Only your friends list, and only for a tower you explicitly joined with. We read it through tinyburg.app, which already holds your linked account, so this study never sees your Nimblebit credentials.
This works for TinyTower and TinyTower Classic today. The study lists Nimblebit's other games too, but cannot yet decode their saves, so there is nothing it can read from them.
Your tower's contents are not part of the study. Bitizens, floors, coins, and bux are none of our business.
A connection between two players, and only when both of them have joined. If someone in your friends list has not joined, they are not recorded, not even anonymously.
We also store a count: how many friends you had in total, against how many were taking part. That count is what lets an analysis say how much of the real network it is looking at. It names nobody.
Your dashboard draws your circle as a graph: your towers, the people mutually friended with them, and the connections among that group. Everyone in it is taking part, and nobody outside your own circle appears.
That last part is worth being exact about, because it goes one step past a list of your friends: if two people in your circle are friends with each other, the graph shows that. Both of them joined the study, and both are already shown to you individually, but neither was asked specifically about this. If you would rather not be drawn in someone else's picture, withdrawing is what stops it.
Nimblebit gives each game its own player numbering, so the same friend code can belong to two different people in two different games. The study treats a tower as a game and a code together, and never connects one game's player to another's.
Taking part with one tower says nothing about your others. Each is joined, read, and withdrawn on its own.
Because a connection needs both people, this is a picture of the network among volunteers, not a picture of TinyTower. Statistics computed on it directly, like how many friends the average player has, will be skewed by who chose to take part.
That is a genuine limitation and not something we can fully correct for. It is why the friend counts above are recorded, and why any published analysis should state the sampling rate alongside its results.
Withdrawing removes every record that mentions you: the connections, the history behind them, the counts, and the permission itself. The published views are refreshed as part of the same operation, so you disappear from exports too.
Deletion runs as a durable job, which means it finishes even if something crashes partway through, and it records that it completed. It does not quietly half-happen.
Only the Tinyburg account that proved it owns a tower can enroll it or withdraw it. Knowing somebody's friend code does not let you take part on their behalf, and it does not let you delete their data either.