FEB22005(X): Econometrics 2

Module 6 – Video 3:
Multinomial Variables
Andreas Pick

Erasmus University Rotterdam

Econometric Institute
Ordered Logit Model

The Ordered Logit [OL] model is used when

1 yi IS ordered (numerical values may not have direct meaning!)

2 xi is constant for all alternatives (so again xi )

The goal is the same as for MNL and CL:

→ Model the probability that individual i chooses for alternative j:

pij = P(yi = j),

with again 0 ≤ pij ≤ 1 for all i and j, and pi1 + pi2 + . . . + pim = 1 for all i

Other steps are different, and the ordered logit model is best
understood through a latent variable interpretation

ERASMUS SCHOOL OF ECONOMICS 1/9

Latent Variable
Use simple case (one explanatory variable), and introduce a
continuous latent variable yi∗ ,
yi∗ = β0 + β1 xi + εi , with E[εi ] = 0,
and relate this to yi through


 1 if −∞ < yi∗ ≤ τ1 ,

2 if τ1 < yi∗ ≤ τ2 ,




..



 .
yi =

 j if τj−1 < yi∗ ≤ τj ,
..



.




if τm−1 < yi∗ < ∞

m

The probability of a certain category becomes

P(yi = j) = P(τj−1 < yi∗ ≤ τj ) = P(yi∗ ≤ τj ) − P(yi∗ ≤ τj−1 ),
with τ0 = −∞ and τm = +∞, and satisfies the two usual restrictions
ERASMUS SCHOOL OF ECONOMICS 2/9
Ordered Logit Model
We have

P(yi∗ ≤ τj ) = P(β0 + β1 xi + εi ≤ τj )
= P(εi ≤ τj − (β0 + β1 xi ))
= F (τj − (β0 + β1 xi )),

where F (·) is the cdf of εi

Taking the logit form (where Λ(·) = exp(·)/(1 + exp(·))), gives

exp(τj − (β0 + β1 xi ))
P(yi∗ ≤ τj ) = Λ(τj − (β0 + β1 xi )) = ,
1 + exp(τj − (β0 + β1 xi ))

and thus

P(yi = j) = Λ(τj − (β0 + β1 xi )) − Λ(τj−1 − (β0 + β1 xi ))

ERASMUS SCHOOL OF ECONOMICS 3/9

Identification

Again there is an identification issue, as τ1 , . . . , τm−1 and β0 are not all

identified
→ Normalization required, such as β0 = 0

The model then becomes:


Λ(τ1 − β1 xi )
 for j = 1
P(yi = j) = Λ(τj − β1 xi ) − Λ(τj−1 − β1 xi ) for j = 2, . . . , m − 1

1 − Λ(τm−1 − β1 xi ) for j = m


If m = 2 this is also a “binary” logit model!

ERASMUS SCHOOL OF ECONOMICS 4/9

Marginal Effects
In the model

P(yi = j) = Λ(τj − β1 xi ) − Λ(τj−1 − β1 xi )

the Marginal Effects are

∂P(yi = j)
= (λ(τj−1 − β1 xi ) − λ(τj − β1 xi ))β1
∂xi
with
λ(τj − β1 xi ) = Λ(τj − β1 xi )(1 − Λ(τj − β1 xi ))

See:
Effect of xi on P(yi = j) depends on value of xi (as in binary logit)
∂P[yi =j]
The sign of ∂xi may differ from the sign of β1
The distance between β1 xi and the nearest two thresholds
matters (τj−1 and τj )

ERASMUS SCHOOL OF ECONOMICS 5/9

Odds Ratio

The Odds Ratio here is defined as the probability of getting at most

outcome j, relative to the probability of getting higher than j:

P(yi ≤ j) Λ(τj − β1 xi )
= = exp(τj − β1 xi ),
P(yi > j) 1 − Λ(τj − β1 xi )

and the log odds ratio is thus

τj − β1 xi

See:
For a given value of xi , if τj is large it is more expected that a
“lower” category is chosen
If β1 > 0, then when xi increases the probability that observation i
falls in a “higher” category increases

ERASMUS SCHOOL OF ECONOMICS 6/9

Example: Sovereign Credit Rating
Credit rating (AAA, AA, . . . , CCC, CC, C, etc.) denotes
credit-worthiness of companies and countries
Dataset for 62 countries over 12 years (2000 through 2011), with
variables:
Numerical credit rating (17 possible values, low number = low
credit rating; 1 = near default; 17 = very safe)
Current account balance
GDP growth
Government balance
Government debt
Inflation
Unemployment

Use ordered logit to model numerical credit rating with other variables

ERASMUS SCHOOL OF ECONOMICS 7/9

Parameter Estimates
Dependent Variable: NUMERIC_RATINGS
Method: ML - Ordered Logit (Newton-Raphson / Marquardt steps)
Sample: 1 744
Included observations: 744
Number of ordered indicator values: 17
Convergence achieved after 6 iterations
Coefficient covariance computed using observed Hessian

Variable Coefficient Std. Error z-Statistic Prob.

CURRENT_ACCOUNT_BALAN 0.019416 0.009964 1.948641 0.0513

GDP_GROWTH -0.138765 0.020703 -6.702648 0.0000
GOVERNMENT_BALANCE 0.095411 0.019366 4.926648 0.0000
GOVERNMENT_DEBT -0.000812 0.002355 -0.344586 0.7304
INFLATION -0.216752 0.017406 -12.45246 0.0000
UNEMPLOYMENT -0.123795 0.016789 -7.373415 0.0000

Limit Points

LIMIT_2:C(7) -7.390673 0.388450 -19.02604 0.0000

LIMIT_3:C(8) -6.440761 0.339189 -18.98871 0.0000
LIMIT_4:C(9) -5.482925 0.301131 -18.20777 0.0000
LIMIT_5:C(10) -5.094022 0.288221 -17.67399 0.0000
LIMIT_6:C(11) -4.830007 0.280566 -17.21521 0.0000
LIMIT_7:C(12) -4.378903 0.268702 -16.29648 0.0000
LIMIT_8:C(13) -3.855644 0.256729 -15.01837 0.0000
LIMIT_9:C(14) -3.416481 0.248790 -13.73236 0.0000
LIMIT_10:C(15) -3.102689 0.244056 -12.71302 0.0000
LIMIT_11:C(16) -2.825181 0.240339 -11.75500 0.0000
LIMIT_12:C(17) -2.577619 0.237995 -10.83054 0.0000
LIMIT_13:C(18) -2.191119 0.234893 -9.328166 0.0000
LIMIT_14:C(19) -1.870511 0.232429 -8.047676 0.0000
LIMIT_15:C(20) -1.726804 0.231520 -7.458544 0.0000
LIMIT_16:C(21) -1.454210 0.230163 -6.318166 0.0000
LIMIT_17:C(22) -1.244376 0.229092 -5.431766 0.0000

Pseudo R-squared 0.102386 Akaike info criterion 4.667775

Schwarz criterion 4.804152 Log likelihood -1714.412
Hannan-Quinn criter. 4.720344 Restr. log likelihood -1909.965
LR statistic 391.1061 Avg. log likelihood -2.304317
Prob(LR statistic) 0.000000

ERASMUS SCHOOL OF ECONOMICS 8/9

Probabilities and Log Odds Ratios
1.0 .08 1.0
.07
0.8 0.8
.06

FITTEDPROB10

FITTEDPROB17
.05
FITTEDPROB1

0.6 0.6
.04
0.4 .03 0.4

.02
0.2 0.2
.01
0.0 .00 0.0
‐10 0 10 20 30 40 50 60 ‐10 0 10 20 30 40 50 60 ‐10 0 10 20 30 40 50 60

Inflation Inflation Inflation

8 12 14
6 10 12
4 8 10
2 6 8
LOGODDS10

LOGODDS17
LOGODDS1

0 4 6
‐2 2 4
‐4 0 2
‐6 ‐2 0
‐8 ‐4 ‐2
‐10 ‐6 ‐4
‐10 0 10 20 30 40 50 60 ‐10 0 10 20 30 40 50 60 ‐10 0 10 20 30 40 50 60

Inflation Inflation Inflation

ERASMUS SCHOOL OF ECONOMICS 9/9