Many investigators have found that divorce risks decrease as you move
from groups with little educational or social capital to groups with m
ore. This negative educational gradient fits with the notion that peop
le with more education are better at selecting spouses and better at m
aking a marriage work. Other investigators have found a positive gradi
ent, often in populations where the situation is dominated by the indi
vidual's ability to handle the divorce process and to cope with the ec
onomic and other problems that follow in the wake of a divorce. The si
gn of the educational gradient in divorce risks seems to depend on the
balance between countervailing influences, Information about the grad
ient over a few educational levels is about as much as you can expect
to get from the interview data of a normal-sized general survey. With
access to the data from a full-coverage system of the population and e
ducational registers of a sizeable population like that of Sweden, edu
cational effects can be studied in much greater detail. We begin to ta
p this source in the present paper. When we do, the educational gradie
nt in divorce risks turns out to be too slippery a basis for the gener
al theories that have been developed around it so far, at least in a p
opulation where it is reasonably easy to get a divorce and where the h
urtful consequences to the divorcees are more limited than elsewhere.
There has been no uniform relation between educational level and divor
ce risk of Swedish women at the various educational levels during the
1970s and 1980s; developments in recent decades in Swedish first-marri
age divorce risks have been much more favourable to the more highly ed
ucated than to women with less education, and the result is that the e
ducational gradient has become negative as we leave the 1980s. The edu
cational gradient changed sign correspondingly between cohorts born in
the mid-1940s and cohorts from the mid-1960s. In a society such as Sw
eden, it may be more important to explain the trends in divorce risks
by educational level than to explain the gradient of educational effec
ts. Most normal-sized data sets are too small to permit the inclusion
of secular changes in the effect of education on divorce rates, so ana
lysts risk working with a seriously mis-specified model if real educat
ional impacts change over time.