Autocorrelation

• Meaning
• Consequences
• Detection
• Remedial Measures

Violation of OLS assumptions

Last lecture–heteroscedasticity and cross-sectional data

Today–autocorrelation and time series data

Meaning

The CLRM 𝑌! = 𝛽" + 𝛽# 𝑋! + 𝜀!

What happens when the following assumption related to the error term is
violated?

𝐶𝑜𝑣+𝜀! , 𝜀$ - = 𝐸+𝜀! , 𝜀$ - = 0; 𝑖 ≠ 𝑗

We say there is autocorrelation or serial correlation in the error term.

Autocorrelation refers to interdependence among the observations in the error
term.

• Autocorrelation: auto+correlation
• “correlation between members of series of observations ordered in time
[as in time series data] or space [as in cross-sectional data]”
• For instance, if one disturbance is large and positive then another
disturbance too is large and positive!
• Autocorrelation can be either positive or negative.

Page 1 of 13
Examples: output and labor and capital inputs, household incomes and
expenditures (we don’t expect the correlation but there may be)

Why is it important to not have autocorrelated error term?

• Presence of autocorrelation would mean that Yi depends not just on εi but

also on εj (j≠i).

Some plausible patterns of autocorrelation

Page 2 of 13
No autocorrelation

What reasons could potentially lead to autocorrelation?

• Inertia
o most macro indicators (for instance, price indexes, production,
employment) move in tandem with general macroeconomic
conditions (business cycles)
o intertia (to change) -> past outcomes have strong influence on the
current ones
o Earthquake, tsunami, war

• Specification error
o incorrect functional form
o omission of relevant variables
o Note that in such cases, the appropriate procedue is to address
misspecification (and not autocorrelation or heteroscedasticity)

• Cobweb phenomenon: supply reacts to price with some time delay

(because supply decisions take time to implement), farmers decision to
decide about acreage depends on the prices prevailing in the last year

Page 3 of 13
 o Suppose in the period t, Pt turns out to be lower than Pt-1, in the
next period (t+1) farmers may decide to produce less. This leads to a
cobweb pattern with farmers overproducing in one period and
underproducing in the another. In such a situation the error terms
would not be random.

• Presence of lag variables (lagged dependent variable) in the model; habit

persistence (psychological, technological, and institutional reasons)

• Data imputation: absence of data means interpolation or extrapolation;

census population is available only decadally

• Transformation: while discussing heteroscedasticity, we studied first

difference transformation

Consequences

OLS in presence of autocorrelation: are they still BLUE?

• OLS are still linear and unbiased

• But, they are no longer the best or efficient (similar to heteroscedasticity)

Consider the CLRM 𝑌% = 𝛽" + 𝛽# 𝑋% + 𝜀%

and, 𝜀% = 𝜌𝜀%&" + 𝜗% with -1<ρ<1

Here, νt satisfies the usual CLRM properties (i.e., iid)

Under the above specification,

Page 4 of 13
 '!
𝑉𝑎𝑟 +𝛽9# - ≠ ∑ )"!
unless ρ = 0

Thus, OLS gives us biased estimate of variance of parameters. Conventionally

computed confidence intervals and the t and F tests are unreliable.

As was the case with heteroscedasticity, the bias does not disappear even with an
increase in the sample size.

What if we disregard autocorrelation and go ahead with OLS?

• With autocorrelated disturbances, 𝜎; # is likely to underestimate 𝜎 # (see

Gujarati for a proof). We are likely to overestimate R2
• As mentioned above, variance of the parameters are biased and
hypotheses testing might give us misleading results. A priori it is difficult to
tell whether the bias will be positive or negative

What if we use the variance formula listed above?

• The OLS estimator is still inefficient

The BLUE estimator in presence of Autocorrelation

Like the case with heteroscedasticity, we have another estimator– known as

Generalized Least Squares (GLS)–with variance lower than the OLS. This estimator
is BLUE.

Page 5 of 13
The correction factors generally pertain to the first observation in the sample

According to Kennedy, the relative efficiency of GLS vis-à-vis OLS (i.e., the ratio of
two variances) is roughly

1 − 𝜌#
1 + 𝜌#

Detection

Graphical methods

A visual examination of the residuals (or its square)

Gives us clue not just about autocorrelation but also about heteroscedasticity
and specification error

• time sequence plot OR standardized residuals, also their properties

(descriptive statistics)
• plot residuals against their lagged values

Page 6 of 13
Formal Methods

1. Durbin-Watson Test: The DW d-statistic is defined as follows:

∑*%+#(𝜀̂% − 𝜀̂%&" )#
𝑑=
∑*%+"(𝜀̂% )#

0≤d≤4

It has been shown that

𝑑 ≈ 2(1 − 𝜌;)

where

∑ 𝜀̂% 𝜀̂%&"
𝜌; =
∑ 𝜀̂% #

Remember, 𝜀% = 𝜌𝜀%&" + 𝜗%

Page 7 of 13
Decision Rule

• d=2 means no autocorrelation

• d=0 means perfect positive correlation
• d=4 means perfect negative correlation
• what if the value of d lies in between?

Durbin and Watson derived two critical values (a lower bound dL and an upper
bound dU). If the estimated d lies outside these, a decision can be made regarding
the presence of autocorrelation. The critical values depend on the level of
significance, sample size, and the number of explanatory variables.

Steps:

• S1: Run the OLS and get residuals

• S2: Compute d
• S3: Find out dL and dU values, given sample size, level of significance, and k
• S4: Follow the decision rule

Page 8 of 13
What if we end up in no man's land (i.e., indecisive zone)

• Modified d-test (one can use the upper limit dU as the critical value; beyond
the scope of this course)
• Breush-Godfrey Test

Assumptions underlying DW test

• The regression model includes the intercept term

• The error term exhibits first-order autoregressive scheme. The test cannot
be used to detect higher-order autoregressive schemes
• The regression model does not include the lagged value(s) of the
dependent variable as the explanatory variables

2. Breush-Godfrey Test (or Lagrange Multiplier Test):

Takes care of the last two assumptions in the DW test

Steps:

S1: Estimate 𝑌% = 𝛽" + 𝛽# 𝑋% + 𝜀% and get 𝜀̂%

S2: Run an auxiliary regression of 𝜀̂% on X and p lagged values of 𝜀̂% . Get R2 from
this regression. Note the number of observations to run the regression reduces to
n-p

S3: It has been shown that (n-p) times R2 follows the chi-square distribution with
'p' df. That is,

(𝑛 − 𝑝). 𝑅 # ~𝜒,#

Decision rule: If the LHS exceeds the critical chi-square value, reject the Null that
there is no autocorrelation.

Page 9 of 13
Limitation: The value of p, the length of the lag, cannot be specified a priori. Some
experimentation is inevitable.

Remedial Measures

Use GLS

The CLRM 𝑌% = 𝛽" + 𝛽# 𝑋% + 𝜀%

with 𝜀% = 𝜌𝜀%&" + 𝜗% ; -1<ρ<1

Case 1: when ρ is known

𝑌%&" = 𝛽" + 𝛽# 𝑋%&" + 𝜀%&"

𝜌𝑌%&" = 𝜌𝛽" + 𝜌𝛽# 𝑋%&" + 𝜌𝜀%&"

Or,
𝑌% − 𝜌𝑌%&" = 𝛽" (1 − 𝜌) + 𝛽# (𝑋% − 𝜌𝑋%&" ) + 𝜀% − 𝜌𝜀%&"

𝑌% ∗ = 𝛽" ∗ + 𝛽# ∗ 𝑋% ∗ + 𝜗%

The above regression equation is known as generalized difference equation. We

can now apply OLS to the transformed variables in the above equation.

Case 2: when ρ is not known

A: Treat it as equal to 1

May be a good approximation even if ρ is quite close to 1. In this case, we can use
the first difference

∆𝑌% = 𝛽# ∆𝑋% + ∆𝜀%

∆𝑌% = 𝛽# ∆𝑋% + 𝜗%

Page 10 of 13
Note that

• the above model does not have an intercept term

• The first difference is appropriate when the coefficient of autocorrelation is
very high or the Durbin–Watson d is quite low
• useful when d < R2

B: Estimate it from DW Statistics

𝜌; = 1 − 𝑑/2

Now apply the GLS

C: Estimate it from Residuals

𝜀% = 𝜌𝜀%&" + 𝜗%

Now apply the GLS

The above methods of first estimating ρ and then using the same in GLS are
known as Feasible GLS (FGLS)

D: Iterative Methods

• Cochrane–Orcutt two-step procedure

• Hildreth–Lu method

Approximation, starting with an initial value of ρ; more useful when higher order
autocorrelation is present

Page 11 of 13
Correcting the OLS standard errors

Newey –West standard errors: Using the HAC (heteroscedasticity- and

autocorrelation-consistent) standard errors developed by Newey and West

Example, relationship between wages and productivity (output), data on the

business sector of the United States for the period 1960-2005

Note that the slope coefficient is statistically significant

R-squared is pretty high

The d-statistics seems closer to zero; Note that the sample size = 46, alpha = 0.05,
k=1 (excluding the intercept), dL = 1.475 and dU = 1.566. The estimated value is
smaller than dL, therefore, we do not reject the Null of positive autocorrelation.

An option is to use Newey-West standard errors

Page 12 of 13
Kennedy notes that the above correction eliminates the asymptotic bias though it
may not completely eliminate the small sample bias.

Page 13 of 13