Motivated by the need to analyze the National Longitudinal Surveys (NLS) data, we propose a new semiparametric longitudinal mean-covariance model in which the effects on dependent variable of some explanatory variables are linear and others are nonlinear, while the within-subject correlations are modeled by a non-stationary autoregressive error structure. We develop an estimation machinery based on least squares technique by approximating nonparametric functions via B-spline expansions, and establish the asymptotic normality of parametric estimators as well as the rate of convergence for the nonparametric estimators. We further advocate a new model selection strategy in the varying-coefficient model framework, for distinguishing whether a component is significant and subsequently whether it is linear or nonlinear. Besides, the proposed method can also be employed for identifying the true order of lagged terms consistently. Monte Carlo studies are conducted to examine the finite sample performance of our approach and an application of real data is also illustrated.