It seems like you have a classical bayesian problem. You have some sort of prior distribution, a distribution over years of birth, your prior distribution is bimodal with peaks at the two years, you can probably use a convolution of two normal distributions to model this variable. Then have it spit out a posterior distribution after you feed in some data.
The real problem that I have with this analysis is it seems your features aren't particularly good. It is true these vars might have information about birth year, for example for the 20th century the average age of first marriage has steadily been increasing. But I suspect that the signal is going to be fairly weak. Essentially, if I tell you that I got married at age 24, had my first child at 26, and that my older brother is 3 years older than me and my younger sister is 2 years younger than me, can you tell me in what year was I born, 1956 or 1989?
I suspect that without additional data this information that I provided would be completely useless, mostly because it is a very noisy signal. That information could apply equally to someone born in 1956 or 1989. It isn't very helpful.
Essentially, what I am saying is that when you update your prior, it isn't going to change very much. (Your posterior distribution would look very similar to the prior distribution.) Instead of doing some mustache twirling over what is the right algorithm to crack this problem, I think a much more fruitful exercise would be to think up some better features.
