PI: Aviv Nevo
One of the most basic notions in economics is that of a demand function. This function maps prices, as well as other features of products (e.g. advertising), to quantities demanded of these products. Knowledge of the demand function, or a reasonable estimate of it, allows us to answer questions regarding the competitiveness and efficiency of markets; sensitivity to changes in price and advertising; potential demand for new products; and the private and social gains from these products. Modern markets make this estimation a challenging task. In particular, when one wants to analyze the demand for a product, the prices and availability of substitutes should be considered. For most products this means considering hundreds of potential substitutes, which imposes a large number of parameters to be estimated.
Berkeley's recent work in this area uses widely available aggregate data to estimate demand for an unrestricted number of differentiated products, allowing for heterogenous consumer tastes. The results are used to analyze mergers, evaluate the value of new products, and treat biases in the Consumer Price Index (CPI). The goal of ongoing research is threefold. First, the Department aims to study estimates' sensitivities to various assumptions and further develop the methodology. Second, we want to increase computation time. Although the current computation is faster than any previously done, it is still time consuming and can be further improved--primarily by improving the ability to use multi-processor machines and several machines on a shared network. Here, PentiumPro machines using WindowsNT have a great advantage over Unix-based workstations. Finally, these two goals relate to the means to achieve the ability to model and address complex, real-world situations. The ultimate goal of this research is to generate a tool that is both useful modelling real-world situations and useful in answering the questions raised by the models.
The project is important for many reasons. From a scientific point of view, this project extends our ability to study and simulate industry behavior in a realistic and useful way. Additionally, analysis of the effects of pricing decisions, marketing activities, and introduction of new brands are highly valued by companies in many industries. Furthermore, evaluations of mergers and of the Consumer Price Index biases are of great interest to government agencies, and--more importantly--have great implications for the well-being of practically any U.S. consumer. For example, the biases in the CPI, as estimated by a recent report of an advisory committee to Congress, cost the public $135 billion a year!
For both private companies and government agencies, it is important what an analysis method can deliver and what time, cost, and user-friendliness are required to implement the method. This project strives to improve all these dimensions, and welcomes the assistance of the Intel-based systems requested in this proposal to realize these improvements. The computational burden required to achieve these potential benefits is almost unmatched in the social sciences and is comparable to programs in the natural sciences. These computational costs will increase as the model is extended to deal with realistic situations. Improvements to the method through the use of more efficient algorithms and better programming have been employed, and the primary potential for future gains is in improving the computerized system.
Compared to previous Department work using either SunSparc workstations or super-computers, Economics has been able to achieve higher computing power by using a network of Pentium and PentiumPro machines with a WindowsNT operating system. A significant increase in power was achieved by "parallel processing" the problem. The method used is general, requires a simple network, and is ideal for use with WindowsNT machines. Initial results of this project will be available within several months of receiving the required Intel equipment, which includes several multi-processor (quad- and dual- PentiumPro) machines connected by a local network (detailed in the budget attachment).
February 1999