When does the boundedness of the dependent variable become problematic in linear regression? Linear regression assumes that the dependent variable ranges from $-\infty$ to $\infty$. Many (most? all?) real DVs do not actually have such a range. For instance, the weight of adult male humans can't, even in theory, be negative and, for practical purposes, is bounded by about 50 and 1500 pounds (or roughly 30 and 1000 kg ... I am not concerned with he exact bounds).
When does the boundedness become problematic and what problems does it cause?
EDIT For the assumption of unboundedness see e.g. Berry Understanding regression assumptions from Sage.
 A: There is an issue as you approach a bound; it might not be an issue far from a bound.


*

*Linearity. 
Clearly if you can't cross a boundary but a linear relationship would cross the boundary you can encounter problems with linearity - the true relationship must at some point stop increasing or decreasing at the same rate as you move the predictors in a way that takes you closer to the boundary (or it must flatten abruptly); typically relationships start to curve noticeably before you get really close. This curving is familiar in binomial (e.g. logit and probit links) and Poisson regression models, but it also occurs in models for continuous variables.

*Constant variance
In addition as the conditional mean approaches the boundary, the variation about the mean will naturally reduce; there's less room between the mean and the boundary but the position of the mean prevents too much variation away from it on the other side from the boundary (after all, that would pull the mean up).
A bound often comes with a relationship between mean and variance. Many positive or non-negative variables have spread that continues to increase as the mean increases, even when far from the boundary; it's not always the case that issues related to having a bound only occur near that bound.

*Skewness
The shape of the conditional distribution of the response would typically become more skewed away from the boundary (e.g. becoming more right skew as you approach a boundary from above); as the mean squeezes closer and closer to the bound, the observations on that side will tend to be closer to the mean than the ones on the other side with don't get "squeezed" in the same way.
Naturally the degree to which these things are an issue depends on a host of factors (the kind of model, the nature of the variable(s), the nearness of the bound(s), and much else besides -- even the tolerance of the person doing the fitting); a more specific situation might allow a some investigation on that part.
