EC226

(continued) 1
THE UNIVERSITY OF WARWICK
Summer Examinations 2012/2013
Econometrics 1
Time Allowed: 3 Hours, plus 15 minutes reading time during which notes may be made (on
the question paper) BUT NO ANSWERS MAY BE BEGUN.
Answer ALL EIGHT questions in Section A and ANY THREE questions from Section B.
Section A carries 52 marks in total. Each of the questions in Section B is worth 16 marks.
Answer Section A questions in one booklet and Section B questions in a separate booklet.
Statistical Tables and a Formula Sheet are provided. Approved pocket calculators are allowed.
Read carefully the instructions on the answer book provided and make sure that the particulars
required are entered on each answer book.
Section A
1. A model for BMI was estimated by OLS on 1832 individuals aged above 50 year old
(robust standard errors reported in brackets).
2
ln( ) 1.006 0.665ln( ) 0.086[ln( ) ] 0.048
(0.055) (0.265) (0.032) (0.011)
i i i i i
BMI age age M e = + + +
(1)
RSS=8.874, TSS=9.556
BMI Body Mass Index.
age Age in years.
M =1 if individual is male
(a) At the 5% level, calculate the power of the test that the coefficient on M is equal
to zero, against a 2-sided alternative, given the true value of the coefficient is
0.020.
(2 marks)
(b) At the 1% significance level test the joint significance of the above coefficients.
(2 marks)
(c) At the 5% significance level test that the elasticity of BMI with respect to age is
zero for an individual aged 60, given the covariance between the coefficient on
2
ln( ) and [ln( ) ]
i i
age age is -0.006.
(2.5 marks)
EC226
(continued) 2
2. Consider a model for regional unemployment in England. Using data for 298 regions,
estimating the model by OLS yielded the results:
ln( ) 0.62 0.171ln( ) 0.218ln( ) 0.117 0.037 ln( )
(0.41) (0.087) (0.008) (0.048) (0.019)
i i i i i i i
ue w y SE SE w e = + + +
(1)
2
0.092 R = , 6.147 RSS = , ln = natural log
where ue = regional unemployment rate (%); w = regional real wages; y = regional
gdp, SE = dummy variable (1 if region is in South East England, 0 if not); L = dummy
variable (1 if region is London, 0 if not).
(a) Interpret the coefficients on the variables: ln( ) w ,
i
SE , ln( ) SE w .
(3 marks)
(b) The variables L and ln( ) w L were added into equation (1). Interpret the
coefficients on these variables as well as that on the variable ln( ) SE w .
(3.5 marks)
3. You are interested in performance of students in a Maths exam (E), which is measured in
%. Based on a random sample of 386 individuals you estimate the following model:
2
ln( ) 3.81 0.512ln( ) 0.142[ln( )] 0.025
(0.95) (0.136) (0.061) (0.006)
i i i i i
E R R Male e = + + +
(1)
where, R = Hours of revision on Maths, measured between 1 and 50 (with an average of
8 hours), and Male = 1 if individual is male and 0 otherwise.
(a) Sketch the nature of the relation specified in equation (1) on a graph of exam
performance against hours of revision.
(3.5 marks)
(b) Including a variable for performance in a test on Mathematical aptitude (A),
taken 6 months earlier, yielded the following results:
2
ln( ) 2.95 0.108ln( ) 0.002[ln( )] 0.024 0.562ln( )
(0.51) (0.026) (0.019) (0.007) (0.192)
i i i i i i
E R R Male A e = + + + +
What insights can you offer to explain the markedly different results?
(3 marks)
EC226
(continued) 3
4. Consider the following model:
1 1 2 2
, 1, ,
i i i i
y x x u i n o | | = + + + = (1)
Discuss the consequences of estimating equation (1) by OLS, in terms of the OLS
estimators, b
1
and b
2
, and the standard errors of these estimators, in each of the following
circumstances:
(a)
2i
x is divided by 100.
(1.5 marks)
(b)
2
1 1
2
2 1
1,
V( )=
1, ,
i
i n
u
i n n
o
o
=

= +

(1.5 marks)
(c)
i
y is measured with error.
(1.5 marks)
(d)
2 1
3
i i
x x = .
(2 marks)
5. Answer the following questions:
a) Describe one assumption underlying Ordinary Least Squares that we commonly rely on
which, if violated, yields either biased or inconsistent parameter estimates. Be sure to
explain why your answer is correct.
(3.5 marks)
b) What two assumptions must be satisfied for a variable to be an appropriate instrument in
an Instrumental Variables estimation?
(3 marks)
EC226
(continued) 4
6. Economists are very interested in the impact macroeconomic shocks have on time series
data. The figure below is a graph of the monthly Brazilian inflation rate from January, 1974
to June, 1993 (234 observations).
The data look like they may be a unit root, so we run a Dickey-Fuller test with 3 lags and get
the following output:
Augmented Dickey-Fuller test for unit root Number of obs = 230
-------- Interpolated Dickey-Fuller --------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
Z(t) -5.451 -3.997 -3.433 -3.133
MacKinnon approximate p-value for Z(t) = 0.0000
(Question 6 continued)
EC226
(continued) 5
(Question 6 continued)
a) Based on the results above, what do you conclude about the presence of a unit root in the
Brazilian inflation rate at a 5% significance level?
(2 marks)
b) Notice in the figure that the inflation rate drops sharply at several points in the late 1980s
and early 1990s. This reflects five government ``shock plans'' introduced with the
explicit goal of reducing the (hyper-) inflation that was then occurring. Do you think
your previous unit root test is appropriate given the presence of these shock plans? Why
or why not?
(2 marks)
c) Suppose (as is the case) you knew exactly in which months the government shock plans
were in effect. How might you test for a unit root in the Brazilian inflation rate for those
months in which the shock plans weren't in effect?
(2.5 marks)
7. Answer both of the following questions:
a) Why might it be important to include a time trend or seasonal dummies in an
econometric model? In particular, what could go wrong if you were to fail to include
them?
(3.5 marks)
b) Provide an empirical example not used in class to illustrate how failing to include a time
trend or seasonal dummies could cause the problem you described in your answer above
(you can make it up as long as it makes good economic sense).
(3 marks)
EC226
(continued) 6
8. A major topic within industrial organization is the study of the boundaries of the firm. Why
are some products exchanged in markets while others are ``traded'' within firms? And what
decides the type of firm, be it a corporation, a partnership, or some other organizational
structure?
Government regulation can sometimes play a role. There are two main types of firms in
insurance markets: mutual insurance companies and stock insurance companies. Mutual
insurance companies are (mutually) owned by their policyholders, while stock insurance
companies are owned by their shareholders. One interesting development in the US
insurance market is that mutual companies share of the market has fallen from something
like 75% as recently as 1940 to something like 15% today.
To analyze this question, a researcher collected data on 881 new firm formations between
1900 and 1949. Some key summary statistics for his data are below:
Variable | Obs Mean Std. Dev. Min Max
----------------+--------------------------------------------------------
mutual | 881 .2338252 .4235027 0 1
mutual_favored | 881 .4971623 .500276 0 1
reorg | 881 .2587968 .4382226 0 1
reorg_x_favored | 881 .1169126 .3214986 0 1
const | 881 1 0 1 1
He hypothesized that, in general, firms preferred to be stock insurance companies, but were
sometimes discouraged from doing that if the initial assets that had to be put up when the
firm was founded were higher for stock companies than for mutual companies. To explore
the importance of any such favoritism, he specified the following Dummy Dependent
Variable models:
1 2
_
it it it it
mutual mutual favored reorg u o | |
-
= + + +
1 2 3
_ _ _
it it it it it
mutual mutual favored reorg reorg x favored u o | | |
-
= + + + +
where
it
mutual
-
measured the latent benefit to company i in year t from forming a mutual
insurance company instead of a stock insurance company, _
it
mutual favored was a dummy
variable equal to 1 if it took less capital to start an insurance company as a mutual company
relative to a stock company,
it
reorg was a dummy variable equal to one if the new company
was simply a corporate re-organization of an existing insurance company, and
_ _
it
reorg x favored was the interaction of _
it
mutual favored and
it
reorg .
(Question 8 continued overleaf)
EC226
(continued) 7
(Question 8 continued)
He estimated both of these specifications using a Logit model and obtained the following
results. In the table, the key estimated coefficients are reported with their standard error in
parentheses underneath.
a) What is the marginal effect of a mutual company being favored on the probability a new
company was a mutual company in the results in column (1)? Since whether a mutual
company is favored is a dummy variable, for full marks use the discrete method for
calculating marginal effects.
(4 marks)
b) What is the marginal effect of a mutual company being favored on the probability a new
company that was reorganizing was a mutual company using the results in column (2)?
(NOTE: If you do not have enough time to calculate the marginal effect, you can earn
partial credit by saying how you would calculate it.)
(2.5 marks)
EC226
(continued) 8
Section B
9. Using data from the US, a researcher estimates a model for happiness, measured on a
scale of 1-6, for a sample of 241 full-time working males estimated by OLS.

2
ln( ) 0.86 0.115ln( ) 0.056ln( ) 0.0072[ln( )] 0.056
(0.17) (0.065) (0.024) (0.002) (0.031)
-0.012
(0.005)
i i i i i
i
h w age age m
nc
= + + +
(1)
RSS = 5.826, TSS=6.927
where: h= Happiness, w = Gross monthly pay ($000), age=Age,
m = 1 if married (or cohabiting), 0 otherwise, nc = Number of children.
(a) Find the age level at which happiness is a minimum.
(2 marks)
(b) The researcher reports diagnostic test results for the Jarque-Bera normality test
of 7.56 and a Breusch-Godfrey test of serial correlation of lag 1 of 2.11.
Explain what you conclude?
(3 marks)
(c) Including a variable for
2
[ln( )] w into (1) above yielded a t-ratio on the new
variable of 1.365. Calculate the RSS for this new model along with the
2
R and
2
R for the new model.
(3 marks)
(d) The researcher re-estimated model (1) with 10 new observations added to the
original dataset, but with the addition of 10 new dummy variables one for each
of these last 10 people. The RSS for this model was 5.826. Undertaking an F-
test of this model compared to that in (1) yielded an F-test of 0. Explain what
test the researcher has undertaken. What conclusion do you draw from the
result?
(3 marks)
(e) Trying to remembering the lectures from the module EC226 and the issue with
collinearity the researcher dropped the dummy variable for the last individual
and obtained a RSS of 5.889 and this when compared to model (1) yielded an
F-test of 2.530. What should you now conclude?
(2 marks)
(f) A friend of the researcher explains that pay is likely to be endogenous, but
your problem could be resolved by the fact that the different states in the US
have different income tax rates. Explain the logic behind the friends
observation.
(3 marks)
EC226
(continued) 9
10. On 4
th
June 2012 a random sample of 394 petrol stations were surveyed in Wales and the
price of standard unleaded petrol recorded. The following specification has been
estimated to try and explain variation in petrol prices using OLS
0 1 2 3 4 5
ln( ) ln( ) ln( )
i i i i i i i
petrol d npet den ue A | | | | | | c = + + + + + + (1)
RSS = 5.292.
Where: petrol = Petrol price in , d = Distance in kilometres to the nearest petrol station
owned by a different company, npet = Petrol price charged the nearest petrol station
owned by a different company, den = population density per square kilometre measured
at the petrol station, ue = unemployment rate (measured in %) in the district around the
petrol station, and A = 1 if station located on an A-road (or motorway), 0 otherwise.
(a) Interpret the coefficient
4
| and
5
| .
(2 marks)
(b) Ten dummy variables were added to equation (1), based on 11 regions of Wales,
where Cardiff (located in South Wales) was one region and excluded from the
model as the default. The RSS from this new equation was 5.001. At the 5%
significance level test this model against model (1) being careful to outline the
null and alternative hypotheses.
(3 marks)
(c) Four dummy variables were added to equation (1), for: West Wales (made up of
2 regions), North Wales (made up of 3 regions), East Wales (made up of 2
regions) and Swansea (also located in South Wales and defined as a region). The
RSS from this new equation was 5.113 At the 5% significance level, test this
model against that in (b), being careful to outline the null and alternative
hypotheses.
(4 marks)
(d) Estimating the model in (b) separately for the 96 petrol stations based in Cardiff
and Swansea and the 298 petrol stations based in the rest of Wales yielded RSS
of 1.223 and 3.699, respectively. At the 5% significance level test this model
against the model outlined in (b), being careful to outline the null and alternative
hypotheses.
(4 marks)
(e) Looking at the model in (1), briefly outline why you might be worried about the
reliability of the OLS results.
(3 marks)
EC226
(continued) 10
11. Consider the following model for the performance of secondary schools in each of 228
local authority averaged over a five year period (2004-2008) as the percentage of all its
students gaining 5+ A-C grades in GCSE exams.:
( ) ( )
3 4 5
ln( ) ln ln ln( )
i i i i i i
P N Y F S | | | | | c
1 2
= + + + + + (1)
where P is the percentage of students with 5+ A-C GCSE results averaged over the
period 2004-2008, N is the average family income level in each local authority in 2003,
Y is the average local authority expenditure per pupil over the period 2004-2008, F is
the proportion of the local authority children getting free school meals over the period
2004-2008 and S is the proportion of the children in the local authority whose 2
nd
language is English in 2006.
The results from estimating this equation using Ordinary Least Squares (with robust
standard error in parentheses):
( ) ( ) ln( ) 2.64 0.067ln 0.215ln 0.118ln( ) 0.0153
(0.032) (0.062) (0.044) (0.082)
i i i i i i
P N Y F S e = + + +
(2)
RSS =10.85, ESS = 3.91.
(a) Explain why we may want to estimate the model with robust standard errors. Discuss
the consequences and risks of using and not using robust standard errors.
(3 marks)
(b) Some authors have suggested using local authority density as an instrument for
average local authority expenditure per pupil. Discuss whether or not such a variable
might be a valid instrument.
(3 marks)
(c) Explain how you undertake the J-test for instrument exogeneity. Why cannot the J-
test be conducted to test the appropriateness if the instrument local authority density?
(3 marks)
(d) The OLS coefficient estimate [standard error] on the local authority expenditure
(ln(Y)) variable from estimating (1) is 0.215 [0.062]. The corresponding IV
coefficient estimate [standard error] on this variable, from using two instruments (of
which density is one), is 0.124 [0.102], with an instrument relevance test of 4.81 and
a J-test of 6.62. What do you conclude?
(4 marks)
(e) In 2006 Local Authorities who had more than 5.2% of the students of secondary
school age whose first language was not English were given an extra 10% to spend
per pupil from central Government funds. Briefly layout a model which you could
estimate which attempts to identify the effect of local authority expenditure per pupil
on percentage of students with 5+ GCSE results.
(3 marks)
EC226
(continued) 11
12. This question addresses the economic returns to migration, a key issue facing many
potential migrants seeking a better life in another country. To address this question, several
authors recently collected detailed historical data on the migration of men from Norway to
the United States during the age of mass migration (1850-1913), during which long-
distance migration happened at an exceptionally high rate. Like many economic historians,
they argued that understanding the returns to migration in this period provides insights
about the incentives to migrate today.
They focused on migration of men between Norway and the US largely for data reasons.
Norway has digitized census data from 1865 and 1900 and the authors collected US census
records on Norwegian-born men living in the US in 1900 using the genealogy website
Ancestry.com. They then matched men by name and age from their birth families in
Norway in 1865 to the labor market in either Norway or the US in 1900. Clever!
The basic regression they wished to run was a simple earnings equation:
1 2 3
ln
i i i i i
wage age agesq US u o | | | = + + + +
where ln
i
wage is the log earnings of individual i (given as the mean year-1900 earnings
of i 's occupation in either Norway or the US),
i
age and
i
agesq are the age and its square of
individual i in 1900, and
i
US is a dummy variable equal to one if individual i migrated to
the US by 1900. Summary statistics for the key variables in the analysis are below.
For reasons that will become clear later in the question, the number of men in i 's family is
important and is included in the summary statistics.
Variable | Obs Mean Std. Dev. Min Max
-------------+-------------------------------------------------------------------
wage | 2655 431.7593 224.9274 90.23226 1500
lnwage | 2655 5.946887 .4900244 4.502387 7.313221
age | 2655 43.72618 3.573321 38 50
agesq | 2655 1924.742 313.7008 1444 2500
US | 2655 .1227872 .3282546 0 1
men_in_family| 2655 2.165348 .4245713 2 4
(Question 12 continued overleaf)
EC226
(continued) 12
(Question 12 continued)
They began with a simple OLS regression. Column (1) below reports the results. In the
table, the key estimated coefficients are reported with their standard error in parentheses
underneath.
a) Suppose for the moment that whether someone migrated to the US or not was random.
What does the coefficient on the US dummy,
3

| , imply about the returns to migration in

this sample of Norwegian men?
(3 marks)
Of course, migration is not random: men choose to migrate. The authors decide to try to control
for the non-random migration process by including household effects. Let i continue to index
individuals, but now let j index households. Of course, there can be multiple men per household
(e.g. brothers), some of whom migrated to the US and some of whom didnt. Here is their new
econometric model:
1 2 3 4
2
ln
J
ij ij ij ij j j ij
j
wage age agesq US d u o | | | |
=
= + + + + +

where each of the variables is defined as above (but now includes a j subscript) and
j
d is a
dummy variable for each of the J households in the data (with associated parameter
4 j
| ). There
are 1,259 households for the 2,655 men, thus 1,258
4
| s.
(Question 12 continued overleaf)
EC226
(continued) 13
(Question 12 continued)
b) It turns out this regression is essentially a fixed-effects regression. Suppose you wanted
to estimate
1
| ,
2
| , and
3
| without cluttering up the results estimating all the
4
| 's.
How would you do that? In particular, how would you transform the data so that you
didn't have to include those 1,000+ household dummy variables? (If you find it difficult
to describe what you would do using words, feel free to simply define new variables and
write down the new regression equation and how you would estimate it.)
(4 marks)
c) Note in the sample statistics above that there are at least two men in each family in the
estimation dataset. Why is this necessary?
(3 marks)
Column (2) above reports the estimates of
1
| ,
2
| ,and
3
| for the fixed-effects regression.
d) Compare the estimated returns to migration (
3

| ) in each specification.
i) What effect does including household fixed effects in your regression do to the
estimated returns to migration?
(2 marks)
ii) Lets call the unobserved heterogeneity,
j
d , household ability. What does your
answer to part (i) imply about the correlation between
ij
US and this unobserved
household ability? In particular, are migrants drawn from households that have
relatively high or relatively low unobserved household ability? For full marks, use
the bias formula to justify your answer.
(4 marks)
EC226
(continued) 14
13. Economists are very interested in whether economic development promotes the development
of democratic institutions. Each of the 34 existing OECD countries are democratic, while
many poor parts of the world are non-democratic. A keystone of both political economics
and development is that higher per-capita income causes democracy.
Whether or not that is true is a difficult question, however. To explore it, a group of
researchers recently collected a dataset on measures of democracy and income per capita
across a large number of countries. This question analyzes the data for the Ivory Coast (Cote
d'Ivoire). Democracy is measured by the Freedom House Political Rights Index in which a
country scores highly if political rights come closest to the ideals suggested by a checklist of
questions (e.g. are there free and fair elections, etc.). Income per capita is measured by GDP
per capita. The authors have annual data from 1972 to 2000 for the Ivory Coast on these
measures.
a) As we often do with time-series data, we worry that one or both of these series is a unit
root. Which type of Unit Root test (Model A, B, or C) would you use for these data?
Why?
(2 marks)
(Question 13 continued overleaf)
EC226
(continued) 15
(Question 13 continued)
It turns out that you can reject the hypothesis that the data have a unit root for both series. You
next begin to explore the relationship between democracy (democracy_score) and income per
capita (gdp_percap), running the following regressions:
1 1
_ log_ _
t t t
democracy score gdp percap u o |

= + +
1 1 3 3
_ log_ _ log_ _
t t t t
democracy score gdp percap gdp percap u o | |

= + + +
1 1 3 3
1
_ log_ _ log_ _
_
t t t
t t
democracy score gdp percap gdp percap
democracy score u
o | |

= + + +
+
Summary statistics for the important variables are:
Variable | Obs Mean Std. Dev. Min Max
---------------------+---------------------------------------------------------------
democracy_score | 29 .1839081 .0516557 .1666667 .3333333
log_gdp_percap | 29 7.743131 .1511998 7.531815 8.022284
The regression results from these specifications are in Columns (1)-(3) below. In the table, the
key estimated coefficients are reported with their standard error in parentheses underneath.
(Question 13 continued overleaf)
EC226
(End) 16
(Question 13 continued )
b) Unfortunately, we haven't done any checking yet for serial correlation. How would you
do this for the model in column (1)? Briefly describe the steps you would undertake,
being sure to write out any regression equations you use in the process as well as your
null and alternative hypotheses.
(4 marks)
c) At the bottom of each column in the row labelled ``BG'' is the test statistic for the
Breusch-Godfrey test with a one-year lag length. Test the null hypothesis of no serial
correlation for your column-(1) results.
(3 marks)
d) What is the long-run impact of a change in the log GDP per capita on democracy in your
column (3) estimates?
(3 marks)
You'll note that all the coefficients in column (3) are individually significant at the 5% level,
suggesting we've captured some important dynamics between democracy and income per capita
in the Ivory Coast.
e) Carefully look again at the raw data in the figures above. Do you think the dynamic
behavior we've identified is spurious (i.e. a statistical accident) or legitimate? Why or
why not?
(4 marks)
17
EC 226 Econometrics 1: Formulas
TwoVariableLinear RegressionModel:
i i i
Y X o | c = + +
Least Squares estimates:
2
( )( ) / ( )
t t t
b X X Y Y X X =

a Y bX =
Estimation of the error variance:
2 2 2 2
/ ( 2) { ( ) ( ) }/ ( 2)
e t t
s SSE n Y Y b X X n = =

Tests on regression slope coefficient:
0 0
: H | | = ,
0
( ) /
b
t b s | = , where
2
/ ( )
b e t
s s X X =

Standard Error of Prediction of Y

n+1
given X
n+1
:
2
0
0 0 2
1 ( )
( | ) {1 }
( )
e
t
X X
se Y X s
n X X

= + +

MultipleRegressionModel:
1 1, 2 2, ,
..
t t t k k t t
Y X X X o | | | c = + + + + +
Estimation of the error variance:
2
2
1
1 1
n
i
i
e
e
RSS
s
n k n k
=
= =

R-squared:
2
1
RSS
R
TSS
= ; R-bar-squared:
2
/ ( 1)
1
/ ( 1)
RSS n k
R
TSS n

=

Tests on regression coefficients:

(i) Single coefficient is equal to some hypothesised value:
0 0
: , 1, 2..,
i i
H i k | | = = ,
0 1
( ) / ~
i
i i b n k
t b s t |

=
(ii) All slope coefficients are equal to zero:
0 1 2
: .. 0
k
H | | | = = = = ,
2
, 1 2
1
~
1
k n k
R n k
F F
R k


=

(iii) A sub-set of coefficients are equal to zero:

0 1 2
: .. 0
j
H | | | = = = = ,
*
, 1
( ) 1
~
j n k
SSE SSE n k
F F
SSE j


=
18
(iv) The coefficients from the first sub-period are equal to those from the second sub-period:
1 2 1 2 1 2
0 1 1
: , , ,
k k
H o o | | | | = = = ,
1 2
1, 2( 1) 1 2
[ ( )] /( 1)
~
( ) /[ 2( 1)]
R
k n k
RSS RSS RSS k
F F
RSS RSS n k
+ +
+ +
=
+ +
Durbin-Watson Test statistic:
2
1
2
2
1
( )

n
i i
i
n
i
i
u u
d
u

=
=

Durbins h-statistic:
2
2
Estimate of variance of coefficient on lagged dependent variable , 1 ,
1 2
c
c
n d
h r r s
ns
= = =

Unit Root Tests:

Augmented Dickey-Fuller: Null hypothesis
0
: 0 H > , Alternative hypothesis
1
: 0 H <
Model A:
1
1
q
t t j t j t
j
y y y o c

=
A = + A +

Model B:
1
1
q
t t j t j t
j
y y y o c

=
A = + + A +

Model C:
1
1
q
t t j t j t
j
y t y y o o c

=
A = + + + A +

Test statistic:

~ MacKinnon critical values
( ) se

Error Correction Model (ECM):

0 1 1 1 1 1

t t t t t t
y x x y s u o o o

A = + A + A + A + +
LimitedDependent VariableModel
Linear Probability Model: Pr[ 1]
i i
Y X | ' = =
Logit Model:
exp( )
Pr[ 1] ( ) ( )
1 exp( )
i
i i i
i
X
Y F X X
X
|
| |
|
'
' ' = = = = A
' +
Probit Model:
1/2 2
Pr[ 1] ( ) (2 ) exp( / 2) ( )
i i
X
i i i
Y F X z dz X
|
| t |
'

' ' = = = = u
}
Interpreting coefficients:
( ) ( ( ))
( )
i i i
j
ji i i
E Y F X
X X
c c |
|
c c |
'
=
`
'
)
,
where
( ( ))
( )
( )
i
i
i
F X
f X
X
c |
|
c |
'
' =
'
and ( )
i
f X | ' is the probability density function.
For logit model: ( ) ( ) ( )[1 ( )]
i i i i
f X X X X | | | | ' ' ' ' = = A A
For probit model:
1/2 2
( ) ( ) (2 ) exp( / 2)
i i
f X X z | | | t

' ' = =