Since its economic reform, China has changed significantly as it makes its transition from a centrally-planned to a consumer-oriented economy and thus has gradually increased household income and changed consumption patterns in urban China. This study attempts to provide an in-depth understanding of heterogeneous consumer patterns in urban China by developing a multi-stage censored demand system using household data. Specifically, this study develops an economic model considering heterogeneous consumption patterns across households and commodity groupings and estimates econometric models of a Quadratic Almost Ideal Demand System (QAIDS) using household data. Three methodologies are integrated including constructing a multi-stage demand system, incorporating demographic variables using the 'ordinary budget share scaling and translation' (OBSSAT), and employing a two-step estimator to deal with zero consumption problems. This study covers three provinces in China, Shandong, Jiangsu, and Guangdong, and uses household data from 1998 provided by the National Bureau of Statistics (NBS). Based on the Chinese food guide pyramid, a three-level utility tree is constructed dividing 18 food items into five subgroups. An empirical analysis is conducted by estimating econometric models to examine the impact of the potential factors, e.g., income and demographic variables, on food demand. The results show the uniqueness of this study in three dimensions. First, using the OBSSAT helps answer the question of "how to break down the heterogeneous consumption patterns in urban China?" In addition, our findings also show that China should be treated as several markets instead of one. Second, the QAIDS has not previously been applied to the study of food demand in urban China. Our results show that the QAIDS is superior to the AIDS; however, the degree of importance for the quadratic term decreases as demographic and censoring effects are considered in a demand system. Finally, 18 food items are broken down into five food subgroups and are estimated by a multi-stage censored QAIDS. Including this large food bundle in a demand system provides us detailed information of the relationship among food items.