Ordered Choice Model and Preference Heterogeneity

1. Traditional ordered choice model

The traditional model is an underlaying random utility ot latent resgression
model,

0
yi = xi + "i ; i = 1; :::; N (1)
in which the latent variable yi ranging from 1 to 1 is mapped to an
observed variable yi . The variable yi is thought of as incomplete information
about an underlying yi according to the measurement model:

yi = 1 if 0 = 1 yi < 1
= 2 if 1 yi < 2
=
= J if j 1 yi < J =1 (2)
Also can be expressed in the following way:

yi = j if j 1 yi < j for j = 1 to J (3)

The model obtains the unknow marginal utilities, , as well as J + 2 un-
know threshold parameters, j , all to be estimated a sample of n observations,
indexed by i = 1; :::; n. The data consists of the covariates, xi and the observed
discrete outcome, yi = 1; 2; :::; J. With respect to disturbance, "i is a contin-
uous disturbance wiht cdf, F ("i jxi ) = F ("i ) and with density f ("i ) = F 0 ("i )
The probability associate with observed outcomes are given for the following
equation:

0 0
Pr [yi = mjxi ] = Pr "i j xi Pr "i
xi ; j = 0; 1; :::; J j 1
(4)
with a full set of normalizations, the likelihood function for estimation of the
model parameters is given for:

0 0
Pr [yi = mjxi ] = F j xi F j 1 xi ; j = 0; 1; :::; J (5)

The partial e¤ects in the ordered choice model are:

0 0
@ Pr [yi = mjxi ] @F j xi @F j 1 xi
= (6)
@xk @xk @xk
0 0
= k f j 1 xi f j xi (7)

1
 2. Random parameters and heteregeneity in the ordered choice
model

I start from the base case, assume homogeneous parameters,

0 0
Pr [yi = mjxi ] = F j xi F j 1 xi > 0; j = 0; 1; :::; J (8)

In order to model heterogeneity in the utility function across individuals, I

assume that random parameters follow a distribuction, so that:

i = +W i (9)

where is the mean paramater. W is a diagonal matrix whose elements

are standard disviation and i unobserved stochastic term, and i follows a
distibution i ( ; ). Then the standard expression of (1) can expressed:
0
yi = i xi + "i (10)
The conditional probability outcome m is

0 0
Pr [yi = mjxi ; i] =F j i xi F j 1 i xi (11)

where i vary with unobserved terms. However, the probability in (11) con-
tains the unobserved random terms i . The terms that enter to the log likelihood
function for estimation purposes must be unconditional on the unobservables.
Then, they are integrated out, thus, I have the next equation:

Z
0 0
Pr [yi = mjxi ] = F j i xi F j 1 i xi f ( i) d i (12)
i

The model is estimated by simulated maximun likelihood. Then, the simu-

lated log likelihood similated is given for:

log Ls ( ; ; W )

XN R
1X 0 0
= F j ( +W ir ) xi F j 1 ( +W ir ) xi (13)
n=1
R r=1

2
 where ir is of R multivariate random draws for simulation. The parameters
estimated must be transformed to yield estimates of the marginal changes, that
is, to determine how a marginal change in one regressor changes the distribution
of all the outcome probabilities.
Taking the partial derivate of (13) with respect to xi yields the marginal
e¤ect,

Z " 0 0
#
@ Pr [yi = mjxi ] @F j i xi @F j 1 i xi
= f ( i) d i
@xk @xk @xk
i
Z
0 0
= f j 1 i xi f j i xi ik f ( i) d i (14)
i

Where ik = k + W ik . Similarly to log likelihood function, the partial

e¤ect must be computed by simulation. Interpretation using the marginal e¤ects
can be misleading when an independent variable is a dummy variable. Hence,
it is more appropiated to calculate the discrete change which is the change in
the predicted probability for a change in xi from the start value 0 to the end
value 1.