects.
Helpful Materials: [Oxford Causal Inference Course (19.1)]
The video was made when I was an MS student (2024): https://youtu.be/4pFR_6fwyzM
This is a rough analysis I did on my own (I’m aware that there are many points that could be criticized). So, please don't take it too seriously.