Ordered Logit Models – Basic & Intermediate Topics

Richard Williams, University of Notre Dame, https://www3.nd.edu/~rwilliam/

Last revised February 9, 2021

This is adapted heavily from Menard’s Applied Logistic Regression analysis; also, Borooah’s Logit and Probit:
Ordered and Multinomial Models; Also, Hamilton’s Statistics with Stata, Updated for Version 7.

For a more detailed discussion with additional examples, see Williams, R. A., & Quiroz, C. (2019). Ordinal
Regression Models. In P. Atkinson, S. Delamont, A. Cernat, J.W. Sakshaug, & R.A. Williams (Eds.), SAGE
Research Methods Foundations. doi: 10.4135/9781526421036885901.
https://methods.sagepub.com/Foundations/ordinal-regression-models

We have talked about the analysis of dependent variables that have only two possible values, e.g.
lives or dies, wins or loses, gets an A or doesn’t get an A. Of course, many dependent variables
of interest will have more than two possible categories. These categories might be unordered
(doesn’t move, moves South, moves East) or ordered (high, medium, low; favors more
immigration, thinks the level of immigration is about right, favors less immigration). We will
briefly discuss techniques for handling each of these.

Ordinal Regression

As Menard notes, when dependent variables are measured on an ordinal scale, there are many
options for their analysis. These include

• Treating the variable as though it were continuous. In this case, just use OLS regression
or the other techniques we have discussed for continuous variables. Certainly, this is
widely done, particularly when the DV has 5 or more categories. Since this is probably
the easiest approach for readers to understand, sometimes the other approaches are tried
just to confirm that the use of OLS does not seriously distort the findings.

• Ignoring the ordinality of the variable and treating it as nominal. i.e. use multinomial
logit techniques like those we will discuss later. The key problem here is a loss of
efficiency. By ignoring the fact that the categories are ordered, you fail to use some of the
information available to you, and you may estimate many more parameters than is
necessary. This increases the risk of getting insignificant results. But, your parameter
estimates still should be unbiased.

• Treating the variable as though it were measured on an ordinal scale, but the ordinal scale
represented crude measurement of an underlying interval/ratio scale. For example, the
categories “High, Medium, Low” might be rough measures for Socio-economic status or
intelligence. Ordered logit models can be used in such cases, and they are the primary
focus of this handout.

Menard cautions that choosing the correct option requires careful judgment. In other words,
don’t just assume that because Stata has a routine called ologit, or that the SPSS pulldown
menu for Ordinal Regression brings up PLUM, that these are necessarily the best way to go.

Ordered Logit Models – Basic & Intermediate Topics Page 1

Ordered Logit/ Proportional Odds Models. Having made that caution, I’ll now explain how
the ordered logit models estimated by SPSS PLUM and ologit work. The ordered logit model
fit by ologit is also known as the proportional odds model. The terms parallel lines model and
parallel regressions model are also sometimes used, for reasons we will see in a moment.

1. In the ordered logit model, there is an observed ordinal variable, Y.

2. Y, in turn, is a function of another variable, Y*, that is not measured.

a. In the ordered logit model, there is a continuous, unmeasured latent variable Y*,
whose values determine what the observed ordinal variable Y equals.

b. The continuous latent variable Y* has various threshold points. (κ is the Greek
small letter Kappa.) Your value on the observed variable Y depends on whether or not you have
crossed a particular threshold. For example, when M = 3

Yi = 1 if Y*i is ≤ κ1

Yi = 2 if κ1 ≤ Y*i ≤ κ2

Yi = 3 id Y*i ≥ κ2

For example, it might be that if your score on the unobserved latent variable Y* was 37 or less,
your score on Y would be 1; if your Y* score was between 37 and 53, Y would equal 2; and if
your Y* score was above 53, Y would equal 3.

Put another way, you can think of Y as being a collapsed version of Y*, e.g. Y* can take on an
infinite range of values which might then be collapsed into 5 categories of Y.

3. So, what does Y* equal? How do you estimate this model?

a. In the population, the continuous latent variable Y* is equal to

K
Y *i = ∑ β k X ki + ε i = Z i + ε i
k =1

Note that there is a random disturbance term, which, in this case, has a standard logistic
distribution (mean of 0 and variance of 3.29; a N(0, 1) distribution is also often used). This
reflects the fact that relevant variables may be left out of the equation, or variables may not be
perfectly measured.

b. The Ordered Logit Model estimates part of the above:

K
Z i = ∑ β k X ki = E (Y *i )
k =1

c. Note that, because of the random disturbance term, the unmeasured latent variable
Y* can be either higher or lower than Z. By way of analogy, the typical person with 12 years of

Ordered Logit Models – Basic & Intermediate Topics Page 2

education might make $30,000 a year; but any specific person with 12 years of education may
make more than that or less than that. Because of the disturbance term, i.e. because Z is not a
perfect measure of Y*, you will incorrectly classify some cases as falling within one range when
they actually fall within another. But, because you know the distribution of the error term, you
can also estimate what the probability of error is.

d. The K βs and the M-1 κs are parameters that need to be estimated. Once you have
done so, using the corresponding sample estimates for each case you compute
K
Zi = ∑ βk X k
k =1

Note that there is no intercept term. You then use the estimated M-1 cutoff terms to estimate the
probability that Y will take on a particular value.

OPTIONAL: The formulas are

exp( X i β − κ j )
P (Y
=i > j) = , j 1 , 2, ..., M − 1 , which implies
1 + [exp( X i β − κ j )]

exp( X i β − κ1 )
P(Yi = 1) = 1 −
1 + [exp( X i β − κ1 )]

exp( X i β − κ j −1 ) exp( X i β − κ j )
P(Yi =
j) = − 2,..., M − 1
j=
1 + [exp( X i β − κ j −1 )] 1 + [exp( X i β − κ j )]

exp( X i β − κ M −1 )
P (=
Yi M
= )
1 + [exp( X i β − κ M −1 )]

In the case of M = 3, these equations simplify to

1
P (Y = 1) =
1 + exp( Z i − κ1 )

1 1
P (Y = 2) = −
1 + exp( Z i − κ 2 ) 1 + exp( Z i − κ1 )

1
P (Y = 3) = 1 −
1 + exp( Z i − κ 2 )

4. Hence, using the estimated value of Z and the assumed logistic distribution of the
disturbance term, the ordered logit model can be used to estimate the probability that the
unobserved variable Y* falls within the various threshold limits.

Ordered Logit Models – Basic & Intermediate Topics Page 3

NOTE: As Long points out, you can also motivate the ordered logit model by thinking of it as a
nonlinear probability model, i.e. you predict the probability of a 1, a 2, etc. You don’t have to
rely on the notion of an underlying y*, and some prefer not to.

Example. In Statistics With Stata, Updated for Version 7, Hamilton presents a fascinating
example that shows that the space shuttle Challenger disaster of January 28, 1986, might have
been averted had NASA officials heeded the warning signs. Data cover the first 25 flights of the
U.S. space shuttle. For each flight, the following variables are measured:

Distress The number of “thermal distress incidents” in which hot gas damaged
the joint seals of a flight’s booster rockets. Damage to the joint seals
helped lead to the Challenger disaster. This is the DV. It is coded 1 =
None, 2 = 1 or 2, and 3 = 3 plus.

Temp The calculated joint temperature at launch time. Temperature depends

largely on weather. Colder temperatures cause the rubber o-rings
sealing the booster rocket joints to become less flexible and hence
more likely to have problems.

Date Date, measured in days elapsed since January 1, 1960 (an arbitrary
starting point). The rationale for this variable is that undesirable
changes in the shuttle program and aging hardware may have caused
launches to become more risky across time.

Here is the data:

flight distress temp date z (computed)

1 none 66 7772 14.09598
2 1 or 2 70 7986 14.10568
3 none 69 8116 14.70623
4 MISSING 80 8213 13.11785
5 none 68 8350 15.64853
6 1 or 2 67 8494 16.29509
7 none 72 8569 15.67466
8 none 73 8642 15.74117
9 none 70 8732 16.55703
10 1 or 2 57 8799 19.03107
11 3 plus 63 8862 18.19784
12 3 plus 70 9008 17.46397
13 none 78 9044 16.19526
14 none 67 9078 18.21411
15 3 plus 53 9155 20.89439
16 3 plus 67 9233 18.72344
17 3 plus 75 9250 17.3923
18 3 plus 70 9299 18.42019

Ordered Logit Models – Basic & Intermediate Topics Page 4

 19 1 or 2 81 9341 16.65107
20 1 or 2 76 9370 17.61325
21 none 79 9407 17.2147
22 3 plus 75 9434 17.99692
23 1 or 2 76 9461 17.91227
24 3 plus 58 9508 21.18747
25 (Challenger) MISSING 31 9524 25.92117

Here is what Stata’s ologit gives you when Distress is regressed on Date and Temp.
Challenger and one flight with MD are excluded, yielding an N of 23 cases.
. use https://www3.nd.edu/~rwilliam/statafiles/shuttle2.dta, clear
(First 25 space shuttle flights)

. * Date has been added to shuttle2.dta, but here is the command that created it.
. * gen date = mdy( month, day, year)

. ologit distress date temp

Iteration 0: log likelihood = -24.955257
Iteration 1: log likelihood = -18.871284
Iteration 2: log likelihood = -18.79755
Iteration 3: log likelihood = -18.79706
Iteration 4: log likelihood = -18.79706

Ordered logistic regression Number of obs = 23

LR chi2(2) = 12.32
Prob > chi2 = 0.0021
Log likelihood = -18.79706 Pseudo R2 = 0.2468

------------------------------------------------------------------------------
distress | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
date | .003286 .0012662 2.60 0.009 .0008043 .0057677
temp | -.1733752 .0834475 -2.08 0.038 -.3369293 -.0098212
-------------+----------------------------------------------------------------
/cut1 | 16.42813 9.554822 -2.298978 35.15524
/cut2 | 18.12227 9.722302 -.933092 37.17763
------------------------------------------------------------------------------

Here is how to interpret the results:

* The model chi-square is 12.32 with 2 d.f. This is highly significant, and tells us that date
and/or temp has a significant effect on the number of thermal distress incidents.

* McFadden R2 (aka pseudo R2) is

Pseudo R2 = Model L2/ DEV0 = 12.32/49.91 = .247

(Remember, DEV0 = -2 * LL0 = -2 * -24.955257 = 49.91)

* The positive coefficient for DATE means that the likelihood of distress incidents did
increase with time. Similarly, the negative coefficient for TEMP implies that colder tempers
increased the likelihood of having distress incidents.

Ordered Logit Models – Basic & Intermediate Topics Page 5

* The threshold parameters of 16.4281 and 18.1223 tell us the following. Since there are
three possible values for Y (i.e. M = 3), the values for Y are

Yi = 1 if Y*i is ≤ 16.4281

Yi = 2 if 16.4281 ≤ Y*i ≤ 18.1223

Yi = 3 if Y*i ≥ 18.1223

As usual, we can look at the sign and significance of coefficients when interpreting them, but it
helps to plug in some hypothetical or real data values to get a better feel for the coefficients’
meaning.

Shuttle flight 13: temperature was 78 on launch date and date equaled 9044. Hence, for Flight
13, we compute

Z i = (.003286 * 9044) − .173375 * 78 = 16.195334

Note that this value is less than the lowest threshold estimate of 16.4281. For Flight 13, we can
next compute

1 1
P(Y = 1) = = = .5579
1 + exp( Z i − κ1 ) 1 + exp(16.195334 − 16.4281)
1 1
P(Y = 2) = −
1 + exp( Z i − κ 2 ) 1 + exp( Z i − κ1 )
1 1
= − = .8729 − .5579 = .315
1 + exp(16.195334 − 18.1223) 1 + exp(16.195334 − 16.4281)
1 1
P (Y = 3) = 1 − = 1− = .1271
1 + exp( Z i − κ 2 ) 1 + exp(16.195334 − 18.1223)

Hence, for Flight 13, which occurred more than a year earlier than Challenger under much
warmer conditions, the most likely outcome was that there would be no damage to the booster
joints. In fact, Flight 13 did not have any problems.

That is, for Flight 13, our estimate of Z is 16.195334 . This is our “best guess” for the value of
Y*i, and this value places Flight 13 in the Y = 1 threshold. But, because of the random
disturbance term, there is at least some chance that Y*i is larger than Zi, i.e. because of other
unmeasured influences there is some chance that Flight 13 was more at risk than our estimates
indicate. If so, this could move Flight 13 into one of the higher threshold categories, e.g. the Y*
value for Flight 13 could actually be 17, in which case Y would equal 2; or it might be 18.5, in
which case Y = 3. (Of course, it is also possible that the risk for Flight 13 was less than was
estimated.)

Given that our estimate of Z for Flight 13 is very close to the Cutoff point for Y = 1, it is not
surprising that we find P(Y = 1) = .5579, P(Y = 2) = .315, P(Y = 3) = .1271. That is, Y = 1 is the

Ordered Logit Models – Basic & Intermediate Topics Page 6

most likely value for Flight 13, but Y = 2 and Y = 3 also have fairly high probabilities. If the Z
value were much smaller, e.g. Z = 2, then it would be much less likely that Y* actually fell into a
higher threshold range, and we would find that P(Y = 1) would be much higher.

Shuttle flight 25, Challenger: Remember, Challenger’s own data was not used when
calculating these parameters. Hence, it would have been possible for a NASA official to use
these numbers on launch day to predict the likelihood of a problem. On Challenger’s Launch
Date, Date equaled 9524, and the temperature at launch time was 31 Fahrenheit (the previous
coldest launch had been at 53 Fahrenheit). Hence, for Challenger,

Z i = (.003286 * 9524) − .173375 * 31 = 25.9212

Note that this value is much higher than the upper threshold estimate of 18.1223 presented by
ologit. Using the formulas presented earlier and the threshold estimates, we can now compute
the probabilities of Challenger falling into each of the three different distress categories:

1 1
P(Y = 1) = = = .0000754
1 + exp( Z i − κ 1 ) 1 + exp(25.9212 − 16.4281)
1 1 1 1
P(Y = 2) = − = −
1 + exp( Z i − δ 2 ) 1 + exp( Z i − κ 1 ) 1 + exp(25.9212 − 18.1223) 1 + exp(25.9212 − 16.4281)
= .000410 − .0000754 = .0003346
1 1
P(Y = 3) = 1 − = 1− = .99959
1 + exp( Z i − κ 2 ) 1 + exp(25.9212 − 18.1223)

Hence, based on the experience from the previous 23 flights, there was virtually no chance that
Challenger would experience no damage to its joint seals. Indeed, it was a virtual certainty that
Challenger would experience 3 or more damage incidents.

In summary, in the case of Challenger, our estimate of Z is 25.9212, which is far above the upper
threshold limit. It is possible that the actual risk faced by Challenger was less than this, e.g.
maybe Challenger’s score on Y* was really only 23. But, it is very unlikely that the Y* value for
Challenger was actually ≤ 16.4281. It is also possible, but still not very likely, for Challenger that
16.4281 ≤ Y* ≤ 18.1223. If, say, the estimate of Z for Challenger had been 18.5, there would
have been a much better chance that the true value of Y* fell into one of the lower threshold
ranges.

Incidentally, if you run OLS regression instead, the predicted value for Challenger is 4.63 and
the predicted value for Flight 13 is 1.65. The Challenger estimate of 4.63, of course, isn’t a
legitimate value for Y, but it is consistent with the finding that launching on that day was very
risky.

Predicted Values. Stata also makes it easy to get the predicted Z values and the predicted
probability of each of the three possible outcomes. Recall that, by default, Stata’s predict
command usually computes values for all cases in the data, not just those that were included in
the analysis. Sometimes, this can be dangerous; but in this case it is nice because it gives us
predicted values for Challenger, case 25.

Ordered Logit Models – Basic & Intermediate Topics Page 7

. * Compute the Z value
. predict z, xb

. * Get the predicted probabilities for each of the three possible outcomes.
. * Specify one var for each outcome
. predict none onetwo threeplus, p
. list flight temp date distress z none onetwo threeplus

+-------------------------------------------------------------------------------+
| flight temp date distress z none onetwo threep~s |
|-------------------------------------------------------------------------------|
1. | STS-1 66 7772 none 14.09598 .9115049 .0709674 .0175277 |
2. | STS-2 70 7986 1 or 2 14.10568 .9107192 .0715853 .0176955 |
3. | STS-3 69 8116 none 14.70623 .8483726 .1198293 .031798 |
4. | STS-4 80 8213 . 13.11785 .9647798 .0285567 .0066635 |
5. | STS-5 68 8350 none 15.64853 .6855931 .236687 .0777198 |
|-------------------------------------------------------------------------------|
6. | STS-6 67 8494 1 or 2 16.29509 .5332105 .3282149 .1385746 |
7. | STS-7 72 8569 none 15.67466 .6799332 .2404531 .0796137 |
8. | STS-8 73 8642 none 15.74117 .6652908 .2500842 .084625 |
9. | STS-9 70 8732 none 16.55703 .4678189 .359285 .1728961 |
10. | STS_41-B 57 8799 1 or 2 19.03107 .0689493 .2182961 .7127547 |
|-------------------------------------------------------------------------------|
11. | STS_41-C 63 8862 3 plus 18.19784 .1455785 .3355387 .5188827 |
12. | STS_41-D 70 9008 3 plus 17.46397 .261954 .3969255 .3411205 |
13. | STS_41-G 78 9044 none 16.19526 .5579556 .3149626 .1270818 |
14. | STS_51-A 67 9078 none 18.21411 .143566 .33349 .522944 |
15. | STS_51-C 53 9155 3 plus 20.89439 .0113597 .04749 .9411502 |
|-------------------------------------------------------------------------------|
16. | STS_51-D 67 9233 3 plus 18.72344 .091512 .2625642 .6459238 |
17. | STS_51-B 75 9250 3 plus 17.3923 .2760438 .398755 .3252012 |
18. | STS_51-G 70 9299 3 plus 18.42019 .1200389 .3060272 .5739338 |
19. | STS_51-F 81 9341 1 or 2 16.65107 .4444934 .3687458 .1867608 |
20. | STS_51-I 76 9370 1 or 2 17.61325 .2341337 .3904446 .3754217 |
|-------------------------------------------------------------------------------|
21. | STS_51-J 79 9407 none 17.2147 .3129053 .3995974 .2874972 |
22. | STS_61-A 75 9434 3 plus 17.99692 .1723883 .3589076 .468704 |
23. | STS_61-B 76 9461 1 or 2 17.91227 .1848028 .3675054 .4476918 |
24. | STS_61-C 58 9508 3 plus 21.18747 .0084984 .0360675 .955434 |
25. | STS_51-L 31 9524 . 25.92117 .0000754 .0003346 .99959 |
+-------------------------------------------------------------------------------+

Note that the values for flights 13 and 25 are the same as we previously computed.

Testing Model Assumptions. There are several ways to test the proportional odds/ parallel
lines assumption of the ordered logit model. We will start with the Brant test, which also helps to
clarify exactly what the assumption is. We will explain the assumption further when we discuss
generalized ordered logit (GOLOGIT) models. Appendix A will show other methods.

Brant Test. If you have downloaded and installed spost13, you can use the brant command
to do Brant’s test of ologit’s parallel regression/ proportional odds assumption:

Ordered Logit Models – Basic & Intermediate Topics Page 8

. brant, detail

Estimated coefficients from binary logits

------------------------------------
Variable | y_gt_1 y_gt_2
-------------+----------------------
date | 0.003 0.006
| 2.17 1.80
temp | -0.173 -0.234
| -1.48 -1.86
_cons | -13.644 -36.845
| -1.34 -1.52
------------------------------------
legend: b/t

Brant test of parallel regression assumption

| chi2 p>chi2 df
-------------+------------------------------
All | 0.83 0.662 2
-------------+------------------------------
date | 0.81 0.369 1
temp | 0.22 0.643 1

A significant test statistic provides evidence that the parallel

regression assumption has been violated.

The insignificant overall chi-square value (given in the row labeled All) suggests that ologit’s
assumptions are met. (Of course, this sample is very small; with larger samples it is not at all
unusual to find that the proportional odds assumption is violated.) brant also gives tests for each
individual independent variable.

The detail option for brant clarifies why the ordered logit model is also sometimes called the
parallel lines or parallel regressions model. In the binary logits, the ordinal variable is
dichotomized. First it is category 1 versus all higher categories; then categories 1 and 2 versus all
higher categories; and so on. If the assumptions of the ordered logit model are met, then the
coefficients (other than the constants) should be the same for each logistic regression, i.e. the
regression lines will be parallel, differing only in their intercepts.

Appendix A: Ordered Logit Models – Intermediate Topics

You can go through most of this on your own. Most of this is just straightforward extensions of
techniques we talked earlier, applying them to ordinal data. This appendix covers additional
ways of testing model assumptions; adjusted predictions and marginal effects; combining
multiple imputation and margins with the mimrgns command; and doing ordinal regression
with svyset data.

Ordered Logit Models – Basic & Intermediate Topics Page 9

Other ways to Test Model Assumptions.

gologit2 LR Test and/or BIC and AIC tests. Stata’s gologit2 command (downloadable
from SSC) provides alternate tests that usually give very similar results, but which can also be
used with probit and other links. Here is an LR test:

. quietly gologit2 distress date temp, npl sto(gologit)

. quietly gologit2 distress date temp, pl sto(ologit)
. lrtest gologit ologit, stats

Likelihood-ratio test LR chi2(2) = 0.77

(Assumption: ologit nested in gologit) Prob > chi2 = 0.6815

-----------------------------------------------------------------------------
Model | Obs ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
ologit | 23 -24.95526 -18.79706 4 45.59412 50.1361
gologit | 23 -24.95526 -18.41366 6 48.82732 55.64029
-----------------------------------------------------------------------------
Note: N=Obs used in calculating BIC; see [R] BIC note

In the above you are contrasting a model in which no variables are constrained to meet the
parallel lines assumption with a model in which all variables are constrained to meet the
assumption. If the LR chi-square test is significant, at least one variable violates the parallel lines
assumption.

Also, note that you get BIC and AIC values, so you could, if you wished, use them for deciding
on whether the parallel lines assumption was met. (Also, as we will see later, gologit2 offers a
means by which the parallel lines assumption is relaxed for some variables but not others).

gologit2 Wald Test. gologit2 also offers a Wald test. You estimate an unconstrained model
(in which no variables are required to meet the parallel lines assumption, and then test whether
the coefficients of the variables in the M - 1 equations are all equal (which means parallel lines
holds). As noted below, this is the ONLY test (at least that I know of) that also works when the
data are svyset. In the shuttle example, distress has 3 possible values, so you can give the
commands

. use https://www3.nd.edu/~rwilliam/statafiles/shuttle2.dta, clear

(First 25 space shuttle flights)
. quietly gologit2 distress date temp
. test [#1 = #2]

( 1) [None]date - [1_or_2]date = 0
( 2) [None]temp - [1_or_2]temp = 0

chi2( 2) = 0.51
Prob > chi2 = 0.7764

omodel command. The omodel command (downloadable from SSC) provides yet a 4rd and
usually similar test, and works with both logit and probit links:

Ordered Logit Models – Basic & Intermediate Topics Page 10

. use https://www3.nd.edu/~rwilliam/statafiles/shuttle2.dta, clear
. omodel logit distress date temp

Ordered logit estimates Number of obs = 23

LR chi2(2) = 12.32
Prob > chi2 = 0.0021
Log likelihood = -18.79706 Pseudo R2 = 0.2468

------------------------------------------------------------------------------
distress | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
date | .003286 .0012662 2.60 0.009 .0008043 .0057677
temp | -.1733752 .0834473 -2.08 0.038 -.336929 -.0098215
-------------+----------------------------------------------------------------
_cut1 | 16.42813 9.554813 (Ancillary parameters)
_cut2 | 18.12227 9.722293
------------------------------------------------------------------------------

Approximate likelihood-ratio test of proportionality of odds

across response categories:
chi2(2) = 0.60
Prob > chi2 = 0.7405

oparallel command. Maarten Buis’s oparallel command (downloadable from SSC)

provides most of these tests and more for non-svy data, and can also provide BIC and AIC tests.
The BIC test can be especially good with large samples, since it is less likely to reject the
ordered logit model than are other tests.
. quietly ologit distress date temp
. oparallel, ic

Tests of the parallel regression assumption

| Chi2 df P>Chi2
-----------------+----------------------
Wolfe Gould | .6008 2 0.741
Brant | .8255 2 0.662
score | .535 2 0.765
likelihood ratio | .7668 2 0.682
Wald | .5063 2 0.776

Information criteria

| ologit gologit difference

------+------------------------------------
AIC | 45.59 48.83 -3.23
BIC | 50.14 55.64 -5.50

All of the tests agree that the proportional odds assumption is not violated with the shuttle data,
but again, this is a very small data set. It is quite common to find that at least one variable
violates the assumption. We will talk more later about what to do when assumptions are violated.

Adjusted Predictions and Marginal Effects. The procedures for adjusted predictions and
marginal effects are very similar for multiple outcome commands like ologit, oprobit,
mlogit, gologit2, and oglm. They are therefore discussed in a separate handout.

Multiple Imputation and mimrgns. (Thanks to Daniel Klein, the author of mimrgns, for
explaining this to me.) If you have given a command like

mi estimate: ologit y i.female x1 x2

Ordered Logit Models – Basic & Intermediate Topics Page 11

and you want to use mimrgns, and you are using Stata 14 or later, you can do something like

mimrgns female, predict(default)

If using Stata 13 or earlier, I think you have to specify an mimrgns command for each outcome,
e.g. something like

mimrgns female, predict(outcome(1) pr)

mimrgns female, predict(outcome(2) pr)
mimrgns female, predict(outcome(3) pr)

Here is a constructed example. Basically it shows that those who have previously had heart
attacks tend to be in worse health today than those who have not, i.e. they are about 14
percentage points more likely to be in poor health and 9 percentage points less likely to be in
excellent health.
. * Artificial example of mimrgns and ologit
. version 14.2
. webuse mheart1s20, clear
(Fictional heart attack data; bmi missing)
. generate health = 4 - irecode(age, 45, 55 , 65)
. label define health 1 "poor" 2 "fair" 3 "good" 4 "excellent"
. label values health health
. tab1 health

-> tabulation of health

health | Freq. Percent Cum.

------------+-----------------------------------
poor | 143 24.07 24.07
fair | 177 29.80 53.87
good | 148 24.92 78.79
excellent | 126 21.21 100.00
------------+-----------------------------------
Total | 594 100.00

. mi estimate: ologit health i.attack i.smokes bmi

Multiple-imputation estimates Imputations = 20

Ordered logistic regression Number of obs = 154
Average RVI = 0.0268
Largest FMI = 0.1314
DF adjustment: Large sample DF: min = 1,127.31
avg = 357,982.87
max = 1551474.12
Model F test: Equal FMI F( 3,20230.1) = 2.41
Within VCE type: OIM Prob > F = 0.0649

------------------------------------------------------------------------------
health | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.attack | -.6983812 .320593 -2.18 0.029 -1.326733 -.0700292
1.smokes | -.1313008 .3147298 -0.42 0.677 -.7481603 .4855587
bmi | .0586439 .038692 1.52 0.130 -.0172725 .1345604
-------------+----------------------------------------------------------------
/cut1 | .1345452 .9746209 0.14 0.890 -1.777505 2.046595
/cut2 | 1.200603 .9810356 1.22 0.221 -.7241132 3.125319
/cut3 | 2.767973 1.01049 2.74 0.006 .7854679 4.750479

Ordered Logit Models – Basic & Intermediate Topics Page 12

. * Adjusted predictions for attack
. mimrgns attack, predict(default)

Multiple-imputation estimates Imputations = 20

Predictive margins Number of obs = 154
Average RVI = 0.0199
Largest FMI = 0.0022
DF adjustment: Large sample DF: min = 3781017.30
avg = 1.43e+08
Within VCE type: Delta-method max = 1.04e+09

Expression : predict()

---------------------------------------------------------------------------------
| Margin Std. Err. t P>|t| [95% Conf. Interval]
----------------+----------------------------------------------------------------
_predict#attack |
1 0 | .2185914 .0397999 5.49 0.000 .140585 .2965979
1 1 | .3580002 .0564998 6.34 0.000 .2472626 .4687379
2 0 | .226214 .0342753 6.60 0.000 .1590356 .2933925
2 1 | .2564113 .0366576 6.99 0.000 .1845637 .3282588
3 0 | .3449389 .0418305 8.25 0.000 .2629527 .4269251
3 1 | .2679858 .0410432 6.53 0.000 .1875426 .348429
4 0 | .2102556 .0410699 5.12 0.000 .1297601 .2907511
4 1 | .1176027 .0301815 3.90 0.000 .058448 .1767573
---------------------------------------------------------------------------------

. * AMES for attack

. mimrgns, predict(default) dydx(attack)

Multiple-imputation estimates Imputations = 20

Average marginal effects Number of obs = 154
Average RVI = 0.0018
Largest FMI = 0.0046
DF adjustment: Large sample DF: min = 896,139.85
avg = 1988287.78
Within VCE type: Delta-method max = 4600516.71

Expression : predict()
dy/dx w.r.t. : 1.attack

------------------------------------------------------------------------------
| dy/dx Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.attack |
_predict |
1 | .1394088 .0646855 2.16 0.031 .0126273 .2661903
2 | .0301972 .0166543 1.81 0.070 -.0024447 .0628391
3 | -.0769531 .0375799 -2.05 0.041 -.1506084 -.0032977
4 | -.0926529 .0426154 -2.17 0.030 -.1761777 -.0091282
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

Ordered Logit: Svyset Data. Most/all of the points made before about using svyset data
continue to hold when using ologit, e.g. you can’t do Likelihood Ratio tests; many post-
estimation commands that you can use when the data are not svyset will not work when they are.

Ordered Logit Models – Basic & Intermediate Topics Page 13

Also, the only test of the parallel lines assumption that I know works with svyset data is the
gologit2 Wald test. In the following example, the variable health has 5 values so we test the
equality of the coefficients in equations 1 through 4.
. webuse nhanes2f, clear
. quietly svy: gologit2 health weight i.female age i.race, npl
. test [#1 = #2 = #3 = #4]

Adjusted Wald test

( 1) [poor]weight - [fair]weight = 0
( 2) [poor]0b.female - [fair]0b.female = 0
[Output deleted]

F( 15, 17) = 21.04

Prob > F = 0.0000

It looks to me like the margins command and some, but not all, spost13 commands continue
to work fine with svyset data, e.g. mtable works but fitstat does not.

Appendix B (Optional): Ordered Logit Models Using SPSS PLUM

Here is what SPSS PLUM gives you when Distress is regressed on Date and Temp. (At least this
is what SPSS used to give. I haven’t tried this in years.) Challenger and one flight with MD are
excluded, yielding an N of 23 cases. (Note that, in keeping with SPSS’s policy of internal
inconsistency, PLUM graciously reports things like DEV0 and McFadden’s R2, numbers which it
prefers to keep secret in its LOGISTIC REGRESSION routine.)
GET
FILE='D:\SOC593\SpssFiles\shuttle2.sav'.

COMPUTE Date = Yrmoda(year,month,day) - yrmoda(1960,1,1) .

PLUM
distress WITH date temp
/CRITERIA = CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5)
PCONVERGE(1.0E-6) SINGULAR(1.0E-8)
/LINK = LOGIT
/PRINT = FIT PARAMETER SUMMARY TPARALLEL
/SAVE = ESTPROB (Plum) .

PLUM - Ordinal Regression

Warnings

There are 46 (66.7%) cells (i.e., dependent variable levels by combinations of

predictor variable values) with zero frequencies.

Ordered Logit Models – Basic & Intermediate Topics Page 14

 Case Processing Summary

Marginal
N Percentage
DISTRESS thermal 1 none 9 39.1%
distress incidents 2 1 or 2 6 26.1%
3 3 plus 8 34.8%
Valid 23 100.0%
Missing 2
Total 25

Model Fitting Information

-2 Log
Model Likelihood Chi-Square df Sig.
Intercept Only 49.911
Final 37.594 12.316 2 .002
Link function: Logit.

Goodness-of-Fit

Chi-Square df Sig.
Pearson 42.858 42 .434
Deviance 37.594 42 .665
Link function: Logit.

Pseudo R-Square
Cox and Snell .415
Nagelkerke .468
McFadden .247
Link function: Logit.

Parameter Estimates

95% Confidence Interval

Estimate Std. Error Wald df Sig. Lower Bound Upper Bound
Threshold [DISTRESS = 1] 16.4281 9.865 2.773 1 .096 -2.908 35.764
[DISTRESS = 2] 18.1223 10.009 3.278 1 .070 -1.495 37.739
Location DATE .003286 .001 6.697 1 .010 .001 .006
TEMP -.173375 .085 4.157 1 .041 -.340 -.007
Link function: Logit.

Other Output Available from PLUM. Notice that we specified /SAVE=ESTPROB (Plum) on
the Plum command. This causes SPSS to save the predicted probabilities as part of the .sav file,
with each variable name starting with the arbitrarily chosen prefix of PLUM. Thus,
Formats plum1_1 plum2_1 plum3_1 (f8.4).
List flight temp date distress plum1_1 plum2_1 plum3_1 .

List
FLIGHT TEMP DATE DISTRESS PLUM1_1 PLUM2_1 PLUM3_1

1 66 7772 1 .9115 .0710 .0175

2 70 7986 2 .9107 .0716 .0177
3 69 8116 1 .8484 .1198 .0318
4 80 8213 . . . .
5 68 8350 1 .6856 .2367 .0777

Ordered Logit Models – Basic & Intermediate Topics Page 15

 6 67 8494 2 .5332 .3282 .1386
7 72 8569 1 .6799 .2405 .0796
8 73 8642 1 .6653 .2501 .0846
9 70 8732 1 .4678 .3593 .1729
10 57 8799 2 .0689 .2183 .7128
11 63 8862 3 .1456 .3355 .5189
12 70 9008 3 .2620 .3969 .3411
13 78 9044 1 .5580 .3150 .1271
14 67 9078 1 .1436 .3335 .5229
15 53 9155 3 .0114 .0475 .9412
16 67 9233 3 .0915 .2626 .6459
17 75 9250 3 .2760 .3988 .3252
18 70 9299 3 .1200 .3060 .5739
19 81 9341 2 .4445 .3687 .1868
20 76 9370 2 .2341 .3904 .3754
21 79 9407 1 .3129 .3996 .2875
22 75 9434 3 .1724 .3589 .4687
23 76 9461 2 .1848 .3675 .4477
24 58 9508 3 .0085 .0361 .9554
25 31 9524 . . . .

Number of cases read: 25 Number of cases listed: 25

Note that PLUM does NOT provide predicted probabilities for Case 25, which was missing on
distress (and which, of course, is the case we most wanted the predicted probabilities for!) Stata
handles this better.

Plum also provides a means of testing whether the assumptions required for its use are
reasonable:

Test of Parallel Linesa

-2 Log
Model Likelihood Chi-Square df Sig.
Null Hypothesis 37.594
General 36.827 .767 2 .682
The null hypothesis states that the location parameters (slope
coefficients) are the same across response categories.
a. Link function: Logit.

If this chi-square value is insignificant (which it is) then the use of Plum is justified. If the value
is significant, then you may want to consider another approach, such as the multinomial or
generalized ordered logit models presented later. SPSS’s Test of Parallel Lines is identical to the
Likelihood Ratio test using gologit2 that was presented earlier.

Ordered Logit Models – Basic & Intermediate Topics Page 16