Answer Key 3 Time Series Data and Serial Correlation

1. (Adapted from C11.5, C12.1) Suppose we are interested in how personal tax exemptions (pe) affect
the general fertility rate (gfr). Use the data in FERTIL3.RAW for this exercise.

i. Estimate the following equation:

gfrt = 0 + 1 pet + 2ww2t + 3pillt + t

where ww2 is a dummy variable for the years 1941-1945, and pill is a dummy variable
that is equal to one for the years 1963 on (after the pill was available). Discuss the
significance of the coefficients and interpret their magnitudes.
reg gfr pe ww2 pill

Source | SS df MS Number of obs = 72

-------------+------------------------------ F( 3, 68) = 20.38
Model | 13183.6215 3 4394.54049 Prob > F = 0.0000
Residual | 14664.2739 68 215.651087 R-squared = 0.4734
-------------+------------------------------ Adj R-squared = 0.4502
Total | 27847.8954 71 392.223879 Root MSE = 14.685

------------------------------------------------------------------------------
gfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
pe | .08254 .0296462 2.78 0.007 .0233819 .1416981
ww2 | -24.2384 7.458253 -3.25 0.002 -39.12111 -9.355684
pill | -31.59403 4.081068 -7.74 0.000 -39.73768 -23.45039
_cons | 98.68176 3.208129 30.76 0.000 92.28003 105.0835

A $1 increase in personal exemptions increases fertility by .08 children per 1000 women, which
is statistically significant at the 1% level. Is it economically significant? A $1 increase is fairly
trivial, so another way to interpret the effect is to look at a 1 standard deviation increase in pe
($65.88 according to the summary stats). This translates into reduction of 5 children per 1000
women, about a fourth of the standard deviation for gfr. The WWII and post-pill periods both
had lower average fertility rates than the earlier periods: the average fertility rate in other periods
(when pe=0) is 98.7, but is 74.5 during WWII and 67.1 after the introduction of the pill.

ii. Fertility may react to personal exemptions with a lag. Reestimate your equation adding 2
lags of pe (pet-1 and pet-2). Are these variables jointly significant? What are the degrees of
freedom of your F-test and why?
tsset year

reg gfr pe ww2 pill l1.pe l2.pe

Source | SS df MS Number of obs = 70

-------------+------------------------------ F( 5, 64) = 12.73
Model | 12959.7886 5 2591.95772 Prob > F = 0.0000
Residual | 13032.6443 64 203.635067 R-squared = 0.4986
-------------+------------------------------ Adj R-squared = 0.4594
Total | 25992.4329 69 376.701926 Root MSE = 14.27

------------------------------------------------------------------------------
gfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
pe | .0726718 .1255331 0.58 0.565 -.1781094 .323453
ww2 | -22.1265 10.73197 -2.06 0.043 -43.56608 -.6869196
pill | -31.30499 3.981559 -7.86 0.000 -39.25907 -23.35091
pe |
L1. | -.0057796 .1556629 -0.04 0.970 -.316752 .3051929
L2. | .0338268 .1262574 0.27 0.790 -.2184013 .286055
 _cons | 95.8705 3.281957 29.21 0.000 89.31403 102.427

Note that this is the same as results using the transformed variables included in the dataset:

reg gfr pe ww2 pill pe_1 pe_2

Source | SS df MS Number of obs = 70

-------------+------------------------------ F( 5, 64) = 12.73
Model | 12959.7886 5 2591.95772 Prob > F = 0.0000
Residual | 13032.6443 64 203.635067 R-squared = 0.4986
-------------+------------------------------ Adj R-squared = 0.4594
Total | 25992.4329 69 376.701926 Root MSE = 14.27

------------------------------------------------------------------------------
gfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
pe | .0726718 .1255331 0.58 0.565 -.1781094 .323453
ww2 | -22.1265 10.73197 -2.06 0.043 -43.56608 -.6869196
pill | -31.30499 3.981559 -7.86 0.000 -39.25907 -23.35091
pe_1 | -.0057796 .1556629 -0.04 0.970 -.316752 .3051929
pe_2 | .0338268 .1262574 0.27 0.790 -.2184013 .286055
_cons | 95.8705 3.281957 29.21 0.000 89.31403 102.427

. test pe_1 pe_2

( 1) pe_1 = 0
( 2) pe_2 = 0

F( 2, 64) = 0.05
Prob > F = 0.9480

No, these variables are not jointly significant.

iii.
ρ^
What are the first order autocorrelations ( 1 ) for gfr and pe? What do these suggest
about possible unit root(s)? What does this suggest about your OLS results in (i)?

You can regress gfr on lagged gfr, pe on lagged pe. Here is an alternative approach:

corrgram pe

-1 0 1 -1 0 1
LAG AC PAC Q Prob>Q [Autocorrelation] [Partial Autocor]
-------------------------------------------------------------------------------
1 0.9471 0.9479 67.314 0.0000 |------- |-------
2 0.8776 -0.2493 125.94 0.0000 |------- -|
3 0.8103 0.0594 176.64 0.0000 |------ |
4 0.7414 -0.0868 219.7 0.0000 |----- |

. corrgram gfr

-1 0 1 -1 0 1
LAG AC PAC Q Prob>Q [Autocorrelation] [Partial Autocor]
-------------------------------------------------------------------------------
1 0.9447 0.9777 66.977 0.0000 |------- |-------
2 0.8730 -0.3048 124.99 0.0000 |------ --|
3 0.8072 0.1752 175.3 0.0000 |------ |-
4 0.7345 -0.2981 217.56 0.0000 |----- --|
The first order autocorrelation is very close to 1, suggesting that there may be a possible unit root. Unless
gfr and pe are cointegrated, this implies that our results may be driven by spurious correlation.

iv. Re-estimate (i) using first differences—that is, changes in gft and changes in pe. (Do not
difference ww2 and pill.) How does the effect of pe compare with your estimates in levels in
(i)?
reg cgfr cpe ww2 pill

Source | SS df MS Number of obs = 71

-------------+------------------------------ F( 3, 67) = 4.10
Model | 197.055018 3 65.6850059 Prob > F = 0.0098
Residual | 1072.52736 67 16.007871 R-squared = 0.1552
-------------+------------------------------ Adj R-squared = 0.1174
Total | 1269.58238 70 18.1368911 Root MSE = 4.001

------------------------------------------------------------------------------
cgfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cpe | -.0935102 .0325581 -2.87 0.005 -.1584965 -.0285239
ww2 | 5.131562 2.249415 2.28 0.026 .6417119 9.621413
pill | -1.987072 1.04927 -1.89 0.063 -4.081424 .1072791
_cons | -.4703762 .6034214 -0.78 0.438 -1.67481 .7340579

Note this is the same as

reg D.gfr D.pe ww2 pill

Source | SS df MS Number of obs = 71

-------------+------------------------------ F( 3, 67) = 4.10
Model | 197.055018 3 65.6850059 Prob > F = 0.0098
Residual | 1072.52736 67 16.007871 R-squared = 0.1552
-------------+------------------------------ Adj R-squared = 0.1174
Total | 1269.58238 70 18.1368911 Root MSE = 4.001

------------------------------------------------------------------------------
D.gfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
pe |
D1. | -.0935102 .0325581 -2.87 0.005 -.1584965 -.0285239
ww2 | 5.131562 2.249415 2.28 0.026 .6417119 9.621413
pill | -1.987072 1.04927 -1.89 0.063 -4.081424 .1072791
_cons | -.4703762 .6034214 -0.78 0.438 -1.67481 .7340579

Now we see that a 1 dollar change in personal exemptions leads to a .09 reduction in the growth rate of
fertility, very different than the results from before. This suggests that some of the positive result from
before may have been driven by spurious correlation--an overall reduction in both the birth rate and
exemptions over time. Here are two graphs to give you a sense of what is going on:
 250
200
150
100
50
0

1920 1940 1960 1980

1913 to 1984
births per 1000 women 15-44 real value pers. exemption, $
100
50
0
-50

1920 1940 1960 1980

1913 to 1984
change in gfr: gfr - gfr_1 pe - pe_1

In the first graph, the two series move more or less together. In the second graph, we can see that the
annual changes often have an inverse relationship. (You can also see that differencing these series leads
to stationarity).

v. Reestimate (ii) using first differences of gft, pe, and lagged pe. (Again, do not difference
ww2 and pill.) Interpret the coefficients and comment on their statistical significance.
reg cgfr cpe cpe_1 cpe_2 ww2 pill

Source | SS df MS Number of obs = 69

-------------+------------------------------ F( 5, 63) = 5.29
Model | 372.894284 5 74.5788568 Prob > F = 0.0004
Residual | 888.565534 63 14.1042148 R-squared = 0.2956
-------------+------------------------------ Adj R-squared = 0.2397
Total | 1261.45982 68 18.5508797 Root MSE = 3.7556
------------------------------------------------------------------------------
cgfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cpe | -.0751636 .0323566 -2.32 0.023 -.1398232 -.010504
cpe_1 | -.0513865 .0331632 -1.55 0.126 -.1176579 .0148848
cpe_2 | .0882556 .0279766 3.15 0.002 .0323488 .1441624
ww2 | 4.839225 2.831973 1.71 0.092 -.8200213 10.49847
pill | -1.676145 1.004766 -1.67 0.100 -3.684009 .3317186
_cons | -.6502546 .5817652 -1.12 0.268 -1.81282 .5123105

Now we see that while change in personal exemptions are negatively related to changes in fertility in a
given year, after two years, a change in fertility is associated with increased fertility, indicating that the
effect may take some time to appear. This lagged model makes sense, given that outcomes fertility
decisions take time to be realized.

vi. Add a linear time trend to the model in (v). Is a time trend necessary in the first-difference
equation?

It turns out that whether a time trend is significant depends on the specification. The first specification
includes dummies for the post ww2 and pill periods. This essentially shifts the time trend in changes up
after ww2 and down again after the pill:

. reg cgfr cpe cpe_1 cpe_2 ww2 pill t

Source | SS df MS Number of obs = 69

-------------+------------------------------ F( 6, 62) = 5.79
Model | 453.089851 6 75.5149751 Prob > F = 0.0001
Residual | 808.369968 62 13.0382253 R-squared = 0.3592
-------------+------------------------------ Adj R-squared = 0.2972
Total | 1261.45982 68 18.5508797 Root MSE = 3.6108

------------------------------------------------------------------------------
cgfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cpe | -.0618201 .0315717 -1.96 0.055 -.124931 .0012908
cpe_1 | -.039124 .0322664 -1.21 0.230 -.1036236 .0253757
cpe_2 | .0951653 .0270425 3.52 0.001 .0411081 .1492225
ww2 | 3.250812 2.797162 1.16 0.250 -2.340636 8.84226
pill | -4.888069 1.615706 -3.03 0.004 -8.117819 -1.658319
t | .0944006 .0380635 2.48 0.016 .0183127 .1704884
_cons | -3.141163 1.149618 -2.73 0.008 -5.439216 -.8431097

If we do not include those time period dummies, then t has no effect. If you look back and the graph of
changes above, you will see that there does not appear to be a time trend overall, but that the WWII and
post 63-pill periods do have some large outliers. In general, time trends are often not necessary to include
after differencing the data, but that is not always the case: there may be a reason why CHANGES in the
variable trend.

. reg cgfr cpe cpe_1 cpe_2 t

Source | SS df MS Number of obs = 69

-------------+------------------------------ F( 4, 64) = 4.88
Model | 294.856737 4 73.7141842 Prob > F = 0.0017
Residual | 966.603082 64 15.1031732 R-squared = 0.2337
-------------+------------------------------ Adj R-squared = 0.1859
Total | 1261.45982 68 18.5508797 Root MSE = 3.8863

------------------------------------------------------------------------------
cgfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
 cpe | -.0348352 .0272856 -1.28 0.206 -.0893445 .019674
cpe_1 | -.0131442 .0278616 -0.47 0.639 -.0688042 .0425158
cpe_2 | .11109 .0272773 4.07 0.000 .0565974 .1655826
t | .0078781 .0242282 0.33 0.746 -.0405233 .0562796
_cons | -1.267445 1.046219 -1.21 0.230 -3.357507 .822617

vii. Using the model in (vi), test for whether there is AR(1) serial correlation in the errors.
. reg cgfr cpe cpe_1 cpe_2 ww2 pill

Source | SS df MS Number of obs = 69

-------------+------------------------------ F( 5, 63) = 5.29
Model | 372.894284 5 74.5788568 Prob > F = 0.0004
Residual | 888.565534 63 14.1042148 R-squared = 0.2956
-------------+------------------------------ Adj R-squared = 0.2397
Total | 1261.45982 68 18.5508797 Root MSE = 3.7556

------------------------------------------------------------------------------
cgfr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cpe | -.0751636 .0323566 -2.32 0.023 -.1398232 -.010504
cpe_1 | -.0513865 .0331632 -1.55 0.126 -.1176579 .0148848
cpe_2 | .0882556 .0279766 3.15 0.002 .0323488 .1441624
ww2 | 4.839225 2.831973 1.71 0.092 -.8200213 10.49847
pill | -1.676145 1.004766 -1.67 0.100 -3.684009 .3317186
_cons | -.6502546 .5817652 -1.12 0.268 -1.81282 .5123105

. estat dwatson

Durbin-Watson d-statistic( 6, 69) = 1.561962

. estat durbinalt

Durbin's alternative test for autocorrelation

---------------------------------------------------------------------------
lags(p) | chi2 df Prob > chi2
-------------+-------------------------------------------------------------
1 | 3.143 1 0.0763
---------------------------------------------------------------------------
H0: no serial correlation

This is somewhat borderline. The p-value of .0763 indicates that there is a 7 percent chance that
we would observe the autocorrelation that we do even if the truth were no serial correlation.
Although this is not 5% or smaller, given the way the test is structured, we err on the side of
accepting the null. We might want to do more to look at the model (including t, for example.)

2. Suppose we are interested in how laws and economic conditions might affect driving behavior. Use
TRAFFIC2.RAW (monthly observations from CA from Jan 1981-Dec 1989) to answer these questions.

a. The variable prcfat is the percentage of accidents resulting in at least on fatality. Note that
this variable is a percentage, not a proportion. What is the average of this variable over this
period?

sum prcfat

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
prcfat | 108 .8856363 .0997777 .7016841 1.216828

Note that the max is greater than one because there are a few periods with many fatalities per
accident (i.e., more than one person died on average in accidents)
 b. Run a regression of prcfat on a linear time trend, 11 monthly dummies (set January as your
base month), wkends, unem, spdlaw, and beltlaw. Discuss the estimated effects of unem,
spdlaw, and beltlaw. Do the signs and magnitudes make sense to you?

reg prcfat t wkends unem spdlaw beltlaw feb-dec

Source | SS df MS Number of obs = 108

-------------+------------------------------ F( 16, 91) = 14.44
Model | .764228387 16 .047764274 Prob > F = 0.0000
Residual | .301019769 91 .00330791 R-squared = 0.7174
-------------+------------------------------ Adj R-squared = 0.6677
Total | 1.06524816 107 .00995559 Root MSE = .05751

------------------------------------------------------------------------------
prcfat | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
t | -.0022352 .0004208 -5.31 0.000 -.0030711 -.0013993
wkends | .0006259 .0061624 0.10 0.919 -.011615 .0128668
unem | -.0154259 .0055444 -2.78 0.007 -.0264392 -.0044127
spdlaw | .0670877 .0205683 3.26 0.002 .0262312 .1079441
beltlaw | -.0295053 .0232307 -1.27 0.207 -.0756503 .0166397
feb | .0008607 .0289967 0.03 0.976 -.0567377 .0584592
mar | .0000923 .0274069 0.00 0.997 -.0543481 .0545327
apr | .0582201 .0278195 2.09 0.039 .0029601 .11348
may | .0716392 .0276432 2.59 0.011 .0167293 .1265492
jun | .1012618 .0280937 3.60 0.001 .0454571 .1570665
jul | .1766121 .0272592 6.48 0.000 .122465 .2307592
aug | .1926117 .0274448 7.02 0.000 .1380959 .2471274
sep | .1600164 .028203 5.67 0.000 .1039947 .2160381
oct | .1010357 .0276702 3.65 0.000 .0460722 .1559991
nov | .013949 .0281436 0.50 0.621 -.0419548 .0698528
dec | .0092005 .027858 0.33 0.742 -.046136 .064537
_cons | 1.029799 .1029523 10.00 0.000 .8252964 1.234301

Higher speed limits are estimated to increase the percent of fatal accidents, by .067 percentage points.
This is a statistically significant effect. The new seat belt law is estimated to decrease the percent of fatal
accidents by about .03, but the two-sided p-value is about .21.

Interestingly, increased economic activity also increases the percent of fatal accidents. This may be
because more commercial trucks are on the roads, and these probably increase the chance that an accident
results in a fatality.

c. Test the errors for AR(1) serial correlation.

Here is how you could conduct the test using the Breusch-Godfrey method: After getting the OLS
uˆ uˆt on uˆt 1 , t  2,...,108. (Included an intercept, but that is
residuals, t , run the regression
unimportant.)

tsset t
predict error, resid
reg error l.error

Source | SS df MS Number of obs = 107

-------------+------------------------------ F( 1, 105) = 8.91
Model | .023534599 1 .023534599 Prob > F = 0.0035
Residual | .277302016 105 .002640972 R-squared = 0.0782
-------------+------------------------------ Adj R-squared = 0.0695
 Total | .300836614 106 .002838081 Root MSE = .05139

------------------------------------------------------------------------------
error | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
error |
L1. | .2816415 .0943463 2.99 0.004 .0945701 .4687129
_cons | .0002981 .0049684 0.06 0.952 -.0095533 .0101496

uˆ
The coefficient on t 1 is ̂  .281 (se = .094). Thus, there is evidence of some positive serial
correlation in the errors (t  2.99). Note that this test is only valid if there are no concerns about the
endogeneity of the regressors. A strong case can be made that all explanatory variables are strictly
exogenous. Certainly there is no concern about the time trend, the seasonal dummy variables, or wkends,
as these are determined by the calendar. It is seems safe to assume that unexplained changes in prcfat
today do not cause future changes in the state-wide unemployment rate. Also, over this period, the policy
changes were permanent once they occurred, so strict exogeneity seems reasonable for spdlaw and
beltlaw. (Given legislative lags, it seems unlikely that the dates the policies went into effect had anything
to do with recent, unexplained changes in prcfat.

d. Re-estimate the model accounting to serial correlation.

Couple of options:

. arima prcfat t wkends unem spdlaw beltlaw feb-dec, ar(1)

(setting optimization to BHHH)

Iteration 0: log likelihood = 168.775
Iteration 1: log likelihood = 168.81849
Iteration 2: log likelihood = 168.85351
Iteration 3: log likelihood = 168.8576
Iteration 4: log likelihood = 168.85947
(switching optimization to BFGS)
Iteration 5: log likelihood = 168.86032
Iteration 6: log likelihood = 168.8641
Iteration 7: log likelihood = 168.86419
Iteration 8: log likelihood = 168.86439
Iteration 9: log likelihood = 168.8644
Iteration 10: log likelihood = 168.8644
Iteration 11: log likelihood = 168.8644

ARIMA regression

Sample: 1 - 108 Number of obs = 108

Wald chi2(17) = 191.09
Log likelihood = 168.8644 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
| OPG
prcfat | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
prcfat |
t | -.0021501 .0005349 -4.02 0.000 -.0031986 -.0011017
wkends | .0006179 .0054809 0.11 0.910 -.0101244 .0113602
unem | -.0132124 .0068961 -1.92 0.055 -.0267285 .0003037
spdlaw | .0641854 .0298944 2.15 0.032 .0055935 .1227773
beltlaw | -.0248763 .0338792 -0.73 0.463 -.0912783 .0415257
feb | -.0008934 .0361499 -0.02 0.980 -.071746 .0699591
mar | -.0011545 .0306679 -0.04 0.970 -.0612625 .0589534
apr | .0575487 .0315385 1.82 0.068 -.0042656 .1193629
may | .0718131 .0263242 2.73 0.006 .0202186 .1234076
jun | .1007231 .0252368 3.99 0.000 .0512599 .1501863
jul | .174789 .0225581 7.75 0.000 .1305759 .2190022
 aug | .1919537 .0308971 6.21 0.000 .1313964 .252511
sep | .159901 .0287129 5.57 0.000 .1036247 .2161772
oct | .1007692 .0243195 4.14 0.000 .0531039 .1484345
nov | .0133081 .0275731 0.48 0.629 -.0407342 .0673504
dec | .0085411 .0281464 0.30 0.762 -.0466249 .0637071
_cons | 1.009016 .1103693 9.14 0.000 .7926966 1.225336
-------------+----------------------------------------------------------------
ARMA |
ar |
L1. | .2859698 .1117903 2.56 0.011 .0668649 .5050747
-------------+----------------------------------------------------------------
/sigma | .0506463 .0041883 12.09 0.000 .0424375 .0588551

Or this one:
. prais prcfat t wkends unem spdlaw beltlaw feb-dec

Iteration 0: rho = 0.0000

Iteration 1: rho = 0.2816
Iteration 2: rho = 0.2884
Iteration 3: rho = 0.2887
Iteration 4: rho = 0.2887
Iteration 5: rho = 0.2887

Prais-Winsten AR(1) regression -- iterated estimates

Source | SS df MS Number of obs = 108

-------------+------------------------------ F( 16, 91) = 10.16
Model | .495019683 16 .03093873 Prob > F = 0.0000
Residual | .277020817 91 .003044185 R-squared = 0.6412
-------------+------------------------------ Adj R-squared = 0.5781
Total | .7720405 107 .007215332 Root MSE = .05517

------------------------------------------------------------------------------
prcfat | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
t | -.0021487 .0005479 -3.92 0.000 -.003237 -.0010604
wkends | .0006166 .0050041 0.12 0.902 -.0093234 .0105567
unem | -.0131807 .0071065 -1.85 0.067 -.0272969 .0009354
spdlaw | .0641361 .0267953 2.39 0.019 .0109105 .1173616
beltlaw | -.024816 .0301099 -0.82 0.412 -.0846256 .0349936
feb | -.0009093 .0244211 -0.04 0.970 -.0494188 .0476001
mar | -.0011624 .0262733 -0.04 0.965 -.0533512 .0510264
apr | .0575507 .0275861 2.09 0.040 .0027543 .112347
may | .071829 .0278625 2.58 0.012 .0164836 .1271745
jun | .1007237 .0280584 3.59 0.001 .0449891 .1564583
jul | .1747688 .0272886 6.40 0.000 .1205634 .2289741
aug | .1919517 .0275896 6.96 0.000 .1371484 .246755
sep | .1599066 .0283165 5.65 0.000 .1036594 .2161538
oct | .1007763 .0277006 3.64 0.000 .0457525 .1558002
nov | .013306 .0273035 0.49 0.627 -.040929 .0675411
dec | .0085443 .0245607 0.35 0.729 -.0402424 .0573311
_cons | 1.00872 .1016073 9.93 0.000 .8068894 1.21055
-------------+----------------------------------------------------------------
rho | .2887052
------------------------------------------------------------------------------
Durbin-Watson statistic (original) 1.430031
Durbin-Watson statistic (transformed) 1.994739

e. Compute the first order autocorrelations (

ρ^ 1 ) for unem and prcfat. What do these
suggest about possible unit root(s)?
corrgram unem
 -1 0 1 -1 0 1
LAG AC PAC Q Prob>Q [Autocorrelation] [Partial Autocor]
-------------------------------------------------------------------------------
1 0.9410 0.9573 98.306 0.0000 |------- |-------
2 0.8928 0.0854 187.65 0.0000 |------- |
3 0.8601 0.1641 271.35 0.0000 |------ |-
4 0.8366 0.0972 351.3 0.0000 |------ |

. corrgram prcfat

-1 0 1 -1 0 1
LAG AC PAC Q Prob>Q [Autocorrelation] [Partial Autocor]
-------------------------------------------------------------------------------
1 0.7077 0.7094 55.611 0.0000 |----- |-----
2 0.4439 -0.1127 77.696 0.0000 |--- |
3 0.1802 -0.1834 81.37 0.0000 |- -|
4 -0.0559 -0.1727 81.728 0.0000 | -|

The first order autocorrelation for prcfat is .709, which is high but not necessarily a cause for concern.
For unem,
ˆ1  .950 , which is cause for concern in using unem as an explanatory variable in a
regression.

f. Estimate the model in (ii) using first differences for unem and prcfat (Do not difference the
month or policy variables.) Compare your results to those in (ii).

. reg D.prcfat t wkends D.unem spdlaw beltlaw feb-dec

Source | SS df MS Number of obs = 107

-------------+------------------------------ F( 16, 90) = 2.94
Model | .213030831 16 .013314427 Prob > F = 0.0006
Residual | .406915412 90 .004521282 R-squared = 0.3436
-------------+------------------------------ Adj R-squared = 0.2269
Total | .619946244 106 .005848549 Root MSE = .06724

------------------------------------------------------------------------------
D.prcfat | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
t | .0001433 .0004849 0.30 0.768 -.00082 .0011067
wkends | .0068097 .0072276 0.94 0.349 -.0075492 .0211685
unem |
D1. | .0125342 .0161094 0.78 0.439 -.01947 .0445385
spdlaw | -.0071825 .0237979 -0.30 0.763 -.0544612 .0400962
beltlaw | .0008251 .0265048 0.03 0.975 -.0518312 .0534814
feb | .0346228 .037046 0.93 0.352 -.0389755 .1082211
mar | .0419346 .0389248 1.08 0.284 -.0353964 .1192656
apr | .0985703 .0382988 2.57 0.012 .022483 .1746577
may | .0568102 .0374416 1.52 0.133 -.0175742 .1311946
jun | .0540339 .0347738 1.55 0.124 -.0150503 .1231182
jul | .0878394 .0331103 2.65 0.009 .02206 .1536187
aug | .0589255 .0396686 1.49 0.141 -.0198832 .1377342
sep | .0065431 .0379741 0.17 0.864 -.068899 .0819852
oct | -.0323897 .0352025 -0.92 0.360 -.1023255 .0375462
nov | -.0591083 .0354151 -1.67 0.099 -.1294666 .01125
dec | .0272794 .0363245 0.75 0.455 -.0448856 .0994445
_cons | -.126868 .1048114 -1.21 0.229 -.3350941 .0813581

This regression basically shows that the change in prcfat cannot be explained by the change in unem or
any of the policy variables. It does have some seasonality, which is why the R-squared is .344.

3. Use the data in PHILIPS.RAW for this exercise. (This follows several of the examples in
Wooldridge, but using the full set of the data, rather than only through 1996.)
 The Phillips curve posits a relationship between unemployment and inflation:

inf t−inf et =β 1 ( unemt −μ 0 ) +e t

Here inf et is the expected rate of inflation for year t that was formed in year t-1. The above formulation
posits that there is a relationship between unanticipated inflation (deviations from expectations) and
cyclical unemployment—deviations of unemployment in year t from the natural rate of unemployment,
μ0. One assumption of this model is that the natural rate of unemployment is constant.
Under the adaptive expectations model, current expected values of inflation depend on recently observed
inflation, resulting in the following:
inf t−inf t−1=β 0 + β 1 ( unemt ) +e t = ∆ inf t= β0 + β 1 ( unemt ) +e t

where β 0=−β 1 μ 0

a. Estimate this equation. Interpret the coefficients. Using your estimates, calculate the
natural rate of unemployment.

. reg inf unem

Source | SS df MS Number of obs = 56

-------------+------------------------------ F( 1, 54) = 3.58
Model | 31.599858 1 31.599858 Prob > F = 0.0639
Residual | 476.815691 54 8.8299202 R-squared = 0.0622
-------------+------------------------------ Adj R-squared = 0.0448
Total | 508.415549 55 9.24391907 Root MSE = 2.9715

------------------------------------------------------------------------------
inf | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
unem | .5023782 .2655624 1.89 0.064 -.0300424 1.034799
_cons | 1.053566 1.547957 0.68 0.499 -2.049901 4.157033

Note that the coefficient here is positive—not what we’d expect.

---------------------------------------------
b. Obtain the residuals from this estimation. Is there evidence of serial correlation in these
residuals?

. dwstat

Durbin-Watson d-statistic( 2, 56) = .8014823

This is far below 2, and well below the lower bound—see the Appendix
table D I handed out. So evidence of positive serial correlation.

Breusch Godfrey method:

. tsset year
time variable: year, 1948 to 2003
delta: 1 unit

. predict error, resid

. reg error l.error

Source | SS df MS Number of obs = 55

 -------------+------------------------------ F( 1, 53) = 27.91
Model | 155.221317 1 155.221317 Prob > F = 0.0000
Residual | 294.721637 53 5.5607856 R-squared = 0.3450
-------------+------------------------------ Adj R-squared = 0.3326
Total | 449.942953 54 8.33227692 Root MSE = 2.3581

------------------------------------------------------------------------------
error | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
error |
L1. | .5724722 .1083545 5.28 0.000 .3551407 .7898038
|
_cons | -.1118079 .3179895 -0.35 0.727 -.7496141 .5259983
------------------------------------------------------------------------------
uˆ
The coefficient on t 1 is ̂  .572 (se = .108). Again, there is evidence of positive serial correlation in
the errors (t  5.28).

a. Re-estimate this model accounting for serial correlation using the Prais-Winsten method
of FGLS. Comment on the difference in coefficient estimates.
prais inf unem

Iteration 0: rho = 0.0000

Iteration 1: rho = 0.5721
Iteration 2: rho = 0.7350
Iteration 3: rho = 0.7792
Iteration 4: rho = 0.7871
Iteration 5: rho = 0.7883
Iteration 6: rho = 0.7885
Iteration 7: rho = 0.7885
Iteration 8: rho = 0.7885
Iteration 9: rho = 0.7885

Prais-Winsten AR(1) regression -- iterated estimates

Source | SS df MS Number of obs = 56

-------------+------------------------------ F( 1, 54) = 8.39
Model | 38.377534 1 38.377534 Prob > F = 0.0054
Residual | 246.917431 54 4.57254502 R-squared = 0.1345
-------------+------------------------------ Adj R-squared = 0.1185
Total | 285.294965 55 5.18718118 Root MSE = 2.1384

------------------------------------------------------------------------------
inf | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
unem | -.7139659 .2897858 -2.46 0.017 -1.294951 -.1329804
_cons | 7.999443 2.048343 3.91 0.000 3.892762 12.10612
-------------+----------------------------------------------------------------
rho | .7885234
------------------------------------------------------------------------------
Durbin-Watson statistic (original) 0.801482
Durbin-Watson statistic (transformed) 1.913928

Note that the sign of the coefficient has flipped!

d. Then estimate the adaptive expectations model: ∆ inf t= β0 + β 1 ( unemt ) +e t

Obtain the residuals from this estimation? Is there evidence of serial correltion in these
residuals? If there is, re-estimate the model.

reg cinf unem

Source | SS df MS Number of obs = 55

-------------+------------------------------ F( 1, 53) = 6.13
 Model | 32.6324798 1 32.6324798 Prob > F = 0.0165
Residual | 282.055894 53 5.32180932 R-squared = 0.1037
-------------+------------------------------ Adj R-squared = 0.0868
Total | 314.688374 54 5.82756247 Root MSE = 2.3069

------------------------------------------------------------------------------
cinf | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
unem | -.5176487 .209045 -2.48 0.017 -.9369398 -.0983576
_cons | 2.828202 1.224871 2.31 0.025 .3714212 5.284982
------------------------------------------------------------------------------

. predict cerror, resid

(1 missing value generated)

. reg cerror l.cerror

Source | SS df MS Number of obs = 54

-------------+------------------------------ F( 1, 52) = 0.08
Model | .300505001 1 .300505001 Prob > F = 0.7798
Residual | 197.891764 52 3.80561085 R-squared = 0.0015
-------------+------------------------------ Adj R-squared = -0.0177
Total | 198.192269 53 3.73947677 Root MSE = 1.9508

------------------------------------------------------------------------------
cerror | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cerror |
L1. | -.0326971 .1163578 -0.28 0.780 -.2661861 .2007919
|
_cons | .1674464 .2654783 0.63 0.531 -.3652747 .7001676
------------------------------------------------------------------------------

Note that in this model, there is no evidence of serial correlation, the coefficient on unem
is still negative (as in the AR adjusted model).
e. An alternative model (the expectations augmented Phillips curve) allows the natural
rate of unemployment to depend on past levels of unemployment. Reestimate the above
model using changes in unemployment rather than levels as the independent variable.
Comment on the difference between your results here and those in (a).
The estimated equation in first differences is


inf t  .072  .833 unemt

(.306) (.290)
n = 55, R2 = .135
The coefficient on unem has the sign that implies an inflation-unemployment
tradeoff, and the coefficient is quite large in magnitude. In fact, the estimated
coefficient is not statistically different from –1, which would imply a one-for-one
tradeoff.

f. Compute a first order autocorrelation for unem. In your opinion, is the root close to
one?
The first order autocorrelation of unem is about .75. This is one of those tough cases: the
correlation between unemt and unemt-1 is large, but it is not especially close to one.
 g. Based on what you found from your various estimation results using different models
and on the autocorrelations in errors and in the series, explain the pattern of your
results. What would you conclude about which model is most appropriate?

The levels regressions need to be adjusted for autocorrelation. However, the results for both the Phillips
curve estimation (with the change in inflation as the dependent variable) and the augmented Phillips
equation (with first differences of both variables) do not indicate autocorrelation in the errors. The main
question is whether the correlations are driven by unit root processes in the variables of interest. Again,
given an autocorrelation in unem of .75, this is sort of borderline.