Understand methods and results
A short guide to interpreting causal estimates in aitiolin.
aitiolin is an open-source causal inference prototype for private and educational use. Its Reviewed comparison feature helps you define a causal question, compare estimation methods, examine uncertainty and assess the assumptions behind a result.
Important: A numerical estimate is not proof of causation. The validity of an interpretation depends on the data, the proposed causal graph and the assumptions used in the analysis.
These explanations apply to Reviewed comparison. The existing inference, agent and robustness controls use separate settings.
1. What are you comparing?
Before estimating an effect, specify:
- Treatment and control: Which two conditions or values are being compared?
- Outcome: What is being measured? For a binary outcome, which event counts as 1?
- Target population: Are you estimating the Average Treatment Effect (ATE), the effect among treated observations (ATT), or the effect among controls (ATC)?
- Adjustment variables: Which potential confounders should be included, based on your causal graph?
aitiolin also lets you review how categorical variables and missing values are handled. These choices affect the interpretation of the result.
2. How should you interpret an effect?
Suppose an analysis reports an estimated effect of −0.05 on the probability of customer churn.
This represents a difference of five percentage points, not a relative reduction of five percent. For example, a change from 20% to 15% is a reduction of five percentage points.
However, this result does not automatically mean that changing a customer's contract will reduce churn. That conclusion requires credible causal assumptions, including an appropriate graph and adjustment strategy.
The numbers above are illustrative, not results from an aitiolin analysis.
3. Which estimation methods are available?
The Reviewed comparison feature supports several approaches:
On smaller screens, scroll the table horizontally to view both columns.
| Method | What it does |
|---|---|
| Linear regression | Estimates an average predicted outcome difference between selected treatment conditions. |
| Logistic regression | Estimates a difference in predicted probabilities for a binary outcome. |
| Propensity-score weighting | Uses estimated treatment probabilities to construct weighted comparisons. |
| Propensity-score matching | Compares observations with similar estimated treatment probabilities. |
| Descriptive difference | Calculates an unadjusted difference between groups. This is not presented as an adjusted causal estimate. |
Where supported, aitiolin provides confidence intervals to describe statistical uncertainty.
A confidence interval does not account for every possible source of error, such as an incorrect causal graph or unmeasured confounding. Matching-specific confidence intervals are not currently implemented in the Reviewed comparison feature.
4. What should you check before interpreting a result?
A useful causal analysis goes beyond the estimated effect. Review the following evidence:
- Causal graph: Are the proposed relationships and their timing plausible?
- Overlap: Are comparable treatment and control observations available?
- Covariate balance: Are relevant measured differences between the groups adequately addressed?
- Robustness and sensitivity: How dependent is the conclusion on the chosen estimator, graph and assumptions?
Warnings highlight potential problems. The absence of warnings does not establish that the causal assumptions are correct.
5. Can you reproduce the analysis?
aitiolin records relevant analysis information, including the graph, treatment definition, configuration, cleaning history, software environment and results.
You can export recorded analyses as JSON or ZIP bundles. These exports do not include the original CSV data. Restoration requires matching local data and a review of the saved assumptions.
Data are processed and stored by the Django server. Processing is local when that server runs on your own computer. The Reviewed comparison feature does not require an external AI provider.