Thread by @H_Shawberry

Out today: Our research in @PsychScience shows that daily smartphone use is both consistent over time and unique to an individual when studying 28,692 days of logged app usage from 780 people. A 🧵 doi.org/10.1177/095679…

We found that people exhibited consistent patterns in their application usage behaviours on a day-to-day basis E.g., always using Facebook the most and the calculator app the least each day.

We also found that daily smartphone use is unique to the individual, and can therefore identify them from a crowd. Watch this video to learn about this further: vimeo.com/677915730/7c82…

Must thank @crest_research @ProfPaulTaylor @davidaellis Stacey Conchie @lancspsychres @PsychologyLancs for making this work possible. Read on for more details:

When studying behavioural consistency, we took inspiration from interactionalist approaches. We analysed the use of 21 apps (used by at least 25% of the sample). For each day of data, we generated 21 z-scores (one per app). These 21 scores form a behavioural profile for that day.

Following the procedure of Shoda, Mischel, and Wright (1994) we studied the similarities of these profiles using ipsative correlations. Apps were ranked from the most used to the least used each day, and were then compared to see if the apps appeared in the same order each day.

The higher the correlation coefficient, the more similar the profile shapes were. You can compare two days of data from the same person (for a within-subject comparison) or two days of data from two individuals (for a between-subject comparison)

Please view our shiny app (next tweet) showing examples of behavioural profiles, and ipsative correlations between two days of data. It demonstrates both within and between subject comparisons across two variables: pickup behaviour profiles and duration behaviour profiles.

You can press “click here” buttons repeatedly to build a distribution behaviouralanalytics.shinyapps.io/AppUseProfiles/ The “normative profile” represents the sample wide average daily use of an app. Points above or below shows how an individual deviates from the norm.

We conducted 440 million between subject comparisons and 440 million within subject comparisons for the daily app pickups variable and daily app duration variable respectively (see graph)

If you examine two randomly selected days from the same person, this showed greater similarity in application use patterns than if you randomly selected two days that belonged to two different people. Effect sizes were large (Pickups: d = 1.70, Durations: d = 1.46)

To measure uniqueness in people’s app use, we fed 4,680 days of app usage data into random forest models. This equated to 6 days of data for each participant. As each day of data was paired with one of the 780 users, the model learnt people’s normal daily app use patterns.

We then tested whether models could identify an individual when provided with only a single day of smartphone app use data that was anonymous and not yet paired with a user. Models could identify the correct person from a day of anonymous data 1/3 of the time! (chance = 1/780)

It was possible to view the top 10 most likely individuals that a specific day of data belonged to. Around 75% of the time, the correct user would be among the top 10 most likely candidates.

Therefore, we have shown that logs of smartphone behaviours (and likely other digital traces) are consistent over time, and unique! Simple digital meta data can be used to identify a user, even when not logged in to their digital accounts.

You can view our open science page here for processed data, code, preregistrations, supplementary materials, and wiki osf.io/xvd6s/

This work is also being presented this week at #SPSP2022 in the On Demand Symposium: Analyzing Digital Human Behavior: The Shape of Psychology to Come