Fixed Effects and Panel Data

The one idea: If you observe the same units repeatedly over time, you don't need to measure every confounder — you can difference away anything about a unit that stays constant across periods, known or unknown, just by comparing each unit to its own earlier self. Fixed effects turn "I'm missing a confounder" into "I don't need to find it, as long as it doesn't change."

⚠️ The trap: "We have years of history on each account, so we can just control for account in a regression and call it causal." That feels right and it's incomplete. A fixed-effects model only controls for confounders that are constant for each unit across the whole panel — it does nothing for a confounder that changes over time along with the treatment. A promotion that coincides with a manager change, or a price cut that coincides with a competitor's move, still biases the estimate, fixed effects or not.

Every method so far in this track — matching, the backdoor criterion, regression adjustment — needs you to find and measure the confounders. Panel data hands you a different lever entirely, when it's available: the same units observed at multiple points in time. A panel might be 50 customer accounts each measured monthly for two years, or 30 stores each tracked weekly, or 200 sales reps each observed quarter over quarter. The repeated-observation structure means you can compare each unit to itself across time, and that self-comparison automatically cancels out anything about the unit that doesn't change — including confounders you never measured and might not even be able to name.

The intuition: your own baseline…