Practically no surveys have forced participation, a government census is about the only one I can think of. Yet we still use them all the time because they have proven to have useful and accurate information. You would have to show the people opting out of the steam survey are some particular sub group that effects the results (e.g. radeon users). That is not a correlation I have ever seen.
There are statistical techniques that can be applied to account for a survey not being a random sample, but we know practically nothing about the steam survey other than what we have. Without raw data we can't even attempt to determine which methods should be applied.
There are a lot of other reasons why Steam surveys are bad for the kind of arguments they get used for and outside of them not being a random sample, almost all of the others come down to not having Valve's methodology which can lead to biased results in any number of ways.
The onus is always on the person presenting the statistical data (or more broadly the claim being asserted) to show that their data is good, not that someone else needs to show it's incorrect. It's the same reason criminal trials don't require you to prove your innocence. It's up to the prosecution to prove your guilt. We don't let a pharmaceutical company put a drug on the market unless or until someone else could prove it isn't safe or that it does cause side effects.
Lack of random sampling is why most surveys are junk and you can easily find or produce one with different results. Just look at how practically every brand can claim to be the "most trusted" or something like that. Obviously that can't be correct, yet surveys which don't use a random sample of the population can produce different results.
Any claims outside of general popularity of cards used by Steam users who respond to hardware surveys needs a list of included assumptions being made about the data if it's going to be used to make some claim. If people did that they'd start to realize they're dealing with a lot of unknowns and making a lot assumptions that may even build off of each other.