This depends on weather you are worried about being convicted, or dealing with probable cause (in the U.S.).
Let's assume that you are at home. You start up your VPN and connect to your offsite VPN provider. If I am monitoring outgoing traffic (from your house), I know that you just connected to a certain IP address, and that the IP address is a VPN provider. Everything inside the payload of the packet is encrypted.
You then decide while at home, to check your email. I happen to be monitoring outgoing traffic from the VPN provider (which is not encrypted). I record it all using snort, and run Wireshark against the output. I see a connection to your email address and an email written. This may be protected by SSL if it's webmail. If it's regular email, it's likely plain text. If it's not plain text, I can try to intercept it at the receiver. The email is of no legal significance (i.e. you aren't using it to plan something unlawful). However, I make note of the fact that you confuse the use of their, there, and they're. I'll also notice a few idioms you like to use.
Over the course of monitoring outgoing traffic I see your account write several emails. I note patterns of misspellings, and more figures of speech. I collect these over a period of a month or two.
I then put the items that I notice into Wireshark. I add several things that you are known to say. Every time a misspelling occurs, or the use of an idiom (that you use) is found in the content of ANY packet that is outgoing from the VPN service you use, I view it.
Given another month or two I have a lot of data points. Some are sites you went to, others are not. The first thing I do is eliminate all of the data points that exited the VPN service provider while you were NOT on line (i.e. I didn't see you online from home, remember I started by monitoring that connection).
Then I look at the remaining traffic and see if I have any cluster points. Lots of recurring themes. Same subject matter over an over. I compare that to your unencrypted traffic, and your email.
I haven't applied enough filters to isolate you from the noise (people that use the same idioms/spelling errors you do), but I would have a good case for probable cause. If I have enough points of reference, it is just like a fingerprint.
Essentially I'm applying a Bayesian analysis to a corpus of work, to state something about the likelihood that I believe an exemplar to be a member of the set constructed by my suspect. The collection of works that I would compare to comes from any work that the suspect has publicly acknowledged they are responsible for. That analysis is well-known (and there's a whole statistics stackexchange site, too).
I'll let you answer, what would I come up with at this point?