Analytical Framework: We used a multistage budgeting framework which extends the idea of an exhaustive expenditure system to different levels or stages. This framework addresses a common problem in empirical estimation of system demand models requiring a sizeable number of equations, given the wide variety of consumer goods jointly purchased by households (Blundell, Pashardes, and Weber 1993; Fan, Wailes, and Cramer 1995). Specifically, a full demand system containing all consumer goods warrants a huge number of own- and cross-price parameters that are impractical to estimate under the constraint of limited data. one solution is to estimate the model in stages, whereby expenditures on goods belonging to broad food categories are incorporated in the model by estimating them sequentially. We applied a three-stage budgeting approach that estimated food and fish expenditure functions in the first and second stages, respectively. In the third stage, a system of demand equations for fish by species type was estimated using a quadratic extension of the AIDS (QUAIDS) model, which has recently proved popular (Blundell, Pashardes, and Weber 1993; Garcia, Dey, and Navarez 2005). The quadratic nature of the QUAIDS model captures the non-linearity in the consumption behavior of households for goods. At the same time, it relaxes the restriction imposed by linear demand functions regarding the al-location of marginal expenditures among commodities, assuming them to be the same in rich and poor households (Beach and Holt 2001). Such assumptions limit the classification of goods into either necessities or luxuries and deny the possibility that some goods may be luxuries at a low-income level and necessities at a higher income level (garcia, Dey, and Navarez 2005). A schematic diagram depicting the three stage-budgeting frameworks for eight species groups in Bangladesh is displayed in figure 2. In the first stage, the households are assumed to make decisions on how much of the predetermined income is to be allocated for food expenditure. Food expenditure is considered to involve a function of income, prices of food and non-food items, household characteristics (such as family size), and a set of dummy variables [time (months), districts, and urban/rural divide]. In the second stage, each household allocates a portion of the food expenditure for fish consumption, the amount of which is assumed to be dependent on prices of different food items (e.g., fish, cereal, meat, chicken, eggs, milk, vegetables, spices, pulses, oil) and other dummy variables as specified in the first stage. Finally, the expenditures for different types of fish are estimated in the third stage, using the prices of different types of fish, per capita fish expenditures, and dummy variables as mentioned earlier. It is likely to have zero consumption of specific fish categories, which may be due to any of the three broad factors: i) variations in preference across samples (households may simply not consume some species); ii) infrequent purchasing; and iii) misreporting (Keen 1986). In addition, due to the seasonal variability, the supply of a particular fish may not be available in the market; therefore, certain species may not be consumed. Various models for dealing with zero observation problems have been proposed by Deaton and Irish (1984); Keen (1986); Blundell and Meghir (1987); Heien and Wessells (1990). We used the approach applied by Heien and Wessells (1990) to deal with zero observations in the sample, as this is especially related to the AIDS model function. our particular application involves a special case of the censored simultaneous equation model in which the dependent variables are censored by a sub-set of unobservable latent variables. The dependent variables, which are budget shares of the species of fish, are either zero or some positive amount for each household. Those shares of zero values are censored by an unobservable latent variable that induces the decision not to purchase that particular item during the survey period. The decision to buy or not to buy can be indicated by a binary indicator variable, which is a function of the latent variables and is estimated using the Probit model (lee 1978). The assumptions underlying this model (and its proof) are that: i) the individual observations are independently and identically distributed and ii) the error terms are approximately normal with zero mean and a finite variance-covariance matrix that is constant overall observations.