Did it actually cause it?

How many extra weekly signups did the launch bring?

The data. standard_causal_impact_v2.csv, the published sample file for Causal impact: 104 rows; columns week_start, region, weekly_signups, phase, notes. Sample data, not a customer's.

The calculation. Python, standard library only. This exact code ran against the file when this page was built, and printed the output below it.

import csv
rows = list(csv.DictReader(open("standard_causal_impact_v2.csv")))
rows.sort(key=lambda r: r["week_start"])
def series(phase):
    return [float(r["weekly_signups"]) for r in rows
            if r["phase"] == phase and r["weekly_signups"]]
pre, post = series("before launch"), series("after launch")
mean = lambda v: sum(v) / len(v)
print(len(pre), "weeks before,", len(post), "weeks after")
print("naive lift:", round(mean(post) - mean(pre), 1), "a week")
x = range(len(pre))  # straight line fitted to the weeks before
mx, my = mean(x), mean(pre)
b = (sum((i - mx) * (v - my) for i, v in zip(x, pre))
     / sum((i - mx) ** 2 for i in x))
proj = [my + b * (len(pre) + i - mx) for i in range(len(post))]
print("trend before launch:", round(b, 2), "signups a week, per week")
print("lift over the projected trend:",
      round(mean([v - p for v, p in zip(post, proj)]), 1), "a week")

60 weeks before, 44 weeks after
naive lift: 47.6 a week
trend before launch: 0.51 signups a week, per week
lift over the projected trend: 21.1 a week

What it says. Comparing the average week after launch with the average week before says 47.6 extra signups a week. But signups were already rising by 0.51 a week before launch. Projecting that trend forward and measuring against it leaves 21.1 extra a week, 44% of the naive figure. On this reading the rest was already coming.

What it does not say. A straight line is the simplest possible counterfactual, and anything else that changed at launch is mixed into the lift. The library analysis fits the jump and the change in trend by segmented regression, allows for weeks that depend on the week before, and puts a 95% prediction interval around the effect.

Run Causal impact on this file or your own, or open its stored sample report.