Diff-in-Diff is a statistical method used to estimate the effect of a campaign or intervention. It compares how outcomes change over time in a treated group (exposed to the campaign) versus a control group (not exposed).

By looking at the difference in the changes between the two groups, we isolate the campaign's true impact — separating it from normal trends or seasonality.

Each row in your data represents one observation for one region (or group) at a specific time — either before or after the campaign.

For each region, you must provide:

• Group ID (e.g., "Paris", "Region_1")
• Pre (0) or Post (1) campaign indicator
• Treated (1) or Control (0) indicator
• Outcome (e.g., conversions, sales, clicks)

👉 Important: Each region must have at least two rows — one for pre and one for post.

These columns use 0 or 1 to indicate status:

• Pre 0 / Post 1:
  • 0 = before the campaign
  • 1 = after the campaign
• Treated 0 / 1:
  • 1 = group was exposed to the campaign
  • 0 = group was not exposed (control)

• Group ID: This is the name of the region or group you're analyzing — it can be anything like Paris, Store_12, or A.
• Outcome: This is the result you're measuring. It must be a numerical value, such as:
  • Sales
  • Sign-ups
  • Clicks
  • Revenue

Yes — your groups can have multiple rows per time period (pre or post).

However, it's important that each group has the same number of rows per period. For example, if one region has 3 rows pre and 2 rows post, it will be excluded.

💡 This ensures fair comparison and valid statistical inference.

The tool estimates the Average Treatment Effect (ATT), which is the average difference in outcome caused by the campaign per region.

For example: if the ATT = +20, it means that on average, each treated region experienced a 20-unit increase in the outcome compared to what would have happened without the campaign.

The confidence interval gives a range where the true effect likely falls:

• If both ends are above 0, the effect is positive and significant
• If both ends are below 0, the effect is negative and significant
• If the interval includes 0, the effect is not statistically significant

For robust statistical inference, try to include at least:

• 4 treated groups
• 5 control groups

If you have fewer treated groups, a special method (inspired by Conley & Taber) is used to approximate significance.

If there are too few groups overall, the confidence interval will not be available.

If you don’t have enough treated or control regions, the tool will still estimate the effect, but cannot calculate a confidence interval.

In that case:

• You’ll still see the estimated effect
• But we can’t determine if it's statistically significant

For better reliability, aim to include at least 4 treated and 5 control regions.

For the estimate to be reliable, a few assumptions must hold:

• Parallel trends: Control and treated groups should behave similarly before the campaign
• No major external shocks: Avoid events like holidays, weather changes, or national campaigns that could affect results
• Similar structure: Groups should be comparable — not dramatically different in size or behavior

If these assumptions are not met, your estimates may be unreliable. Consider applying more advanced modeling approaches in such scenarios.

The counterfactual line shows what the treated group would have looked like without the campaign, based on the trend observed in the control group.

It’s calculated as:

Counterfactual Post = Treated Pre + (Control Post − Control Pre)

This acts as a "what-if" scenario for visual comparison.

We want this tool to be as accessible and understandable as possible. If you're unsure about any concept or how to format your data, feel free to reach out — we'd be happy to clarify or guide you.