The question is valid, but it has somewhat of a 'backwards' scent.
Normally, you conduct a research (eg, card sorting) with some research goal in mind. The goal will account for the various personas. In other words, you recruit participants based on the fit criteria defined by the research goal which accounts for personas.
Consider for example a site (or on app) that has two main personas - N (for domain novice) and E (for domain experts). You also know that the site usage share between them is 20% for N, and 80% for E. For research validity sake, I'd argue you should have in your card sorting the same amount of participants type N and E (say 12 for each group). In the analysis stage, you separate the two groups - meaning you analyse each separately. Great discrepancies in the results between the two groups could suggest an audience based IA; but if such path is not to be taken, group E (80%) should have more weight on the combined results.
If you only segment your participants after the card sorting, you may run into some troubles - what if 20 participants belong to group N, and only 4 to group E? That's just to warn against such practice.
But I think the general principle to follow is that the primary persona group gets more weight compared to secondary ones (or any additional persona classification).