Growth models are important components of population biology study and are generally essential to adequately assess the impact of fishery. Given a specific functional form, the appropriate estimation of growth parameters depends on the error structure assumed for the data. For example, if variability in size is constant as a function of age, an additive error structure is suitable. However, if the variability in size increases with age, a multiplicative error or variance modeling is appropriate. Variance heterogeneity will typically not influence the parameter estimates significantly, but if ignored it may result in severely misleading the standard error and prediction intervals. The four parameters model formulated by Schnute contains a number of specific growth models that can be used to explain the pattern of growth in small yellow croaker(). We used data transformation and variance modeling to investigate the effect of assuming a different error structure in the model. We used data from stow net surveys conducted between May–September in 2007–2008 and from bottom trawls conducted in the northern region of the East China Sea between October–April in 2007–2009. We used the likelihood ratio test ( distribution) and Akaike’s Information Criterion to quantitatively compare the fit of nested submodels. Error structure had a significant effect on the fitted models. The estimated parameter values for the lognormal error, power variance. Furthermore, relatively small standard errors and narrow confidence intervals suggest that the integration of variance structure in the growth models is more accurate and robust than in the additive models. The log-transforma­tion models and variance structure models fit the data better than the additive models. The funneling observed in the plots of deviance residuals against age for the additive models was reduced in the corresponding plots for the lognormal error and variance structure models. The power variance and exponential variance models yielded significantly different estimates than the additive models