SCM

Forum: help

Monitor Forum | Start New Thread Start New Thread
RE: Strange resultf from systemfit [ reply ]
By: Arne Henningsen on 2015-06-27 06:11
[forum:42334]
Hi Carol

The formulas for argument "inst" must include *all* instrumental variables, i.e. including the exogenous regressors:

inst1 <- ~ probc11 + bbpop + paperpen + buying + rlstate + auto + pi

inst2 <- ~ probc21 + bbpop + paperpen + tv + radio + rlstate + auto + pi

If variables "buying", "tv", "radio", "probc11", and "probc21" are indeed exogenous, you could use all instrumental variables for both of the equations:

inst <- ~ probc11 + probc21 + bbpop + paperpen + buying + tv + radio + rlstate + auto + pi

Best regards,
Arne

Strange resultf from systemfit [ reply ]
By: Carol Ting on 2015-06-26 14:49
[forum:42333]
Hi everyone,

I am fitting a 2-equation system with the systemfit package:

log(ads) = log(clicks) + bbpop + paperpen + buying + rlstate + auto + pi
log(clicks) = log(ads) + bbpop + paperpen + tv +radio + rlstate + auto + pi

Where the variables log(clicks) and log(ads) are endogenous, and they are instrumented by variables probc11 and probc21, respectively.

The dataset is in my public folder https://dl.dropboxusercontent.com/u/14429852/search.csv and I used the following code to perform the analysis but got some strange results below (a large number of standard errors are NA's):

eq1 <- lads ~ lclicks + bbpop + radio + tv + paperpen + rlstate + auto + pi
eq2 <- lclicks ~ lads + bbpop + buying + paperpen + rlstate + auto + pi
system <- list( ads = eq1, clicks = eq2 )
inst1 <- ~ probc11
inst2 <- ~ probc21
instlist <- list( inst1, inst2 )
fit3sls <- systemfit( system, "3SLS", inst = instlist, data = Search )
summary(fit3sls)

RESULTS-------------------------


systemfit results
method: 3SLS

N DF SSR detRCov OLS-R2 McElroy-R2
system 606 589 100710 3219.98 -98.5842 -38.5621

N DF SSR MSE RMSE R2 Adj R2
ads 303 294 10026.1 34.1025 5.83974 -119.0536 -122.3204
clicks 303 295 90683.4 307.4013 17.53286 -96.7416 -99.0609

The covariance matrix of the residuals used for estimation
ads clicks
ads 7.52202 -9.55094
clicks -9.55094 55.42238

The covariance matrix of the residuals
ads clicks
ads 34.1025 85.2243
clicks 85.2243 307.4013

The correlations of the residuals
ads clicks
ads 1.000000 0.832372
clicks 0.832372 1.000000


3SLS estimates for 'ads' (equation 1)
Model Formula: lads ~ lclicks + bbpop + radio + tv + paperpen + rlstate + auto +
pi
Instruments: ~probc11

Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.11322e-02 NA NA NA
lclicks -3.00638e-01 6.52312e+05 0e+00 1.00000
bbpop 1.91237e-03 1.72256e+02 1e-05 0.99999
radio 8.44606e-05 1.72739e+04 0e+00 1.00000
tv 2.01290e-01 NA NA NA
paperpen 1.28808e-02 NA NA NA
rlstate -1.24766e+00 NA NA NA
auto -9.46462e+00 NA NA NA
pi 5.42942e+00 NA NA NA

Residual standard error: 5.839735 on 294 degrees of freedom
Number of observations: 303 Degrees of Freedom: 294
SSR: 10026.135118 MSE: 34.1025 Root MSE: 5.839735
Multiple R-Squared: -119.053649 Adjusted R-Squared: -122.320415


3SLS estimates for 'clicks' (equation 2)
Model Formula: lclicks ~ lads + bbpop + buying + paperpen + rlstate + auto +
pi
Instruments: ~probc21

Estimate Std. Error t value Pr(>|t|)
(Intercept) 4.89695e+00 6.75394e+06 0 1
lads -4.66608e-01 NA NA NA
bbpop -6.42986e-04 NA NA NA
buying 6.98992e-02 NA NA NA
paperpen -2.61266e-02 NA NA NA
rlstate 4.82680e+00 NA NA NA
auto -3.11123e+01 NA NA NA
pi 1.70815e+01 NA NA NA

Residual standard error: 17.532862 on 295 degrees of freedom
Number of observations: 303 Degrees of Freedom: 295
SSR: 90683.370466 MSE: 307.401256 Root MSE: 17.532862
Multiple R-Squared: -96.741634 Adjusted R-Squared: -99.060927


Will really appreciate it if anyone can point out what is going on. Thanks a lot!!!

Carol

Thanks to:
Vienna University of Economics and Business University of Wisconsin - Madison Powered By FusionForge