In: Economics
A modern consumption function is estimated using quarterly data for 1959:2 to 2000:4, a total of 166 observations.
Consumption Real Personal Consumption Expenditures (Billions of 1996 dollars)
Income Real Disposable Income (Billions of 1996 dollars)
Sentiment Index of Consumer Sentiments (1966q1=100)
RealR Real 3 Month Treasury Bill rate (percent)
Wealth Household net worth (Billions of dollars)
Consumption |
Coefficient |
Std. Error |
t-Statistic |
|
Income |
0.781254 |
0.009032 |
||
Sentiment |
1.597824 |
0.329168 |
||
RealR |
-10.95494 |
1.493983 |
||
Wealth _cons |
0.020927 52.97639 |
0.001179 42.76011 |
||
R-squared |
0.999204 |
|||
Adjusted R-squared |
0.999184 |
|||
F-statistic |
50518.28 |
|||
Durbin-Watson stat |
0.607651 |
The estimated model is given. The individual statistical significance and economic meaning is given as below. The t-test for individual significance would have the test statistic as , which will follow t-distribution with df=n-k=161. For 5% significance level, the critical t would be as . Note that the null is and alternate is , and if , then we would fail to reject the null, and otherwise if , we may reject the null.
The model suggests that the consumption
The overall fit is suggested by R-squared and (a bit better by) adjusted R-squared. The R-square are high enough, suggesting that a major proportion (more than 99%) of the variation in consumption is explained by variations in the independent variables. The test of overall significance would have the null and alternate that at least one of the coefficient is statistically significant (statistically different from zero). The F-statistic is given as , and would follow the F-distribution with df=k,n-k-1=4,161. The critical F at 5% significance would be as . Since , we may reject the null, and conclude that the model is overall significant (not all or at least one of the variables are useful in explaining the dependent variable).
The econometric problem posed by the model is that, the model suffers from (serial) autocorrelation (among residuals). This is suggested by the low durbin-watson test statistic d, since d<2. The autocorrelation means that regressors are related as (w is normal - IID with zero mean), and as a consequence, the R-square and individual t-statistics would be useless, and usually overestimated (more than the true parameters). The parameters are also inconsistent, ie the biasness is not removed even if we increase the sample size.
A remedy is to estimate the rho, which can be approxed as , for d be the DW statistic. Then, we have or or or or or or . In this case, we have the residual wt, which is normal - IID with mean zero. Hence, transforming the dependent variable by and the independent variables by after estimating the rho, we may run the regression as , and would have the required true coefficients and their individual significance, along with the true overall significance.