The question probably is why a 32-bit type was deemed sufficient back then, but as others have noted, it is very hard to anticipate the lifetime of software and it was a pragmatic tradeoff between anticipated OS lifetime and necessary storage space. There are other problems as well, though. time_t is a type designed for keeping track of the current time of a computer system, therefore an arbitrary cutoff point in the past is sufficient. But then it got used for all kinds of other things – storing people's birthdays, using it as internal representation for a DateTime type, etc. Those have been problematic ever since because a possible date range of 1901–2038 is not necessarily sufficient for those use cases.
Fact is, though, that for keeping time in a computer a simple scalar that counts units since a reference date is the easiest and most efficient solution. Over time there have been plenty of such reference dates (epochs) [½], the most notable these days probably Unix (1970-01-01), popular spreadsheets (around 1900-01-01) or Windows (1601-01-01).
[½] http://en.wikipedia.org/wiki/Epoch_(reference_date)#Notable_...
I still don't know enough about OSs to know how feasible it is, but back then, the move towards 64 bit OSs seemed like a great opportunity to make this change. Maybe when we move to 128 bit systems we can finally do this? I'd like not to postpone this until the last minute like last time.