Econ 314: Quantitative Economics

Session 10 Tutorial - Autocorrelation

Please print and attempt prior to the tutorial session.

1. What is autocorrelation? Why is it problematic? Explain in your own words.

• Autocorrelation occurs when the error term for the current period in time-series data are
correlated with the error terms for prior periods. Autocorrelation is problematic because it
violates a CLRM assumption, which states that the error terms are uncorrelated over time.

• As a result, OLS estimates are unbiased but inefficient in the presence of autocorrelation and,

• all standard errors, hypothesis tests and confidence intervals are incorrect.

2. Suppose you are interested in explaining variation in monthly ice cream consumption (thousands
of litres) and that you estimate the regression function (standard errors in parentheses)

𝐼𝐼𝐼𝐼𝐼𝐼 �
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡 = 247.93 + 12.22 𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 + 217.39 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑡𝑡

(98.62) (4.35) (81.63)

𝑛𝑛 = 192
𝑅𝑅 2 = 0.9287

a) How many years’ worth of data do you have? How can you tell? Explain.

• There are 16 years’ worth of data. This is because there are 192 monthly observations and
192/12=16.

b) Do you suspect that autocorrelation might be present in this model? If so, what type? Explain.

• Yes. Positive autocorrelation is likely present in these data because ice cream sales in a given
month are likely highly correlated with ice cream sales in a prior month. For example, because
ice cream sales are likely highest in summer months and lowest in winter months, it is likely
to see higher sales in November, December, January, and February and lower sales in June,
July, August, and September. As such, higher values of the error term are likely to follow
higher values of the error term while lower values of the error term are likely to follow lower
values of the error term.

c) You then obtain the following plot:

Page 1 of 4
 Based on the above, does it confirm or negate your answer in (b) above? Explain why or why
not.

• The above confirms my observation of positive autocorrelation. This is because there

seems to be some type of pattern here, where positive values follow positive values, and
negative ones follow negative values.

3. You estimate a model in which you seek to find out the number of shipments sent out for a
shipping company each month and the number of calls they get each month. You obtain
the following results:
regdw shipments calls

Source | SS df MS Number of obs = 100

-------------+------------------------------ F( 1, 98) = 0.02
Model | 6.4073466 1 6.4073466 Prob > F = 0.8923
Residual | 34082.3427 98 347.779007 R-squared = 0.0002
-------------+------------------------------ Adj R-squared = -0.0100
Total | 34088.75 99 344.330808 Root MSE = 18.649

------------------------------------------------------------------------------
shipments | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
calls | -.0136982 .1009199 -0.14 0.892 -.2139706 .1865741
_cons | 101.7095 10.18843 9.98 0.000 81.49094 121.9282
------------------------------------------------------------------------------
Durbin-Watson Statistic = 2.707735

Perform a one-sided test at the 5% level for the presence of autocorrelation. State your
null and alternative hypotheses and show/explain all reasoning.

Page 2 of 4
 H0: ρ ≥ 0 (No negative autocorrelation)
H1: ρ < 0 (Evidence for negative autocorrelation)

α = 0.05 d = 2.708

Critical values: dL = 1.65, hence 4 - dL = 2.35

dU = 1.69, hence 4 - dU = 2.31

d > 4 - dL, hence we reject H0.

• We conclude that there is evidence of negative autocorrelation.

4. You then do the following with the dataset above, obtaining the results that follow:
. gen Dshpmnts = shipments - shipments[_n-1]
(1 missing value generated)
. gen Dcalls = calls - calls[_n-1]
(1 missing value generated)
. reg Dshpmnts Dcalls, noconstant

Source | SS df MS Number of obs = 99

-------------+------------------------------ F( 1, 98) = 22.26
Model | 12178.5118 1 12178.5118 Prob > F = 0.0000
Residual | 53626.4882 98 547.209063 R-squared = 0.1851
-------------+------------------------------ Adj R-squared = 0.1768
Total | 65805 99 664.69697 Root MSE = 23.393

------------------------------------------------------------------------------
Dshpmnts | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Dcalls | -.4181033 .0886264 -4.72 0.000 -.5939795 -.2422271
------------------------------------------------------------------------------
predict et, residuals
(1 missing value generated)

. runtest et, mean

N(et <= -.0260797730374216) = 50
N(et > -.0260797730374216) = 49
obs = 99
N(runs) = 56
z = 1.11
Prob>|z| = .27

a) What remedial measure were you applying? What assumptions did you make?
• I was applying Generalised Least Squares (specifically, the first differences method)
in order to remedy the autocorrelation. I assumed that ρ = 1 in ut = ρut-1 + εt.

b) Was the remedy successful? Explain.

Page 3 of 4
 Using the results from the Run test:
H0: ρ = 0 (No autocorrelation)
H1: ρ ≠ 0 (Autocorrelation exists)

The p-value from the Run test is 0.27. Hence we fail to reject H0 at all the conventional
significance levels (0.01, 0.05, 0.10).
• We conclude that there is no autocorrelation in the corrected model. Thus, the remedy
has been successful.

Page 4 of 4