Econometrics

Eshan Shenolikar

Eshan Shenolikar Econometrics SXCA

 Data Scales
• Four broad categories: ratio scale, interval scale, ordinal scale, and nominal
scale

• Ratio Scale
• Meaningful Quantities: ratio or difference/distance
• Natural ordering
• Meaningful to make statements like: how much big/small, twice, half etc

• Interval Scale
• Allows to state the difference/distance but not ratio
• e.g. temperature, years
Eshan Shenolikar Econometrics SXCA
• Ordinal Scale
• The ordering exists but the distances between the categories cannot be
quantified

• Nominal Scale
• Variables such as gender (male, female) and marital status (married,
unmarried, divorced, separated) simply denote categories

Eshan Shenolikar Econometrics SXCA

 Dummy Variable Regression
• Categorical, qualitative
• Gender, marital status, race etc
• If there is a significant difference between wages earned by women vs men,
whites vs non-whites etc
• Quantify by constructing artificial variables binary in nature: 0 or 1
• Variables that assume such 0 and 1 values are called dummy variables.
• Such variables are thus essentially a device to classify data into mutually
exclusive categories such as male or female

Eshan Shenolikar Econometrics SXCA

 Single Dummy Variable Model
• For example, to indicate gender female (which is 1 if person is female, 0 if person
is male) is better name than gender (unclear what gender = 1 means)
• What would it mean to specify simple regression model where regressor is binary?
Consider
• wage = β0 + δ0female + u
• Under SLR.4 E (u|female) = 0,
• E (wage|female) = β0 + δ0female
• Since female only takes 0 or 1,
• E (wage|female = 0) = β0 + δ0 · 0 = β0
• E (wage|female = 1) = β0 + δ0 · 1 = β0 + δ0
• In other words, average wage for men is β0 and average wage
for women is β0 + δ0
Eshan Shenolikar Econometrics SXCA
 Caution
• If there are k categories, include k-1 dummy variables
• For each qualitative regressor, the number of dummy variables introduced
must be one less than the categories of that variable
• The category for which no dummy variable is assigned is known as the base,
benchmark, control, comparison, reference, or omitted category. And all
comparisons are made in relation to the base category
• The intercept value (β0) represents the mean value of the benchmark category
• Coefficients on dummy variables are called differential intercept coefficients
because they tell by how much the value of the category that receives the
value of 1 differs from the intercept coefficient of the benchmark category
• Selection of base is upto the researcher in case of more than 1 category

Eshan Shenolikar Econometrics SXCA

 ANOVA Models
• When all the explanatory variables observed are dummy variables
• One qualitative variable with 3 categories
• E.g. Average consumption expenditure in 3 different regions- Region1, Region
2 and Region 3
• Specification: Yi = β1 + β2D2i + β3D3i + ϵi
• D2i = 1 : if state is in Region 1, 0 : otherwise
• D3i = 1 : if state is in Region 2, 0 : otherwise
• If both the dummy variables are 0, it would be the estimate for Region 3

Eshan Shenolikar Econometrics SXCA

 Two Qualitative Variables
• 2 qualitative variables, each with 2 categories
• Literacy rate based on gender and area of residence
• Categories: male-female, rural-urban
• Yi = β1 + β2D2i + β3D3i + ϵi
• Yi = Literacy rate
• D2i = Gender ; 1 : female, 0 : otherwise
• D3i = Area ; 1 = urban, 0 : otherwise
• Benchmark = male, rural

Eshan Shenolikar Econometrics SXCA

 ANCOVA Models
• Models having a mix of qualitative and quantitative regressors are termed as
ANCOVA models
• Covariance: Measure of joint variability of 2 random variables
• These allow us to measure the effect of qualitative variables after controlling
for the quantitative regressors
• E.g. : Average consumption expenditure (per capita) in relation to region and
household size
• Yi = β1 + β2D2i + β3D3i + β4 Xi + ϵi
• Thus we get ceteris paribus effect of household size, and controlling for
household size we can check for each region

Eshan Shenolikar Econometrics SXCA

 Distributed Lagged Models
• When the specification includes not only the current but also the lagged values
of the explanatory variables, it as a distributed lag model
• On the other hand when we have the past values or lagged values of the
dependent variable among the explanatory variables it is known as an auto
regressive model
• Distributed lag model:
• Yt = α + β0 Xt + β1Xt−1 + β2 Xt−2 + ut
• Auto regressive model:
• Yt = α + β0 Xt + γYt−1 + ut
• Since the auto regressive model depicts the time path of the dependent
variable in relation to its past values it is also known as dynamic model

Eshan Shenolikar Econometrics SXCA

 Role of Lag in Economics
• Generally the dependence of a variable on the explanatory variable is rarely
instantaneous
• Very often the response is with the time lapse
• This time lapse can be termed as lag
• Yt = α + β0 Xt + β1Xt−1 + β2 Xt−2 + . . . + βk Xt−k + ut
• The above model is a distributed lag model with finite lag length of k time periods
• The coefficient β0 is known as short run or impact multiplier as it gives the
change in mean value of Y following a unit change in X which is observed in the
same time period
dYt
• This can be derived as: i.e. partial derivative of Y with respect to X
dXt

Eshan Shenolikar Econometrics SXCA

• Similarly if previous time periods of the explanatory variable are considered
and the coefficients are added then the partial sum are called interim or
immediate multipliers
• However if we consider a k period model [FDL (k)]
k

• ∑
βi = β0 + β1 + . . . + βk = β
i=0
• The final value β is known as the long run, total or distributed lag multiplier
• In order to gauge the proportion of long-run or total impact felt by a certain
time period:
· βi
• βi = ; standardised
β
Eshan Shenolikar Econometrics SXCA
 Finite Distributed Lag Models
• More variables affect the dependent variable with a lag
• yt = β0 + β1Xt + β2 Xt−1 + ut . . . . . . FDL(2)
• Suppose X = c for all time periods. At time t, X increases by one unit to c + 1
and then returns to c in the forthcoming time periods
• Setting error term to 0, we get the ceteris paribus effect
• yt−1 = β0 + β1c + β2c
• yt = β0 + β1(c + 1) + β2c
• yt+1 = β0 + β1c + β2(c + 1)
• ∴ yt − yt−1 = β1 ; impact propensity or impact multiplier
• ∴ yt+1 − yt−1 = β2
Eshan Shenolikar Econometrics SXCA
• Suppose there is a permanent increase in X
• Xs = c; s < t, Xs = c + 1; s ≥ t
• yt−1 = β0 + β1c + β2c
• yt = β0 + β1(c + 1) + β2c
• yt+1 = β0 + β1(c + 1) + β2(c + 1)
• ∴ yt − yt−1 = β1
• ∴ yt+1 − yt−1 = β1 + β2 ; long run propensity (LRP) or long run multiplier

Eshan Shenolikar Econometrics SXCA

• Suppose there is a permanent increase in X
• Impact propensity is the contemporaneous effect
• For any horizon h, the cumulative effect : β1 + β2 + . . . + βh can be interpreted
as - change in y due to a permanent one unit increase in x
• LRP = cumulative effect post all changes i.e. sum of all coefficients
• LRP = β1 + β2 + . . . + βq

Eshan Shenolikar Econometrics SXCA

 Reasons for Lags
• Psychological reasons
• Force of habit or inertia
• The process of sudden change me involved disutility
• People might consider the change in a given variable to be transitory
instead of permanent and hence might want to stick to their habits
• Technological reasons
• Not possible to change some factors in the short run in reaction to change
in their prices
• Sometimes imperfect knowledge also account for lags
• Institutional reasons
• These include contractual commitments
Eshan Shenolikar Econometrics SXCA
 Estimation
• Finite (lag) distributed-lag model

• Yt = α + β0 Xt + β1Xt−1 + β2 Xt−2 + . . . + βk Xt−k + ut

• Infinite (lag) model

• Yt = α + β0 Xt + β1Xt−1 + β2 Xt−2 + . . . + ut

Eshan Shenolikar Econometrics SXCA

 Ad Hoc Estimation
• Given by Alt & Tinbergen
• OLS can be applied since the explanatory variables are non-stochastic or
uncorrelated with the error term
• Hence all the lagged values of the regressors are also uncorrelated with the
error term
• This method suggests to proceed sequentially:
• reg Yt on Xt
• reg Yt on Xt and Xt-1
• reg Yt on Xt, Xt-1 and Xt-2
• The process stops when the coefficients start becoming insignificant or
changes sign
Eshan Shenolikar Econometrics SXCA

 Drawbacks
• No prior guide on lag length
• Statistical inference is not robust since more no of lags reduces the degrees of
freedom
• df = number of independent observations out of the total no. of
observations (including the intercept)
• df = (n – no. of parameters estimated)
• Inference may be biased on account of strong multicollinearity

Eshan Shenolikar Econometrics SXCA