Cookies on Macro by Mark.We use essential cookies for sign-in, locale, and security. With your permission, we also use optional cookies for preferences, for analytics, and for advertising measurement with the Meta (Facebook) Pixel. Read our cookie policy.
Pooled cross-country or cross-region regression with country and time fixed effects to net out unit-level and aggregate variation.
Graph workspace
The defining visualization for fixed-effects panel estimation is the within-entity relationship: how changes in within each entity relate to changes in , net of entity-level means. Plotting demeaned against demeaned reveals the within variation that identifies , and overlaying entity-specific lines shows the fixed effects.
Pooled OLS estimate
Biased estimate that confounds within-entity and between-entity variation. Inflated (or deflated) by correlation between fixed effects and regressor means.
1.36
same as β
Fixed Effects estimate
Consistent estimate using only within-entity variation. Purged of time-invariant omitted variable bias.
0.4710
same as β
OLS bias
Difference between OLS and the true value. Positive when and negative when , whatever the sign of the true slope: it comes from the omitted fixed effect, not from .
0.8578
same as β
Clustered SE
Standard error that accounts for within-entity serial correlation. Larger than naive (homoskedastic) SE when errors are correlated within entities.
0.1524
same as β
True slope coefficient: 0.5
The real effect that both estimates are trying to recover.
Changes the slope of the demeaned regression line. Both OLS and FE slopes rotate, but FE stays closer to the true value.
FE-regressor correlation (OLS bias source): 0.6
How strongly an entity's baseline level moves with its average x; this is what biases the pooled estimate.
Higher widens the gap between the OLS and FE regression lines. When the two slopes differ only by sampling noise (about 0.58 against 0.48 at the defaults).
Entity heterogeneity (fixed effect SD): 2
How much entities differ from one another in their baseline levels.
Higher spreads the entity clusters further apart in the raw scatter. The FE estimate is unaffected (entity means are removed): it stays at about 0.47 at the defaults whatever is.
How many separate entities the data follows; more entities make the estimate more precise.
More colored clusters appear in the scatter. With more entities the FE estimate settles near the true slope (within about 0.04 of 0.5 from 18 entities up) and the clustered SE becomes trustworthy. With only a handful of clusters it is itself unreliable: on this sample it reads 0.05 at 3 entities, below its value at 30, and jumps between about 0.05 and 0.27 below 10.
How many periods each entity is observed; more periods give more change to learn from.
Each entity's cluster has more points. Within-entity regression lines become better defined. The overall FE estimate precision improves.