Better models by discarding data?