Jamie Monogan
Washington University in St. Louis
1 / 19
Objectives
By the end of this meeting, participants should be able to: Explain why F -ratios are useful in instances of multicollinear predictors. Conduct a direct Granger test between two variables.
2 / 19
Background on Multicollinearity
Example
Imagine that we wish to test prosperity () leads to incumbent re-election success And for we have ve indicators:
1 2 3 4 5
3 / 19
Background on Multicollinearity
Common variance, indicating Unique variance, due to the indicator itself Error variance
2 2 2 = common + unique + error
2 total
This is the theory behind measurement models such as exploratory or conrmatory factor analysis.
4 / 19
Background on Multicollinearity
5 / 19
Background on Multicollinearity
The Eect
If we do a regression of y as a function of x1, x2, x3, x4, & x5 to test the eect of on y, we are likely to see that: One of the xs gets explanatory credit for , while the others may be nonsignicant, wrong signed, or even wrong signed and signicant. The problem, is that after the common variance is modelled in one x, the others are picking up only the screwy relationships between the atheoretical unique variances and y.
6 / 19
Background on Multicollinearity
In this instance it is quite wrong to assert (blindly) that one of the xs inuences y and the others do not. If they tap the same concept, then you cant know which of them represents and which represents meaningless unique variances. The correct approach is not to try to separate eects at all if they are inseparable and to ask the theoretically correct question about y, not the individual xs. How? Block exclusion (F) test.
7 / 19
Background on Multicollinearity
Produce R2 s for an unrestricted model (UR) and a restricted version (R) that drops several parameters. Then a test of the restriction is: Fq,Nk =
2 2 (RUR RR )/q 2 (1 RUR )/(N k)
Where q is the number of dropped parameters, N is the number of observatons, and k is the number of parameters in the unrestricted model. Note: For the special case of 1 parameter, F=t2 for that parameter and p(F)= p(t) and thus we do not need a block exclusion test.
8 / 19
Background on Multicollinearity
Software
R Run the restricted and unrestricted models as separate lm objects (perhaps model.ur and model.r). Then, anova(model.ur, model.r) will oer the F -test on the restrictions. Stata In Stata, simply estimate the unrestricted model and then ask for an F -test on several variables in the model It will still yield a comparison of R2 with and without particular coecient restrictions. To test the exclusion of d, e, and f:
1 2
reg y a b c d e f test d e f
Granger Causality Testing November 10, 2010 9 / 19
Background on Multicollinearity
For the theory xy, nd a z such that you have a causal sequence, xzy Then, if z explains the x-y connection, it follows that controlling z should eliminate all further association. Thus for y = 0 + 1 x + 2 z + , 1 would be zero. Testing if 1 =0 then tests the causal theory.
10 / 19
We are often in the position of having about equally good theories and research programs that posit both xy and yx. So we specify our model according to theory and get contrary results. Researchers on both sides, according to Grangerand later more emphatically Simsare arrogant about theory and data. And since they get contrary results, at least one of them is wrong. So we turn to causality testing in humility, admitting that we do not know the right model and just asking the data to speak to us.
11 / 19
For many of these linkages micro theory could justify causality in either direction. Example: Economic Perceptions Approval
It is plausible that those who think the economy is strong are more likely to approve the president. But also those who approve the president might be likely to distort their economic perceptions in the direction of nding more prosperity than actually is the case.
This is a case (for the macro indicators) where we would rather let the data speak about cause and exogeneity that to settle itmaybe falselyby assumption.
Jamie Monogan (WUStL) Granger Causality Testing November 10, 2010 12 / 19
Granger Causality
Intuition: if xy then perturbing x (x) leads to later changes in y (y) Asymmetry: if xy then perturbing y (y) has no eect on future values of x Denition: A series x may be said to cause a series y if and only if the expectation of y given the history of x is dierent from the unconditional expectation of Y.
E(y| ytk , xtk ) = E (y | ytk )
13 / 19
For k appropriate lags, we model: yt = 0 + 1 yt1 + . . . k ytk + et then we ask whether adding similar information about x will improve our ability to predict y. Thus: yt = 0 + 1 yt1 + . . . + k ytk + 1 xt1 + . . . + k xtk + et The s are uninformative. If the s are jointly signicant, we have established cause.
14 / 19
For k appropriate lags, we model: xt = 0 + 1 xt1 + . . . k xtk + et then we ask whether adding similar information about y will improve our ability to predict x. Thus: xt = 0 + 1 xt1 + . . . + k xtk + 1 yt1 + . . . + k ytk + et The s are uninformative. If the s are jointly zero, then x is exogenous with respect to y.
15 / 19
Software
R Create a tsunion data set with several lags of y & x Run two lm objects with y as the DV: one with lagged xs (y.with.x) and one without (y.without.x) anova(y.with.x, y.without.x), signicant F implies xy Then the reverse: Run two lm objects with x as the DV: one with lagged ys (x.with.y) and one without (x.without.y) anova(x.with.y, x.without.y), nonsignicant F implies y Stata reg y l(1/4).y l(1/4).x test l(1/4).x, signicant F implies xy Then the reverse: reg x l(1/4).x l(1/4).y test l(1/4).y, nonsignicant F implies y
Jamie Monogan (WUStL) Granger Causality Testing
x
November 10, 2010 16 / 19
See Freeman (1983). Both approaches are similar in that they rst control for expected patterns of autodependence in the dependent variable, and then claim causation when the cleaned series are associated with the independent variable. The dierence is that Box-Jenkins modeling uses an empirical technique to identify the best simple model of autodependence whereas Granger modeling is more brute force, including serveral lags of y so that dependence may be modeled out even without knowing exactly what it is.
17 / 19
Theoretical Issues
The Direct Granger test violates the correct specication assumption of OLS, since the correct specication is assumed to be unknown. It relies on overtting to make sure that all the autodependence process is removed from the data. Overtting is harmless as regards unbiasedness, but causes estimator ineciency. Consequently the Granger coecients are not optimal and should not be used. That is, the direct Granger test should be used as a signicance test only, not as a source of structural coecients.
18 / 19
Read Enders 5.5-5.9 and Brandt & Williams chapter 2. Download HWarmsRace.csv (remember: read.csv) Use a direct Granger test to show if there is a causal relationship between the defense expenditures of India and Pakistan. If there is, are the expenditures of one country exogenous to the others?
19 / 19