Econometrics

A randomized experiment is the gold standard for causation — but most of the questions that matter can't be randomized. You can't randomly assign a policy, a recession, or who receives a treatment they chose themselves. Yet the causal question remains: did the change actually cause the effect, or would it have happened anyway?

Econometrics is the family of quasi-experimental methods that answer that question from observational data. Each one is a different trick for approximating the missing experiment — for constructing a credible picture of what would have happened without the treatment.

Difference-in-Differences

When a change hits one group and not another, DiD compares the change over time in the treated group to the change in an untreated control. If the two moved in parallel before the treatment, the gap that opens afterward is the causal effect — as the diverging lines above show.

1. 1Find a treated group and a comparable control that wasn't treated.
2. 2Confirm they moved in parallel before the intervention (the key assumption).
3. 3Measure how each changed after; the difference in those differences is the effect.

Propensity Score Matching

When treated and untreated units differ systematically — treated customers were already bigger spenders — a raw comparison confuses the treatment with the difference. PSM pairs each treated unit with an untreated one that had the same probability of being treated, so like is compared with like.

Instrumental Variables

Sometimes the treatment is tangled with the outcome — they influence each other, or share an unmeasured cause. An instrument is a third variable that nudges the treatment but affects the outcome only through it, letting you isolate the causal effect from the confounded part.

A valid instrument is hard to find and its assumptions are strong, but when one exists it can rescue a causal estimate that ordinary regression can't.

Panel Data: Fixed and Random Effects

When you observe the same units over time — customers across months, regions across years — panel methods control for everything stable about each unit, measured or not.

Fixed effects are the workhorse: by comparing each unit to itself over time, they sweep away every fixed confounder — anything about a customer or region that doesn't change.

Econometrics in the SKARI Statistical Lab

SKARI's Statistical Lab treats econometrics as its own family — the causal-inference toolkit for when you can't run an experiment.

• Difference-in-differences with a pre-trend view
• Propensity score matching to balance measured confounders
• Instrumental variables for endogenous treatments
• Panel fixed- and random-effects models
• Regression discontinuity for cutoff-assigned treatments

Frequently Asked Questions

Isn't this just regression?

Regression is often the engine, but econometrics is about the design around it — matching, instruments, panels — that lets a coefficient carry a causal meaning.

When should I use DiD vs matching?

DiD when you have before-and-after data and a parallel control; matching when you have a snapshot and need to balance groups on measured confounders.

Can these prove causation for certain?

Never with certainty — each rests on assumptions. But a credible design makes a far stronger case than a raw correlation ever could.

Key Takeaways

Econometrics recovers cause where randomization isn't possible. Pick the design that fits your data — difference-in-differences, matching, an instrument, or a panel — build a credible counterfactual, and be honest about the assumptions it rests on.