QE 8,1

FUNCTIONAL FORMS

All models-considered thus far linear in both parameters & variables.

Y = b + b2 x +b3 x-linear in parameters (LIP) & linear variables (LIV) in 1 Y = b + b2 x +b3 2 - linear in parameters ( b b2and b3 x ,1 ), 1 but not linear-all-variables [i.e. xs

(e.g b x , x appears-power- ' 2 ')], . in 3 2

2 enters -power- ' 2 '. b2

2 whilst Y = b + b2 x -not linear in all - parameters si nce 1

But-many economic phenomena-relationship between- variables not linear e.g. if -want-calculate elasticity values good, - slope coefficient gives - absolute changes - one variable given a unit change in the other.

Hence,-using-alternative functional form, - can still use OLS -calculate these elasticities.

But - use OLS, models must - linear - parameters, but not necessarily in their variables.

Although -several models, we-consider:

Log-linear model Semilog models Polynomial regression models Regression through the Origin

QE 8,2

8.1 THE LOG-LINEAR/ LOG-LOG/ DOUBLE LOG MODEL

(a) The Two-variable Model
Yi = AX iB2

- non-linear in variables.

but, taking logarithms,

ln Yi = ln A + B2 ln X 2

This can be estimated as

* Yi * = B1 + B2 X 2i + u i

where Yi * =ln Yi B1 = ln A ,
* X 2i = ln X 2i and

ui is - disturbance term
Model-now linear in parameters (and also in the transformed

variables Y* and X*).

regression can be estimated with OLS and -estimators - BLUE,

provided - usual assumptions hold - transformed model.

QE 8,3

Y Yi * ln Yi i Y Y B2 = = = = * X 2 i X 2i ln X 2i X 2i X 2i

X2 Yi

i.e., B2 measures the elasticity of Y with respect to X2, and thus can - interpreted - the %age change in Y for a given %age change in X.

Thus-in fig. (b) -slope- gives-estimate-price elasticity and since

it- straight line, the elasticity-constant throughout: known - constant elasticity model (use this model only where elasticity - expected constant). Example: Weekly lotto expenditure (Y) in relation to weekly personal disposable income (X) ($).

QE 8,4

The OLS regression based-data above give:

LnYi = -0.672 + 0.7256 lnXi p = (0.2676) (0.0001) r2 = 0.8644 and -results - interpreted as ff:
the expenditure elasticity is 0.73 i.e. if PDI increases by 1%

expenditure on Lotto on the average increases 0.73% (ep<1inelastic demand).

The r2 value of 0.8644 is that 86% of - variation in - log of Y

is explained by - variation in - log of X.

Hypotheses can - tested - normal way and hence slope

coefficient is statistically significant at all (conventional) levels whilst that of the intercept is not.

(b)

The Multiple Log-Linear Model The two-variable model- easily generalised - models with

more - one explanatory variable.

ln Yi = B1 + B 2 ln X 2 + B 3 ln X 3 + ui

The partial slope coefficients B2 and B3 also called partial

elasticity coefficients.

Thus B2 measures the elasticity of Y with respect to X2,

holding the influence of X3 constant and since X3 held constant, called partial elasticity.

QE 8,5

Example: Cobb-Douglas production function:

Y = AL K

where

L is total labour input K is total capital input A, and are parameters.

Then, taking logs,

ln Y = ln A + ln L + ln K

We can model this as:

Yi* = B1 + B2 L* + B3 K i* + u i i
w h e re B1 = in te rc ep t B2 = B3 = Yi* = lnYi L*i = ln Li K i* = ln K i a n d u i is a d istu rb a e te rm nc

OLS regression based-data above give:

Ln Yi = -1.6524 + 0.3397lnLi + 0.8460lnKi

p = (0.014) r2 = 0.995 (0.085) F = 1719.23 (0.000) (0.000)

and results can be interpreted as follows:

Holding capital input constant, if labour input increases by

1%, on the average, output increases by 0.34%. Holding labour input constant, if capital input increases by 1%, on the average, output increases by 0.85%.
Estimated coefficients: labour is individually statistically

significant at 10% level whilst capital is (individually) statistically significant at all levels.
The r2 value of 0.995 is that 99.5% of the variation in the log

of output is explained by the variation in the logs of capital and labour.

Estimated F value-so highly significant that can reject null

hypothesis that labour and capital together have no impact on output

Adding the two elasticity coefficients gives -economic

parameter- returns to scale parameter i.e. response of output to a proportional change in inputs.
Our example- these sum to 1.1857-indicating-increasing

returns to scale (why?).

8.2 COMPARING LINEAR AND LOGLINEAR MODELS

Economic theory does not always specify - particular

functional form of relationship between variables.

How-choose between competing models? Plot the data: if scattergram shows relationship-

linear then linear specification might appropriate and if shows -non-linear relationship then log-linear model-suitable.

This principle-however works only simple case of

two variable regression model, but for multiple regressions other guidelines -needed.

What about choosing models basis of

comparing r2 i.e. choose model gives highest r2?

This approach-own problems? To compare r2 values two models, the

dependent variable must-same form. And if different, then not directly comparable.

Even if-dependent variables both models same still need careful since r2 can always-increased adding more explanatory variables.

Hence instead -focussing mainly on r2 ,

need-consider factors such as :

Relevance of variables included model. Expected signs of coefficients. Their statistical significance. And other derived measures like elasticity coefficients.

8.3

THE SEMILOG MODELS

8.3.1 The log-lin (Growth) Model

Often used to measure growth rates. Consider GDP, Y. The growth rate can be modelled as follows:

Yt = Y0 (1 + r )t
where r is the compound growth rate

ln Yt = B0 + B1t + u t where B0 = ln Y0 and B1 = ln( 1 + r )

Then:

ln Yt = ln Y0 + t ln(1 + r )
This can be modelled as:

The above is called a semilog model because only one variable (in this case the dependent) appears in logarithmic form called LOG-LIN model.

Example: Population of United States (millions of people), 1970-1999.

The OLS regression based-data above give:

Ln(USpop)Yi = 5.3170 + 0.0098t

p = (0.0000) (0.0000) r2 = 0.9996

and -results can be interpreted as follows:

the slope coefficient of 0.0098 means on the average the log of Y (US population) has been increasing at the rate of 0.0098 per year or alternatively, that Y has been increasing at the rate of 0.98% per year.

i.e. in a log-lin model the slope coefficient measures the proportional or relative change in Y for a given absolute change in the explanatory variable, time, in our example.

If this relative change is multiplied by 100, -obtain %age change or growth rate.

8.3.2 The lin-log Model

previous section- considered growth model, - dependent variable was log form but explanatory variable was linear form.

If - dependent variable - linear but - explanatory variable(s) is/are logarithmic, called LIN-LOG model. e.g. we want to find out how expenditure on services (Y) behaves if total personal consumption expenditure (X) increases by a certain percentage. i.e.

Yi = 1 + 2 ln X i + ui

Thus, 2 measures the absolute change in Y if the log of X changes by one unit. Example: Quarterly expenditure on services (Y) and total personal expenditure (X) 1993-11998-3.

The OLS regression based-data above give:

Y = -17907.5 + 2431.69lnX

if -log of X changes by one unit, - absolute change in Y will be 2431.69 billions.

And since a change in the log of a number is a relative change, to calculate the absolute change in Y for a 1% change in X divide the estimated slope coefficient by 100 (i.e.
2 ). 100

In-example: if X changes by 1 %, on the average, Y will change by 24.31 billions.

There is no reason why you cannot have more complex models with more than one log term or why you cannot combine log and linear terms as explanatory variables.

8.4

POLYNOMIAL REGRESSION MODELS

Consider the model:

Yi = 0 + 1X i + 2 X i 2 + 3 X i 3 + ui

These models used extensively in applied econometric studies relating to production and cost functions.

Example:

These polynomial models can be evaluated readily by OLS, since even though the variables are perfectly correlated, the correlation is not linear.

The OLS regression based-data above give:

Yi = 141.77 + 63.48 X i - 12.96 X i 2+ 0.94 X i 3

8.5 REGRESSION THROUGH THE ORIGIN
Yi = 2Xi + ui

In this model the intercept is absent or zero.

If this is the case, the formulae for b2, its variance, and the regression variance are modified as shown on pp. 274 of Gujarati (the modifications are obvious).

However, note the following:

ei need not be zero. R2 can lie outside the range 0-1.

This model should not be used unless there are strong a priori reasons for doing so i.e. it is only appropriate if theory stipulates there should be no intercept.