There are two answers to this question and they point in opposite directions. The interesting part is that the evidence supports both.
The one everybody knows
The widely-cited finding is about personality: people who report trolling more also score higher on a cluster of traits associated with enjoying others' distress. [1] In the work most often quoted for it, that reward — negative social potency — predicted trolling far more strongly than psychopathy or sadism did. [2]
It is worth being precise about how much weight that will carry. It is correlational and self-reported, and this page's evidence grade says so at the top. It rests on 396 adults, 75.9% of them women, answering questions about themselves. [3]
It is also the half of the picture that gets quoted, and it has an unfortunate effect on readers. If trolls are simply a type of person, then the people provoking you are a separate category, permanently, and the only question is how to identify them.
The one that changes what you do
The counterweight is a study that took ordinary users rather than self-selected trolls, and manipulated their circumstances instead. Its authors set it against the prior view that trolling is confined to a vocal and antisocial minority. [4] They proposed two triggers: the person's mood, and the surrounding context of the discussion. [5] Each raised the probability of a user trolling, and together they doubled it. [6]
It was done twice, in two different ways, which is the part worth knowing. The experimental half used a simulated online discussion. [7] The other half went looking for the same mechanisms outside the experiment, in a longitudinal analysis of a large online news discussion community. [8] A finding that survives both a lab and a live forum is worth more than one that only manages the lab.
The sentence that does the most work
The line that should change how you read a bad thread is not the doubling. It is this: in the authors' predictive model, mood and discussion context together can explain trolling behaviour better than an individual's history of trolling. [9]
That is a claim about which of two kinds of information is more useful, and it points away from the personality account.
It does not let anyone predict a particular reply, and this page is not offering a way to. What it changes is where the interesting information sits — and this site's reading of that, which is not a finding of the study, is that the state of a thread deserves more of your attention than the profile of whoever is in it.
Two things follow from that, and both are uncomfortable.
Trolling is partly contagious. Seeing it makes people more likely to do it. A thread does not contain a fixed number of hostile people; it produces more of them as it goes.
You are not exempt. The mechanism is mood plus exposure, and both of those apply to everyone, including the person reading this page while annoyed about an argument they are losing.
What the authors themselves said
Their own conclusion is hedged, and the hedges are the honest part: that ordinary people can, under the right circumstances, behave like trolls. [10] "Can", "under the right circumstances", and results that "combine to suggest" are three separate pieces of caution in one sentence.
The authors' own page for the paper shows one thing that happens to that caution downstream. It lists the press the study received, and one of those headlines reads "There's a troll inside all of us, researchers say". [11] It is a catchier sentence than the one in the abstract. It is also not the same sentence: the hedges are the claim, so a version without them reports something the study did not say.
Why this matters for what to do
If provocation were purely dispositional, the response would be identification: work out who the trolls are, exclude them, done.
If it is substantially situational, then the most useful lever is the thing you control — whether your reply raises the temperature for everyone else reading. That is a much less satisfying conclusion than "some people are just like that", and it is better supported.
It also reframes the advice this site is named after. Declining to reply is not only about starving one person of a reaction. It is about not adding to the exposure that makes the next person more likely to join in — which is the argument the page on that advice has to make on its own evidence.
A second line of work points the same way, and shares one of its authors: the question of whether a conversation heading for trouble can be spotted while it is still developing. How conversations turn toxic covers what that work does and does not show.