Strange Looking Regression Results

[davehill17]

Hi, new member here so apologies if this thread is in the wrong place.

I have created a regression for consumption and GDP, with 64 observations. Having transformed both sets of data to 2005 prices, I have attempted to regress the log of each, with consumption as the dependent variable. These are the results:

Dependent Variable: LOG(HCE)
Method: Least Squares
Date: 03/04/13 Time: 12:15
Sample: 1948 2011
Included observations: 64

Variable Coefficient Std. Error t-Statistic Prob.

C -0.965319 0.052939 -18.23453 0.0000
LOG(GDP) 1.061651 0.006124 173.3454 0.0000

R-squared 0.997941 Mean dependent var 8.189431
Adjusted R-squared 0.997908 S.D. dependent var 0.640675
S.E. of regression 0.029305 Akaike info criterion -4.191338
Sum squared resid 0.053246 Schwarz criterion -4.123872
Log likelihood 136.1228 Hannan-Quinn criter. -4.164760
F-statistic 30048.63 Durbin-Watson stat 0.227328
Prob(F-statistic) 0.000000

Various things about this seem wrong, from the fact that the 'Prob.' column contains all zeroes, as does the Prob(F-statistic) cell, to the ridiculously high t-Statistic for the LOG(GDP) variable. I used the command 'ls log(hce) c log(gdp)'. Does anyone know where I have gone wrong?

If more information is needed, just ask! Many thanks in advance.

[EViews Gareth]

I see nothing done wrong

[trubador]

There is nothing practically wrong here. Both variables are probably nonstationary and Durbin-Watson statistic suggests that residuals might be serially correlated. Consumption is a component of GDP, which usually constitutes the lion share (e.g. 2/3) of expenditures. Therefore, it would not be a coincidence if two series shared a common stochastic drift (or in other words, they are cointegrated). Cointegration Regression or Vector Error Correction methods might provide you with more appropriate estimation frameworks...

[davehill17]

Ok that's brilliant, just my ignorance coming to the fore! Thanks for the help.

What does it mean if the 'Prob.' column contains all zeroes?

[trubador]

It is the associated probability of t-statistic, which tests the significance of the related coefficient against zero. In other words, the probability of not rejecting the hypothesis saying that the value of coefficient is no different than zero.

The format of probability cells are fixed to 4-digit precision in decimals, so it only means that the number is smaller than 0.00005. If you like, you can copy the number at highest precision from the table or just compute it on your own. For instance, you can obtain the same probability value for the constant as follows:

Code: Select all

scalar pval = @tdist(-18.23453,62)