Statistical sampling works if the sampled people answer your question. It also helps if you can find the contact of the samples to begin with. Now, here's some ways the statistical sampling can be substantially biased: people who are still repaying their student loans are in Princeton's database, and they can be contacted; those who didn't take loans (the rich kids) are not, so negative bias. Opposite: Princeton alumni with kids, who hope to get their kids into Princeton, donate. So they are in a database, and they end up in the statistical sampling. This is positive bias. Back to negative bias: some Princeton graduates end up unemployed (generally not because they are unqualified, but because of burnout/depression/family problems). Do you think they are in the mood to answer the phone and disclose their (non-existant) current income?

In principle you can correct for a known bias. You can correct for several, but the more adjustments, the more noise in the estimation.

Now back to my original remark: do you honestly believe only 23% of the Princeton graduates are in the top 1% of the earners?