What is econometric analysis? From question to evidence
Econometric analysis connects an economic question to data, a model and defensible conclusions. Follow a practical workflow with a policy-evaluation example.
Econometric analysis is the process of using economic data, statistical methods and explicit assumptions to investigate an economic question. The finished analysis is more than a fitted model: it connects a target quantity to a dataset, a comparison, an uncertainty calculation and a conclusion whose limits are clear.
If econometrics is the field, an econometric analysis is a particular investigation. This article follows the practical decisions involved rather than cataloguing every estimator.
Define what you want to learn
Consider a hypothetical city offering a public-transport discount to selected neighbourhoods. The research question is whether the offer reduced household car journeys over the following six months.
Before opening the data, specify the outcome and target population. Does “car journeys” mean all trips, commuting trips or distance travelled? Is the target everyone eligible, people who accepted the discount or residents who stayed in the same neighbourhood? These are different quantities.
Also distinguish the purpose. Describing travel among discount users, predicting next month's car use and estimating the effect of offering the discount require different evidence. The data analysis introduction compares those goals.
Choose a comparison before choosing an estimator
Discount users may already prefer public transport. Comparing users with non-users after the offer therefore mixes the programme with existing differences. Comparing a neighbourhood with itself before and after also mixes the offer with fuel prices, weather and other changes over time.
One possible design compares changes in eligible and otherwise comparable ineligible neighbourhoods. This motivates difference-in-differences, but the method name does not validate the design. The analyst must consider whether their untreated outcomes would have followed comparable trends and whether spillovers or anticipation undermine the comparison.
Where allocation was randomised, the analysis would instead start from that assignment mechanism. The appropriate design depends on the actual programme, not on which command is most convenient.
Build an analysis sample you can explain
| Decision | Example in the transport study | What to record |
|---|---|---|
| Unit | Household observed each month | Household and month identifiers |
| Outcome | Number of car journeys in a fixed recall window | Survey question, units and coding |
| Treatment | Eligibility for the discount offer | Assignment rule and announcement date |
| Coverage | Six months before and after | Missing months, attrition and entry rules |
| Exclusions | Records without a valid household identifier | Rule and number excluded |
Preserve raw inputs and create the analysis sample with reproducible transformations. Check that a join has not duplicated households and that the survey question remained consistent. A larger row count after merging files can reflect an error rather than new information.
If households appear repeatedly, the dataset has a panel structure. Repeated records should influence both model specification and the treatment of uncertainty.
Specify, estimate and inspect
Write the model in terms of the research design: outcome, treatment timing, comparison group and any justified controls. Explain why each variable belongs. Avoid choosing controls solely because they produce attractive p-values; a variable affected by treatment can change the causal quantity being estimated.
After estimation, check whether the sample matches the intended one. Inspect missing-value exclusions, group sizes and the variation used to estimate the effect. In a small programme, hundreds of household observations may still correspond to very few independently assigned neighbourhoods.
The methods reference documents supported functions and their options. An estimator's availability is a software fact; whether it answers this programme's question is a research-design judgement.
Quantify uncertainty and challenge the result
Report an estimate in meaningful units, with its confidence interval and inference method. If the offer was assigned at neighbourhood level, independent household standard errors may be inappropriate. See the robust and clustered inference guide.
Then examine plausible threats. Would results change with a different defensible window? Did the groups already diverge before the offer? Did survey response change? Were transport improvements introduced at the same time? Pre-treatment checks can reveal problems, but a reassuring graph cannot prove an unobserved counterfactual.
Keep planned checks distinct from exploratory ones. Record all consequential specification choices, including results that weaken the preferred interpretation. The point is to understand the fragility of the conclusion.
Write the conclusion the design supports
A useful report identifies the population, period, comparison, estimate, uncertainty and main limitations. It distinguishes an effect of offering a discount from an effect of actually using one, and avoids generalising a local six-month result to every city or time horizon.
Save the data definitions, analysis code, package versions and final outputs together. The reproducible workflow guide covers that handoff. To practise the full sequence on teaching data, use the economic question lab and research report lab.
References
- Ouliaris, S. IMF, Finance & Development. Econometrics: Making theory count
- Cunningham, S. Causal Inference: The Remix. Foundational ideas in causal inference
- The Royal Swedish Academy of Sciences (2021). Scientific background: Answering causal questions using observational data