Driftline

The wrong number

Digital-wellbeing tools almost all put one number in front of you: hours and minutes of use. I started from a measurement problem with that number.

It is poorly measured — a 2021 meta-analysis (Parry et al., Nature Human Behaviour) found self-reported phone use correlates only modestly with logged use, and the gap is worst exactly for the "problematic" use these apps care about. And even measured well, it barely predicts anything: Orben & Przybylski (2019), across 350,000+ people, found technology use explains on the order of 0.4% of variance in wellbeing.

This continues a thread that runs through my research: subjective appraisal tends to predict psychological outcomes more strongly than objective behavioral measures — the same structure as fear-of-crime research, where perceived risk predicts fear far better than actual victimization risk. Driftline applies that lens to digital life.

Why the existing apps fall short

I ran a competitive analysis of the intentionality-app category (One Sec, ClearSpace, Opal, ScreenZen, Intently, and others). Two patterns repeat: they measure quantity (the noisy, non-predictive number), and they hand the work back to the user through timers, blocks, and willpower — treating the problem as a discipline failure rather than a product engineered to capture attention.

Importantly, the "pause before you open the app" mechanic is already well-established and even has a published study behind it, so I deliberately did not claim that as a contribution. The real gap is narrower: no app in the category treats appraisal as longitudinal data, pairs an intentionality read with an affect read, or closes the loop into a measurable research instrument.

The concept & design

Driftline replaces the minutes dashboard with two objects. First, an appraisal moment — a two-tap, experience-sampling check-in that captures a session on two axes:

Second, the mirror — instead of reporting volume, it reflects the shape of your sessions back to you. Each session is a point in a 2×2 field, and two perceptual encodings carry the meaning without a legend: warmth = affect (restored sessions glow warm, depleting ones go cold) and blur = drift (sessions you drifted into render literally out of focus; chosen ones are sharp).

The design intent is that the measurement itself is the intervention — a brief, non-judgmental moment of noticing — while the reflection surfaces something non-obvious. A functional interactive prototype of this loop is built (React); the production target is React Native / Expo.

The research design

The concept is only worth building if the underlying hypothesis holds, so I designed a study to test it rather than asserting it: a within-person experience-sampling study — signal-contingent prompts ~4×/day for 14 days, with baseline and exit measures. Because each participant is their own control across pings, every stable confounder (personality, baseline wellbeing, chronic use level) is automatically held constant.

I then wrote a full pre-registration using the field-standard ESM template (Kirtley et al., 2021), which locks the decisions most vulnerable to after-the-fact flexibility:

• 01
  A single primary hypothesis

  Within-person appraisal → wellbeing; secondary hypotheses Holm-corrected, exploratory ones labeled — so no result can be retrofitted into a headline.
• 02
  Person-mean-centering, fixed in advance

  Separates within-person effects (causal-relevant) from between-person effects (confound-prone) — the move that makes the analysis credible.
• 03
  Locked exclusion rules & a stopping rule

  Inclusion thresholds and a fixed sample/stopping rule set before any data exist; no optional stopping.
• 04
  Multilevel mediation

  A 1-1-1 model of use → appraisal → wellbeing — the repeated-measures evolution of the mediation analysis from my master's thesis.

A subtle but critical decision: the appraisal of a session and the wellbeing outcome are measured as separate items with different referents, to avoid a circular model that predicts wellbeing from wellbeing.

The analysis pipeline

I built the full analysis pipeline in R (lme4 / lmerTest) and validated it by simulating a dataset to the pre-registration's exact shape (50 people × 14 days × 4 pings, ~65% compliance) with known true effects, then checking recovery. Every pre-registered estimate landed on its planted value within sampling error.

Pipeline validation — full output

Recovery of known parameters and within/between power, from the reproducible R pipeline on simulated data.

What this demonstrates

Honest status & roadmap

Reflection

Designing the study taught me more than running a quick survey would have. Forcing myself to separate appraisal from outcome, to pre-register a single primary hypothesis, and to validate the analysis on simulated data before touching real participants is a discipline I'll carry into applied research work — it's the difference between a finding you can defend and one you merely hope for.