What does the row estimate look like on the scan of data table with that statistic in place? Anytime the stats give a mis-estimate this far off, I wouldn't expect that plans would be optimal except by luck.
Index Scan using data_pkey on data t (cost=0.57..21427806.53 *rows=58785023* width=131) (actual time=0.024..0.482 *rows=854 *loops=1) Index Cond: (id > 205284974)
