Consider a semiparametric regression model Y = f (theta, X, epsilon), where
f is a known function, theta is an unknown vector, epsilon consists of a r
andom error and possibly of some unobserved variables, and the distribution
F(.) of (epsilon, X) is unspecified. This article introduces, in a general
setting, new methodology for estimating theta and F(.). The proposed metho
d constructs a profile likelihood defined on random-level sets (a random si
eve). The proposed method is related to empirical likelihood but is more ge
nerally applicable. Four examples are discussed, including a quadratic mode
l, high-dimensional semiparametric regression, a nonparametric random-effec
ts model, and linear regression with right-censored data. Simulation result
s and asymptotic analysis support the utility and effectiveness of the prop
osed method.