Comment on G. E. Mizon, "Modelling Relative Price Variability and Aggregate Inflation in the United Kingdom"

Abstract

In most cases simple correlations between economic variables do not provide enough evidence to confirm an economic theory and justify a particular policy. Nowadays, this statement hardly needs any defense. Stylized facts have to be interpreted in a serious manner. Superficial analysis may lead to absurd conclusions. A nice anecdote is quoted in Fisher's (1966, pp. 2-3) book on identification. In Tsarist Russia there was a positive correlation between the geographical distribution of the occurrence of cholera and the geographical distribution of the presence of doctors, since the government had sent doctors to the affected areas. The story goes that distraught farmers in the affected areas killed the doctors. The danger of superficial theory and superficial causal inference remains, even when modern methods are used. Suppose we begin with the superficially simple assumption that doctors' behavior is such that they rush to areas where there are sick people. If we summarize this behavior in a mathematical formula, make use of simulation techniques to generate artificial data from this model and compare them with the real data, we obtain a naive confirmation that doctors are present where diseases occur and that doctors should be prevented from going to areas where people are sick. Clearly, in order to determine the cause of a disease, we need a more profound theory that includes other relevant variables. Moreover, a simple predictive analysis over a relatively short period would have refuted the naive theory. In order to overcome the problems of superficially simple inference, Mizon advocates the general to specific modeling strategy where extensive use is made of instrumental variable estimation and several different testing procedures. These methods are collected in the computer package PCGIVE (General Instrumental Variable Estimation). My comments refer to these topics. Mizon's suggestion that the general to specific modeling strategy is good, while the specific to general modeling strategy is seriously flawed, requires some interpretation. The advantage of the general to specific