How to Find Out What's Actually Affecting Your Sleep and Mood

7 min read · Updated 2026-07-28

To find what moves your sleep or mood, compare the days a factor happened against the days it didn't, then check whether the gap is bigger than the noise. The catch is that testing many factors at once manufactures false positives, so the comparison needs a multiple-testing correction before you act on anything it says.

You already suspect something is wrecking your sleep. Caffeine after 2pm, the late workout, the second glass of wine, the day you skipped lunch. The suspicion is easy. Confirming it is where people give up, because a line chart of sleep hours next to a line chart of anything else mostly just looks like two wiggly lines.

The method that works is boring and it is not new. Split your days into two piles. Days the thing happened, days it did not. Compare the average outcome across the two piles. That is it. Bearable popularized this for symptom tracking and it is the right shape for the question.

The part almost everyone gets wrong

Here is where it falls apart. Say you track twelve factors and check each one against sleep, mood, and weight. That is thirty-six comparisons. At the usual 5% threshold, you would expect roughly two of them to look significant by pure chance even if nothing in your life affects anything.

So you get an insight card telling you that Tuesdays are ruining your sleep, and it is real in the sense that the math checked out, and it is meaningless in the sense that you rolled the dice thirty-six times and this is what came up. Then you rearrange your week around it.

This is the multiple comparisons problem, and it is the single biggest reason personal-tracking insights feel unreliable. The fix is a correction that adjusts the threshold for how many tests you ran. HabitSync uses Benjamini-Hochberg across every comparison in the batch, not per test. A finding is only labeled strong if it survives that correction and its confidence interval excludes zero.

Why sample size decides everything

A four-day pile against a three-day pile can produce a huge gap and mean nothing. Twenty-five days a side producing a small gap can mean quite a lot. Raw difference is the wrong thing to sort by, but it is what most apps show you, because it makes for a better-looking number.

Ranking on the conservative end of the confidence interval fixes that. If the interval for a factor's effect runs from 0.2 to 1.8 hours, the honest headline is 0.2, not 1.0. And if the interval crosses zero at all, the data cannot even establish the direction, so the finding gets pushed down rather than dressed up.

The practical effect is that a flashy four-day fluke stops outranking a pattern you built over a month.

What you need to log

The comparison needs both piles to exist. That sounds obvious and it is the thing people get wrong, because it is tempting to only log the good days. If you record your workouts but not your rest days, the analysis has nothing to compare against.

Two or three weeks of daily logging is roughly where this starts to be worth reading. Before that the piles are too small and everything comes back weak, which is the correct answer, not a bug.

One outcome you genuinely care about. Sleep duration, restedness, or mood are the usual ones.
A handful of candidate factors, not everything you can think of. More factors means a harsher correction and more days needed.
The boring days. A day with nothing notable is data, not a gap.

Reading the result without fooling yourself

Two habits will keep you honest. First, look at the sample size before the effect size. Second, treat a weak label as a genuine result rather than a broken feature. Weak means the evidence is not there yet, and the useful response is more days, not a different app.

It is also worth muting the pairs you have already chased down. Once you know that late caffeine costs you sleep and you have decided what to do about it, having that card resurface every week just buries the things you have not looked at. Muting a factor and outcome pair, or excluding a metric from analysis entirely, is what keeps the list about new information.

This finds patterns, not causes

A comparison like this cannot tell you which direction the arrow points. Bad sleep might be causing low mood, low mood might be costing you sleep, or a rough stretch at work might be driving both. The statistics do not resolve that and no app can resolve it for you from observational data.

What it does give you is a much shorter list of things worth testing deliberately. Change one, hold the rest, watch what happens over the next couple of weeks. That is a real experiment, and it goes a lot faster when you start from three plausible candidates instead of twelve.

The goal is not certainty. It is spending your attention on the two or three things that have some evidence behind them, instead of the one you happened to notice on a bad Tuesday.

Keep reading

Why Your TDEE Calculator Number Is Wrong (and How to Fix It With Real Data) — TDEE calculators give you a population-average estimate. Here's how far off it can be, why it drifts as you diet, and how to replace it with your own data.
How to Tell If a Medication Is Working (and What Else It's Doing) — Many medications take weeks to work, and population side-effect rates say nothing about you. One before/after method answers both questions from your own data.
Why Just Tracking a Habit Can Start Changing It — Noticing an automatic behavior is often what breaks its grip, before you try to change anything. Here's why logging a habit works even on days you don't act.
Why Your Habit Tracker and Your Pill Organizer Should Be the Same App — Most people log habits in one app and medications in another. The interesting answers live between those two datasets - and splitting them hides them.
Never Miss a Dose Without a Single Alarm: The Visibility Method — Reminder apps assume more alarms mean better adherence - until dismissing the alarm becomes the habit. Try routine anchors and visible dose counts instead.
How Long It Takes to Form a Habit: 18 to 254 Days — The 21-day rule has no study behind it. The research people cite found habit formation took 18 to 254 days, averaging 66. Here's what moves you in that range.
Does Habit Tracking Work? What the Evidence Says — In an NIH-funded trial of nearly 1,700 people, those keeping daily food records lost twice as much weight. Here's what tracking does, and where it stops helping.
Why Habits Carry You on Your Worst Days — When willpower runs out, people don't make worse choices - they make more automatic ones. Depletion raised habitual choices 28-32%, good habit or bad.
Why Old Habits Come Back (and What to Do Instead) — Habits don't get erased. Extinction reversed the neural signature of a rat's habit, then it returned the instant retraining began. What that means for relapse.
Why Sharing a Streak Isn't the Same as Sharing Progress — Most habit trackers with friends just mirror checkmarks. HabitSync Groups compare the real numbers behind them, and only what you choose to share.
Why Rigid Habit Trackers Don't Survive Contact With an ADHD Brain — HabitSync wasn't built specifically for ADHD, but flexible goal types, honest miss tracking, and a history that never resets happen to fit exactly where rigid trackers break down.
Why the Scale Stopped Moving When Your Diet Didn't — A stall usually means your maintenance calories moved, not that you lost discipline. Here's how to measure where maintenance is now, using the intake and weight data you already have.
What Separates a Medication Tracker From a Reminder App — Most medication apps are reminder apps that log a checkmark. Here's what to look for if you want to answer whether a medication is working, and what a checkmark can never tell you.
Some Habits Don't Pay Off the Same Day — If you only compare today's habit against today's mood, anything with a lag looks like it does nothing. Here's why delayed effects are common and how to test for them without fooling yourself.
Sharing Progress Without Sharing Your Weight — Accountability groups usually mean handing over numbers you would rather keep private. Here's a sharing model where every metric is off by default and weight can be shared as percent change only.