THE ECONOMETRICS OF MACROECONOMIC MODELLING
Rival models
Financial deregulation in the mid-1980s led to a strong rise in aggregate consumption relative to income in several European countries. The pre-existing empirical macroeconometric consumption functions in Norway, which typically explained aggregate consumption by income, all broke down—that is, they failed in forecasting, and failed to explain the data ex post.
As stated in Eitrheim et al. (20026), one view of this forecast failure is that it provided direct evidence in favour of the rival rational expectations, permanent income hypothesis: in response to financial deregulation, consumers revised their expected permanent income upward, thus creating a break in the conditional relationship between consumption and income. The breakdown has also been interpreted as a confirmation of the relevance of the Lucas critique, in that it was a shock to a non-modelled expectation process that caused the structural break in the existing consumption functions.
Eitrheim et al. (20026) compare the merits of the two competing models: the empirical consumption function (CF), conditioning on income in the long run, and an Euler equation derived from a model for expectation formation. We find that while the conditional consumption function encompasses the Euler equation (EE) on a sample from 1968(2) to 1984(4), both models fail to forecast the annual consumption growth in the next years. In the paper, we derive the theoretical properties of forecasts based on the two models. Assuming that the EE is the true model and that the consumption function is a mis-specified model, we show that both sets of forecasts are immune to a break (i. e. shift in the equilibrium savings rate) that occurs after the forecast have been made. Moreover, failure in ‘before break’ CF-forecasts is only (logically) possible if the consumption function is the true model within the sample. Hence, the observed forecast failure of the CF is corroborating evidence in favour of the conditional consumption function for the period before the break occurred.
However, a respecified consumption function—B&N of the previous section—that introduced wealth as a new variable was successful in accounting for the breakdown ex post, while retaining parameter constancy in the years of financial consolidation that followed after the initial plunge in the savings rate. The respecified model was able to adequately account for the observed high variability in the savings rate, whereas the earlier models failed to do so.
B&N noted the implication that the respecification explained why the Lucas critique lacked power in this case: first, while the observed breakdown of conditional consumption functions in 1984-85 is consistent with the Lucas critique, that interpretation is refuted by the finding of a conditional model with constant parameters. Second, the invariance result shows that an Euler equation type model (derived from, for example, the stochastic permanent income model) cannot be an encompassing model. Even if the Euler approach is supported by empirically constant parameters, such a finding cannot explain why a conditional model is also stable. Third, finding that invariance holds, at least as an empirical approximation, yields an important basis for the use of the dynamic consumption function in forecasting and policy analysis, the main practical usages of empirical consumption functions.
Eitrheim et al. (20026) extend the data set by nine years of quarterly observations, that is, the sample is from 1968(3) to 1998(4). The national accounts were heavily revised for that period. We also extended the wealth measure to include non-liquid financial assets. Still we find that the main results of B&N are confirmed. Empirical support for one cointegrating vector between cht, yht, and wht, and valid conditioning in the consumption function is reconfirmed on the new data. In fact, full information maximum likelihood estimation of a four equation system explaining (the change in) cht, yht, wht, and (pht — pt) yields the same empirical results as estimation based on the conditional model. These findings thus corroborate the validity of the conditional model of B&N.
