Disagree. Put enough anecdotes together and you'll get a pretty good handle on what's going on, even if not up to scientific standards. "Enough", however, might be quite a lot; as in more than ost people would have easy access to.
The problem is that anecdotes cannot, even in principle, be free of several
crippling biases in data collection: they're
inherently unsystematic, unrepresentative, self-selected, apophenic, and prone to confirmation bias, effectively personal publication bias (failure to report—"publish"—null/uninteresting results), and the misinformation effect (reports being altered, sometimes a lot, based on
how the persons are asked questions.) Hence, while it is "data" in the sense that any information whatsoever can be considered "data," it is data of such low and dubious quality that it cannot be treated as fitting or appropriate for any kind of serious analysis.
There's a reason no one would (or should) trust "anecdata" for medical treatment, for example. A friend telling you they got pain relief from some treatment is an anecdote, and useful
as an anecdote. Word-of-mouth reporting has a place and purpose. But that does not mean that getting self-reports from 5000 people about what pain relief worked for them would
ever fly as actual data.
Hence: the plural of "anecdote" is not "data." Yes, it's a pithy phrase lacking in nuance, but it carries the essential message: anecdotes are tainted by tons of bias that cannot be filtered out without doing the
hard, hard work of actual statistical analysis...at which point you may as well have just collected the data systematically to begin with and saved yourself the bother!
Character "career length" is a completely different - but still interesting - question. It would be interesting, if the data exists, to find out how many sessions each species-class combo gets played on average, to see if there's any overall trends toward higher or lower resilience between different options.
Such data
would be super interesting, but almost certainly doesn't exist in systematic form, so we'd be reliant on self reports...which as noted are rather low-quality data.
Further, this would struggle against the effect of campaign failure and/or premature conclusion. I've played in campaigns that
would have run for years, I'm sure, if the DM had not had major life events change their plans, forcing them to end the game. (Injured family member needing care in one case, and pseudo-adopting a nephew who needed a new place to live in another case, and a third where a key player had to stop for work stuff, so we wrapped up that campaign faster than originally intended.) Such effects would presumably not be particularly biased toward any one option (life happens!), but they would act to fuzz out any patterns that might exist, and with them likely being relatively subtle to begin with, anything that makes the picture more blurry is liable to make it too blurry to make out.
