The statistics
Correlations between your habits, your sleep and your mood
This is the part of Crescendo that justifies the rest of it. Tracking is only worth the effort if something reads the data back to you and says something true about it.
What is actually computed
For a habit you either did or did not do, and an outcome measured on a scale — your mood score, your hours asleep — the appropriate measure is a point-biserial correlation. For two continuous measures, such as sleep duration against next-day mood, it is a Pearson correlation. Crescendo computes whichever fits the pair, and the p-value comes from the exact t-distribution rather than a normal approximation, which matters at the sample sizes a single person's data actually reaches.
The output of that is a coefficient between -1 and 1, and a probability. Neither is shown to you on its own, because neither means much alone.
The three floors a finding has to clear
A relationship is reported only if it passes all three of these at once. Any one of them alone is easy to pass by accident.
At least 20 paired days. A correlation computed on six days is not a weak finding, it is noise with a number attached. Twenty is where the estimate starts to be worth reading, and pairing matters — a day counts only when both sides of the comparison were recorded.
A coefficient of 0.25 or stronger in absolute value. A relationship can be statistically detectable and still far too small to change any decision you would make. This floor is about whether the effect is worth your attention, which is a separate question from whether it is real.
A p-value of 0.00625 or lower. That is 0.05 with a Bonferroni correction applied for the eight comparisons run. Testing eight hypotheses at p ≤ 0.05 gives you roughly a one-in-three chance of at least one false positive, and an app that runs dozens of comparisons and reports whatever clears 0.05 is guaranteed to find something. Correcting for the number of tests is the difference between a result and a coincidence.
The hypotheses are fixed in advance
This is the part that is easy to skip and does most of the work. The set of relationships Crescendo tests is decided ahead of time, not searched for after the data arrives.
The alternative — computing every pair and surfacing whatever looks strongest — is called trawling, and it will always find something, because with enough comparisons something always clears any threshold you pick. The correction above only means anything because the number of tests is known in advance rather than chosen to fit the answer.
What it looks like when there is nothing
It says so. "Nothing significant yet" is a designed state, not an empty state — the Insights screen is built to show it, and it is the honest answer most of the time in the first month.
Every card carries its coefficient and its sample size whether it cleared the floors or not, so a near-miss is visible as a near-miss rather than disappearing. An app that always finds a pattern is not measuring anything, and the willingness to return nothing is what makes the times it returns something worth acting on.
At a glance
- Methods
- Pearson and point-biserial correlation
- Significance
- Exact t-distribution p-values
- Minimum sample
- 20 paired days
- Effect-size floor
- |r| ≥ 0.25
- Corrected threshold
- p ≤ 0.00625 (Bonferroni, 8 tests)