Sample size
MDE for Thai marketplace apps
A textbook 2% relative lift is a comfortable default until your daily active base reshapes around tourism weeks, prepaid top-ups, and payday. Sample Size Without Theatre exists because that default keeps walking into Phuket classrooms uninvited.
Minimum detectable effect is not a vibe. It is the smallest change you would still spend engineering time to ship. If you would not rebuild a checkout badge for a 0.4pp move in trial-start, do not write 0.4pp on the form because a calculator asked for a smaller number.
Traffic is not a coin flip here
Marketplace apps in Thailand often see weekday troughs, weekend grocery spikes, and multi-day dead zones around major holidays. Averaging those into a single “daily N” understates variance. Your power calculation then lies in a friendly way: it tells you two weeks is enough when four weeks, excluding Songkran, is the honest runtime.
We ask students to plot daily exposed users for the last quarter before they declare MDE. If the plot looks like a coastline, they do not get to use a flat-mean formula without a caveat in the memo.
Relative versus absolute
Relative lifts flatter low baselines. A 3% relative move on a 2% checkout conversion is 0.06pp. That may be real and still not worth a badge that raises refunds. We prefer absolute pp on the primary when the baseline is small, and we write both when the PM is used to percentages.
Paywall Letters uses the same rule on trial-to-paid. Subscription teams love relative language. Finance pays in baht. The readout should be able to speak both without swapping the decision.
What we tell Floor Week teams
If your warehouse cannot produce a stable daily exposure count, you do not have an MDE problem yet. You have a logging problem. Variant Readout Desk will say that in week one. We would rather delay an intake than watch a team invent a sample size from a broken join.
The calculator is still useful. It is not the author of the bet. The author is the person who will still work here when the holdout ends.