2
$\begingroup$

I have to do a multiple regression in order to predict the GDP, using some (or all) of the variables that I have (consumption, investment, govt expenditures, disposable income, price index, money stock, population, t bill, unemployment, inflation, and interest). The problem is that 7 from the 11 variables that I have are highly correlated with the GDP and between each other (r>0.9).

How can I pick variables to predict the GDP? Should I pick only the one with the highest GDP correlation and then try adding the other variables with lower correlation?

$\endgroup$
  • $\begingroup$ You could perform a factor analysis/principle components analysis on your independent variables and then regress GDP on your principle components since they are independent. $\endgroup$ – StatsStudent Jan 16 '16 at 23:29
2
$\begingroup$

The choice of variables depends on some other criterion you're trying to optimize. For example, if your goal is to make the best possible prediction of GDP given the predictors (best = unbiased with smallest standard error), then including all the predictors would be the way to go. Predictors can't ever add noise (if they were totally unrelated to the outcome their coefficients would be zero), so in a very basic sense more are better.

But we're often asking a slightly different question, such as "what is the best prediction I can get using only K predictors" or "What is the optimal adjusted R^2 value I can obtain?" (Adjusted R^2 starts with the usual R^2 and then "penalizes" for the number of predictors).

The general process you're engaged in is "model selection." There are various algorithms for this (stepwise regression, best subsets regression, etc.) Stepwise has fallen out of favor (see http://en.wikipedia.org/wiki/Stepwise_regression) , and for a relatively small number of predictors a best-subsets algorithm would give you the an optimal tradeoff between model parsimony and predictive power. But if you want to try stepwise then, yes, forward stepwise does what you propose: pick the predictor with the highest R^2, generate the residual from that Y=x1 + e1 model, then pick x2 from the remaining predictors that best explains e1 (e1 = X2 + e2), rinse and repeat.

But again, you might be going for a completely different criterion, such as a model that best explains the processes that actually cause (lead to) GDP. There is a distinction between "explanation" and "prediction." The former would rely on some substantive understanding about what the variables actually mean and how they function in an economy - the latter is just about variables (I as a non-economist can answer your questions about prediction, but not about what variables best EXPLAIN GDP)

Bottom line - the answer to your question depends on a careful consideration of the criteria you're trying to optimize.

$\endgroup$
0
$\begingroup$

Multiple Regression modelling via OLS is useful for cross-sectional data but hardly ever for time series data ( your data ! ) due to potentially correlated observations over time. This fine point ( or not so fine ! ) is hardly ever openly discussed in basic courses as students would just walk out if they knew the limitations of multiple regression when applied to time series data. You might want to look at http://www.autobox.com/cms/index.php/afs-university/intro-to-forecasting/doc_download/24-regression-vs-box-jenkins otherwise entitled "Lies my mother never told me " . In addition to advanced/correct identification of the Transfer Function also known as an ADL OR PDL model one needs to be concerned with level shifts/time trends/seasonal and regular pulses while confirming/ensuring a constant error process and constant parameters over time. Avoid any advice that doesn't mention some of these issues as they are very important.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.