Table pa has 7522676 rows, 4834042 of which have field1 NULL, so it's absolutely not reasonable to expect this to be an optimal strategy.
It occurred to me that even though the majority of values are NULL, there are
1691 unique values in pa.field1, so I suppose it might seem more attractive to the planner than it should do (that's more unique values than there are scdate entries).
I might just set enable_seqscan to false and leave it at that. It makes me unhappy though.