L28 - Modelling Strategies Flashcards
What is the textbook approach to statistical modelling?
- we perform experiments to generate our data
- every time we repeat it the data is effect by a random error
What is a more realistic econometrics approach to modelling?
What is specific to general modelling?
The traditional approach to econometrics modelling was as follows:
- Start with an equation based on economic theory.
- Estimate the equation using an appropriate technique (OLS, IV or something else).
- Check the equation for statistical problems using diagnostic tests for serial correlation, heteroscedasticity etc.
- If the equation fails the tests then respecify it.
- Continue until the equation performs satisfactorily
Problems with the specific to general approach?
- T and F-tests will be unreliable, they wouldn’t be the correct distribution
What is General to Specific Modelling?
General to specific modelling is an alternative modelling strategy.
In GS modelling we:
- Begin with a very general model including as many lags as possible (focused a lot on time-series data)
- Simplify the model by eliminating insignificant variables to obtain a parsimonious specification.
- Rewrite the final specification in error correction form to make it easier to interpret.
What are the advantages of General to Specific modelling?
- Because the initial model is very general, it is less likely that it will be misspecified.
- The specification search procedure allows the data to determine the shape of the distributed lag relationship.
- Because we test for misspecification at each stage, the final model provides a balance between obtaining a well-specified model and a parsimonious specification.
Why do we try to find the most parsimonious specification?
Parsimonious means the simplest model/theory with the least assumptions and variables but with greatest explanatory power
What is a problem with General to Specific Modelling?
- it dangerous to impose lots of restrictions in one go, its best to do it gradually until we get a satisfactory model e.g. if you had many insignificant variables, it might make sense just to get rid of the least insignificant parameter first and then re-test
how can you test for the significance of an AR(1) variable?
We can’t compare this with the unrestricted model using an F-test because the AR(1) model involves a nonlinear restriction.
What is an error-correction form of a model?
- by changing the way the model is written makes it more foreboding but can be interpreted in a more sensible way
- The expression in brackets constitute the equilibrium relationship between Y and X
- The equation has also been separated into the short-run relationship of Y and X (delta) and the level of Y and X which will show the long-run relationship
What are the important features of the error-correction model?
The important features of the error-correction model are:
- A relationship between the growth rate (or differences) of the LHS variable and those of the RHS variables which describes short-run adjustment.
- A relationship in levels between the variables which describes the long-run equilibrium relationship.
- A parameter which describes the speed of adjustment when the variables are away from long-run equilibrium
There is ALWAYS an algebraic transformation which allows us to write an ARDL model in EC form but some of the variables may not be significant when it is estimated.
How do you calculate the equilibrium response of a regression?
in logged form (long-run impact of X on Y):
ηY,X= sum of all X coefficient/1-the lagged Y coefficients
Why use the error correction form of the model?
Why use the error correction form of the model?
- The parameters have more natural economic interpretations than those of the ARDL model.
- The ECM provides a test for the existence of a long-run relationship between the variables (whether the coefficient on the lagged endogenous variable is significantly negative).
but….
the distribution of the estimator for the coefficient on the lagged endogenous variable is non-standard – the critical values are higher than those for the t distribution.
