When the data and anecdotes disagree, the anecdotes are usually right, which means you should examine the data to ensure you're measuring the right thing.