clogit — Conditional (fixed-effects) logistic regression

Description
clogit fits a conditional logistic regression model for matched case – control data, also known as a
fixed-effects logit model for panel data. clogit can compute robust and cluster–robust standard errors
and adjust results for complex survey designs.
See [CM] cmclogit if you want to fit McFadden’s choice model (McFadden 1974).

Quick start
Conditional logistic regression model of y on x with matched case–control pairs data identified by idvar
clogit y x, group(idvar)
Fixed-effects logistic regression model with panels identified by idvar
clogit y x, group(idvar)
Add categorical variable a and report results as odds ratios
clogit y x i.a, group(idvar) or
Same as above, but using sampling probability weight wvar
clogit y x i.a [pweight = wvar], group(idvar) or

Menu
Statistics > Binary outcomes > Conditional logistic regression

1
 2 clogit — Conditional (fixed-effects) logistic regression

Syntax
clogit depvar [ indepvars ] [ if ] [ in ] [ weight ] , group(varname) [ options ]

depvar is treated as binary regardless of values; depvar equal to nonzero and nonmissing (typically equal
to 1) indicates a positive outcome, whereas depvar equal to 0 indicates a negative outcome.

options Description
Model
∗
group(varname) matched group variable
offset(varname) include varname in model with coefficient constrained to 1
constraints(constraints) apply specified linear constraints
SE/Robust
vce(vcetype) vcetype may be oim, robust, cluster clustvar, opg, bootstrap,
or jackknife
nonest do not check that panels are nested within clusters
Reporting
level(#) set confidence level; default is level(95)
or report odds ratios
nocnsreport do not display constraints
display options control columns and column formats, row spacing, line width,
display of omitted variables and base and empty cells, and
factor-variable labeling
Maximization
maximize options control the maximization process; seldom used
collinear keep collinear variables
coeflegend display legend instead of statistics
∗ group(varname) is required.
indepvars may contain factor variables; see [U] 11.4.3 Factor variables.
bayes, bootstrap, by, collect, fp, jackknife, mfp, mi estimate, nestreg, rolling, statsby, stepwise, and svy
are allowed; see [U] 11.1.10 Prefix commands. For more details, see [BAYES] bayes: clogit.
vce(bootstrap) and vce(jackknife) are not allowed with the mi estimate prefix; see [MI] mi estimate.
Weights are not allowed with the bootstrap prefix; see [R] bootstrap.
vce(), nonest, and weights are not allowed with the svy prefix; see [SVY] svy.
fweights, iweights, and pweights are allowed (see [U] 11.1.6 weight), but they are interpreted to apply to groups as a
whole, not to individual observations. See Use of weights below.
collinear and coeflegend do not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.

Options

 Model

group(varname) is required; it specifies an identifier variable (numeric or string) for the matched
groups. strata(varname) is a synonym for group().
offset(varname), constraints(constraints); see [R] Estimation options.
 clogit — Conditional (fixed-effects) logistic regression 3


 SE/Robust

vce(vcetype) specifies the type of standard error reported, which includes types that are derived from
asymptotic theory (oim, opg), that are robust to some kinds of misspecification (robust), that al-
low for intragroup correlation (cluster clustvar), and that use bootstrap or jackknife methods
(bootstrap, jackknife); see [R] vce option.
nonest, available only with vce(cluster clustvar), prevents checking that matched groups are nested
within clusters. It is the user’s responsibility to verify that the standard errors are theoretically correct.

 Reporting

level(#); see [R] Estimation options.

or reports the estimated coefficients transformed to odds ratios, that is, 𝑒𝑏 rather than 𝑏. Standard errors
and confidence intervals are similarly transformed. This option affects how results are displayed, not
how they are estimated. or may be specified at estimation or when replaying previously estimated
results.
nocnsreport; see [R] Estimation options.
display options: noci, nopvalues, noomitted, vsquish, noemptycells, baselevels,
allbaselevels, nofvlabel, fvwrap(#), fvwrapon(style), cformat(% fmt ), pformat(% fmt ),
sformat(% fmt ), and nolstretch; see [R] Estimation options.

 Maximization

maximize options: difficult, technique(algorithm spec), iterate(#), [no]log, trace,

gradient, showstep, hessian, showtolerance, tolerance(#), ltolerance(#),
nrtolerance(#), nonrtolerance, and from(init specs); see [R] Maximize. These options are
seldom used.
Setting the optimization type to technique(bhhh) resets the default vcetype to vce(opg).

The following options are available with clogit but are not shown in the dialog box:
collinear, coeflegend; see [R] Estimation options.
stata.com

Remarks and examples

Remarks are presented under the following headings:
Introduction
Matched case–control data
Use of weights
Fixed-effects logit

Introduction
clogit fits maximum likelihood models with a dichotomous dependent variable coded as 0/1 (more
precisely, clogit interprets 0 and not 0 to indicate the dichotomy). Conditional logistic analysis differs
from regular logistic regression in that the data are grouped and the likelihood is calculated relative to
each group; that is, a conditional likelihood is used. See Methods and formulas at the end of this entry.
4 clogit — Conditional (fixed-effects) logistic regression

Biostatisticians and epidemiologists call these models conditional logistic regression for matched
case – control groups (see, for example, Hosmer, Lemeshow, and Sturdivant [2013, chap. 7]) and fit
them when analyzing matched case – control studies with 1 ∶ 1 matching, 1 ∶ 𝑘2𝑖 matching, or 𝑘1𝑖 ∶ 𝑘2𝑖
matching, where 𝑖 denotes the 𝑖th matched group for 𝑖 = 1, 2, . . . , 𝑛, where 𝑛 is the total number of
groups. clogit fits a model appropriate for all of these matching schemes or for any mix of the schemes
because the matching 𝑘1𝑖 ∶ 𝑘2𝑖 can vary from group to group. clogit always uses the true conditional
likelihood, not an approximation. Biostatisticians and epidemiologists sometimes refer to the matched
groups as “strata”, but we will stick to the more generic term “group”.
Economists and other social scientists typically call the model fit by clogit a fixed-effects logit
model for panel data (see, for example, Chamberlain [1980]). The data used to fit a fixed-effects logit
model look exactly like the data biostatisticians and epidemiologists call 𝑘1𝑖 ∶ 𝑘2𝑖 matched case – control
data. In terms of how the data are arranged, 𝑘1𝑖 ∶ 𝑘2𝑖 matching means that in the 𝑖th group, the dependent
variable is 1 a total of 𝑘1𝑖 times and 0 a total of 𝑘2𝑖 times. There are a total of 𝑇𝑖 = 𝑘1𝑖 + 𝑘2𝑖 observations
for the 𝑖th group. This data arrangement is what economists and other social scientists call “panel data”,
“longitudinal data”, or “cross-sectional time-series data”.
So no matter what terminology you use, the computation and the use of the clogit command is the
same. The following example shows how your data should be arranged to use clogit.

Example 1
Suppose that we have grouped data with the variable id containing a unique identifier for each group.
Our outcome variable, y, contains 0s and 1s. If we were biostatisticians, y = 1 would indicate a case,
y = 0 would be a control, and id would be an identifier variable that indicates the groups of matched
case – control subjects.
If we were economists, y = 1 might indicate that a person was unemployed at any time during a year
and y = 0, that a person was employed all year, and id would be an identifier variable for persons.
If we list the first few observations of this dataset, it looks like
. use https://www.stata-press.com/data/r18/clogitid
. list y x1 x2 id in 1/11

y x1 x2 id

1. 0 0 4 1014
2. 0 1 4 1014
3. 0 1 6 1014
4. 1 1 8 1014
5. 0 0 1 1017

6. 0 0 7 1017
7. 1 1 10 1017
8. 0 0 1 1019
9. 0 1 7 1019
10. 1 1 7 1019

11. 1 1 9 1019
 clogit — Conditional (fixed-effects) logistic regression 5

Pretending that we are biostatisticians, we describe our data as follows. The first group (id = 1014)
consists of four matched persons: 1 case (y = 1) and three controls (y = 0), that is, 1 ∶ 3 matching. The
second group has 1 ∶ 2 matching, and the third 2 ∶ 2.
Pretending that we are economists, we describe our data as follows. The first group consists of 4
observations (one per year) for person 1014. This person had a period of unemployment during 1 year
of 4. The second person had a period of unemployment during 1 year of 3, and the third had a period of
2 years of 4.
Our independent variables are x1 and x2. To fit the conditional (fixed-effects) logistic model, we type
. clogit y x1 x2, group(id)
note: multiple positive outcomes within groups encountered.
Iteration 0: Log likelihood = -123.42828
Iteration 1: Log likelihood = -123.41386
Iteration 2: Log likelihood = -123.41386
Conditional (fixed-effects) logistic regression Number of obs = 369
LR chi2(2) = 9.07
Prob > chi2 = 0.0107
Log likelihood = -123.41386 Pseudo R2 = 0.0355

y Coefficient Std. err. z P>|z| [95% conf. interval]

x1 .653363 .2875215 2.27 0.023 .0898312 1.216895

x2 .0659169 .0449555 1.47 0.143 -.0221943 .1540281

Technical note
The message “note: multiple positive outcomes within groups encountered” at the top of the clogit
output for the previous example merely informs us that we have 𝑘1𝑖 ∶ 𝑘2𝑖 matching with 𝑘1𝑖 > 1 for at
least one group. If your data should be 1 ∶ 𝑘2𝑖 matched, this message tells you that there is an error in the
data somewhere.
We can see the distribution of 𝑘1𝑖 and 𝑇𝑖 = 𝑘1𝑖 + 𝑘2𝑖 for the data of the example 1 by using the
following steps:
. by id, sort: generate k1 = sum(y)
. by id: replace k1 = . if _n < _N
(303 real changes made, 303 to missing)
. by id: generate T = sum(y<.)
. by id: replace T = . if _n < _N
(303 real changes made, 303 to missing)
. tabulate k1
k1 Freq. Percent Cum.

1 48 72.73 72.73
2 12 18.18 90.91
3 4 6.06 96.97
4 2 3.03 100.00

Total 66 100.00
 6 clogit — Conditional (fixed-effects) logistic regression

. tabulate T
T Freq. Percent Cum.

2 5 7.58 7.58
3 5 7.58 15.15
4 12 18.18 33.33
5 11 16.67 50.00
6 13 19.70 69.70
7 8 12.12 81.82
8 3 4.55 86.36
9 7 10.61 96.97
10 2 3.03 100.00

Total 66 100.00

We see that 𝑘1𝑖 ranges from 1 to 4 and 𝑇𝑖 ranges from 2 to 10 for these data.

Technical note
For 𝑘1𝑖 ∶ 𝑘2𝑖 matching (and hence in the general case of fixed-effects logit), clogit uses a recursive
algorithm to compute the likelihood, which means that there are no limits on the size of 𝑇𝑖 . However,
computation time is proportional to ∑ 𝑇𝑖 min(𝑘1𝑖 , 𝑘2𝑖 ), so clogit will take roughly 10 times longer to
fit a model with 10 ∶ 10 matching than one with 1 ∶ 10 matching. But clogit is fast, so computation
time becomes an issue only when min(𝑘1𝑖 , 𝑘2𝑖 ) is around 100 or more. See Methods and formulas for
details.

Matched case–control data

Here we give a more detailed example of matched case – control data.

Example 2
Hosmer, Lemeshow, and Sturdivant (2013, 24) present data on matched pairs of infants, each pair
having one with low birthweight and another with regular birthweight. The data are matched on age of
the mother. Several possible maternal exposures are considered: race (three categories), smoking status,
presence of hypertension, presence of uterine irritability, previous preterm delivery, and weight at the
last menstrual period.
 clogit — Conditional (fixed-effects) logistic regression 7

. use https://www.stata-press.com/data/r18/lowbirth2, clear

(Applied Logistic Regression, Hosmer & Lemeshow)
. describe
Contains data from https://www.stata-press.com/data/r18/lowbirth2.dta
Observations: 112 Applied Logistic Regression,
Hosmer & Lemeshow
Variables: 9 30 Jan 2022 08:46

Variable Storage Display Value

name type format label Variable label

pairid byte %8.0g Case-control pair ID

low byte %8.0g Baby has low birthweight
age byte %8.0g Age of mother
lwt int %8.0g Mother’s last menstrual weight
smoke byte %8.0g Mother smoked during pregnancy
ptd byte %8.0g Mother had previous preterm baby
ht byte %8.0g Mother has hypertension
ui byte %8.0g Uterine irritability
race byte %9.0g race Race of mother

Sorted by:

We list the case – control indicator variable, low; the match identifier variable, pairid; and two of the
covariates, lwt and smoke, for the first 10 observations.
. list low lwt smoke pairid in 1/10

low lwt smoke pairid

1. 0 135 0 1
2. 1 101 1 1
3. 0 98 0 2
4. 1 115 0 2
5. 0 95 0 3

6. 1 130 0 3
7. 0 103 0 4
8. 1 130 1 4
9. 0 122 1 5
10. 1 110 1 5
8 clogit — Conditional (fixed-effects) logistic regression

We fit a conditional logistic model of low birthweight on mother’s weight, race, smoking behavior, and
history.
. clogit low lwt smoke ptd ht ui i.race, group(pairid) nolog
Conditional (fixed-effects) logistic regression Number of obs = 112
LR chi2(7) = 26.04
Prob > chi2 = 0.0005
Log likelihood = -25.794271 Pseudo R2 = 0.3355

low Coefficient Std. err. z P>|z| [95% conf. interval]

lwt -.0183757 .0100806 -1.82 0.068 -.0381333 .0013819

smoke 1.400656 .6278396 2.23 0.026 .1701131 2.631199
ptd 1.808009 .7886502 2.29 0.022 .2622828 3.353735
ht 2.361152 1.086128 2.17 0.030 .2323796 4.489924
ui 1.401929 .6961585 2.01 0.044 .0374836 2.766375

race
Black .5713643 .689645 0.83 0.407 -.7803149 1.923044
Other -.0253148 .6992044 -0.04 0.971 -1.39573 1.345101

We might prefer to see results presented as odds ratios. We could have specified the or option when we
first fit the model, or we can now redisplay results and specify or:
. clogit, or
Conditional (fixed-effects) logistic regression Number of obs = 112
LR chi2(7) = 26.04
Prob > chi2 = 0.0005
Log likelihood = -25.794271 Pseudo R2 = 0.3355

low Odds ratio Std. err. z P>|z| [95% conf. interval]

lwt .9817921 .009897 -1.82 0.068 .9625847 1.001383

smoke 4.057862 2.547686 2.23 0.026 1.185439 13.89042
ptd 6.098293 4.80942 2.29 0.022 1.299894 28.60938
ht 10.60316 11.51639 2.17 0.030 1.261599 89.11467
ui 4.06303 2.828513 2.01 0.044 1.038195 15.90088

race
Black 1.770681 1.221141 0.83 0.407 .4582617 6.84175
Other .975003 .6817263 -0.04 0.971 .2476522 3.838573

Smoking, previous preterm delivery, hypertension, uterine irritability, and possibly the mother’s
weight all contribute to low birthweight. Race of black and race of other are statistically insignificant
when compared with the race of white omitted group, although the race of black effect is large. We can
test the joint statistical significance of race being black (2.race) and race being other (3.race) by using
test:
. test 2.race 3.race
( 1) [low]2.race = 0
( 2) [low]3.race = 0
chi2( 2) = 0.88
Prob > chi2 = 0.6436
 clogit — Conditional (fixed-effects) logistic regression 9

For a more complete description of test, see [R] test. test presents results in coefficients rather than
odds ratios. Jointly testing that the coefficients on 2.race and 3.race are 0 is equivalent to jointly
testing that the odds ratios are 1.
Here one case was matched to one control, that is, 1 ∶ 1 matching. From clogit’s point of view, that
was not important — 𝑘1 cases could have been matched to 𝑘2 controls (𝑘1 ∶ 𝑘2 matching), and we would
have fit the model in the same way. Furthermore, the matching can change from group to group, which
we have denoted as 𝑘1𝑖 ∶ 𝑘2𝑖 matching, where 𝑖 denotes the group. clogit does not care. To fit the
conditional logistic regression model, we specified the group(varname) option, group(pairid). The
case and control are stored in separate observations. clogit knew that they were linked (in the same
group) because the related observations share the same value of pairid.

Technical note
clogit provides a way to extend McNemar’s test to multiple controls per case (1 ∶ 𝑘2𝑖 matching) and
to multiple controls matched with multiple cases (𝑘1𝑖 ∶ 𝑘2𝑖 matching).
In Stata, McNemar’s test is calculated by the mcc command; see [R] Epitab. The mcc command,
however, requires that the matched case and control appear in one observation, so the data will need to
be manipulated from 1 to 2 observations per stratum before using clogit. Alternatively, if you begin
with clogit’s 2-observations-per-group organization, you will have to change it to 1 observation per
group if you wish to use mcc. In either case, reshape provides an easy way to change the organization
of the data. We will demonstrate its use below, but we direct you to [D] reshape for a more thorough
discussion.
In example 2, we used clogit to analyze the relationship between low birthweight and various char-
acteristics of the mother. Assume that we now want to assess the relationship between low birthweight
and smoking, ignoring the mother’s other characteristics. Using clogit, we obtain the following results:
. clogit low smoke, group(pairid) or
Iteration 0: Log likelihood = -35.425931
Iteration 1: Log likelihood = -35.419283
Iteration 2: Log likelihood = -35.419282
Conditional (fixed-effects) logistic regression Number of obs = 112
LR chi2(1) = 6.79
Prob > chi2 = 0.0091
Log likelihood = -35.419282 Pseudo R2 = 0.0875

low Odds ratio Std. err. z P>|z| [95% conf. interval]

smoke 2.75 1.135369 2.45 0.014 1.224347 6.176763

 10 clogit — Conditional (fixed-effects) logistic regression

Let’s compare our estimated odds ratio and 95% confidence interval with that produced by mcc. We
begin by reshaping the data:
. keep low smoke pairid
. reshape wide smoke, i(pairid) j(low 0 1)
Data Long -> Wide

Number of observations 112 -> 56

Number of variables 3 -> 3
j variable (2 values) low -> (dropped)
xij variables:
smoke -> smoke0 smoke1

We now have the variables smoke0 (formed from smoke and low = 0), recording 1 if the control mother
smoked and 0 otherwise; and smoke1 (formed from smoke and low = 1), recording 1 if the case mother
smoked and 0 otherwise. We can now use mcc:
. mcc smoke1 smoke0
Controls
Cases Exposed Unexposed Total

Exposed 8 22 30
Unexposed 8 18 26

Total 16 40 56
McNemar’s chi2(1) = 6.53 Prob > chi2 = 0.0106
Exact McNemar significance probability = 0.0161
Proportion with factor
Cases .5357143
Controls .2857143 [95% conf. interval]

difference .25 .0519726 .4480274

ratio 1.875 1.148685 3.060565
rel. diff. .35 .1336258 .5663742
odds ratio 2.75 1.179154 7.143667 (exact)

Both methods estimated the same odds ratio, and the 95% confidence intervals are similar. clogit pro-
duced a confidence interval of [ 1.22, 6.18 ], whereas mcc produced a confidence interval of [ 1.18, 7.14 ].

Use of weights
With clogit, weights apply to groups as a whole, not to individual observations. For example, if
there is a group in your dataset with a frequency weight of 3, there are a total of three groups in your
sample with the same values of the dependent and independent variables as this one group. Weights must
have the same value for all observations belonging to the same group; otherwise, an error message will
be displayed.
 clogit — Conditional (fixed-effects) logistic regression 11

Example 3
We use the example from the above discussion of the mcc command. Here we have a total of 56
matched case – control groups, each with one case matched to one control. We had 8 matched pairs in
which both the case and the control are exposed, 22 pairs in which the case is exposed and the control is
unexposed, 8 pairs in which the case is unexposed and the control is exposed, and 18 pairs in which they
are both unexposed.
With weights, it is easy to enter these data into Stata and run clogit.
. clear
. input id case exposed weight
id case exposed weight
1. 1 1 1 8
2. 1 0 1 8
3. 2 1 1 22
4. 2 0 0 22
5. 3 1 0 8
6. 3 0 1 8
7. 4 1 0 18
8. 4 0 0 18
9. end
. clogit case exposed [w=weight], group(id) or
(frequency weights assumed)
Iteration 0: Log likelihood = -35.425931
Iteration 1: Log likelihood = -35.419283
Iteration 2: Log likelihood = -35.419282
Conditional (fixed-effects) logistic regression Number of obs = 112
LR chi2(1) = 6.79
Prob > chi2 = 0.0091
Log likelihood = -35.419282 Pseudo R2 = 0.0875

case Odds ratio Std. err. z P>|z| [95% conf. interval]

exposed 2.75 1.135369 2.45 0.014 1.224347 6.176763

Fixed-effects logit
The fixed-effects logit model can be written as

Pr(𝑦𝑖𝑡 = 1 ∣ x𝑖𝑡 ) = 𝐹 (𝛼𝑖 + x𝑖𝑡 β)

where 𝐹 is the cumulative logistic distribution

exp(𝑧)
𝐹 (𝑧) =
1 + exp(𝑧)

𝑖 = 1, 2, . . . , 𝑛 denotes the independent units (called “groups” by clogit), and 𝑡 = 1, 2, . . . , 𝑇𝑖 denotes

the observations for the 𝑖th unit (group).
12 clogit — Conditional (fixed-effects) logistic regression

Fitting this model by using a full maximum-likelihood approach leads to difficulties, however. When
𝑇𝑖 is fixed, the maximum likelihood estimates for 𝛼𝑖 and β are inconsistent (Andersen 1970; Chamber-
lain 1980). This difficulty can be circumvented by looking at the probability of y𝑖 = (𝑦𝑖1 , . . . , 𝑦𝑖𝑇𝑖 )
conditional on ∑𝑇𝑡=1
𝑖
𝑦𝑖𝑡 . This conditional probability does not involve the 𝛼𝑖 , so they are never esti-
mated when the resulting conditional likelihood is used. See Hamerle and Ronning (1995) for a succinct
and lucid development. See Methods and formulas for the estimation equation.

Example 4
We are studying unionization of women in the United States by using the union dataset; see [XT] xt.
We fit the fixed-effects logit model:
. use https://www.stata-press.com/data/r18/union, clear
(NLS Women 14-24 in 1968)
. clogit union age grade not_smsa south black, group(idcode)
note: multiple positive outcomes within groups encountered.
note: 2,744 groups (14,165 obs) omitted because of all positive or
all negative outcomes.
note: black omitted because of no within-group variance.
Iteration 0: Log likelihood = -4521.3385
Iteration 1: Log likelihood = -4516.1404
Iteration 2: Log likelihood = -4516.1385
Iteration 3: Log likelihood = -4516.1385
Conditional (fixed-effects) logistic regression Number of obs = 12,035
LR chi2(4) = 68.09
Prob > chi2 = 0.0000
Log likelihood = -4516.1385 Pseudo R2 = 0.0075

union Coefficient Std. err. z P>|z| [95% conf. interval]

age .0170301 .004146 4.11 0.000 .0089042 .0251561

grade .0853572 .0418781 2.04 0.042 .0032777 .1674368
not_smsa .0083678 .1127963 0.07 0.941 -.2127088 .2294445
south -.748023 .1251752 -5.98 0.000 -.9933619 -.5026842
black 0 (omitted)

We received three messages at the top of the output. The first one, “multiple positive outcomes within
groups encountered”, we expected. Our data do indeed have multiple positive outcomes (union = 1) in
many groups. (Here a group consists of all the observations for a particular individual.)
The second message tells us that 2,744 groups were “omitted” by clogit. When either union = 0
or union = 1 for all observations for an individual, this individual’s contribution to the log likelihood is
zero. Although these are perfectly valid observations in every sense, they have no effect on the estimation,
so they are not included in the total “Number of obs”. Hence, the reported “Number of obs” gives the
effective sample size of the estimation. Here it is 12,035 observations — only 46% of the total 26,200.
 clogit — Conditional (fixed-effects) logistic regression 13

We can easily check that there are indeed 2,744 groups with union either all 0 or all 1. We will
generate a variable that contains the fraction of observations for each individual who has union = 1.
. by idcode, sort: generate fraction = sum(union)/sum(union < .)
. by idcode: replace fraction = . if _n < _N
(21,766 real changes made, 21,766 to missing)
. tabulate fraction
fraction Freq. Percent Cum.

0 2,481 55.95 55.95

.0833333 30 0.68 56.63
.0909091 33 0.74 57.37
.1 53 1.20 58.57
(output omitted )
.9 10 0.23 93.59
.9090909 11 0.25 93.84
.9166667 10 0.23 94.07
1 263 5.93 100.00

Total 4,434 100.00

Because 2481 + 263 = 2744, we confirm what clogit did.

The third warning message from clogit said “black omitted because of no within-group variance”.
Obviously, race stays constant for an individual across time. Any such variables are collinear with the 𝛼𝑖
(that is, the fixed effects), and just as the 𝛼𝑖 drop out of the conditional likelihood, so do all variables that
are unchanging within groups. Thus, they cannot be estimated with the conditional fixed-effects model.
There are other estimators implemented in Stata that we could use with these data, such as
cloglog ... , vce(cluster idcode)
logit ... , vce(cluster idcode)
probit ... , vce(cluster idcode)
scobit ... , vce(cluster idcode)
xtcloglog ...
xtgee ... , family(binomial) link(logit) corr(exchangeable)
xtlogit ...
xtprobit ...

See [R] cloglog, [R] logit, [R] probit, [R] scobit, [XT] xtcloglog, [XT] xtgee, [XT] xtlogit, and [XT] xtpro-
bit for details.
 14 clogit — Conditional (fixed-effects) logistic regression

Stored results
clogit stores the following in e():
Scalars
e(N) number of observations
e(N drop) number of observations dropped because of all positive or all negative outcomes
e(N group drop) number of groups dropped because of all positive or all negative outcomes
e(k) number of parameters
e(k eq) number of equations in e(b)
e(k eq model) number of equations in overall model test
e(k dv) number of dependent variables
e(df m) model degrees of freedom
e(r2 p) pseudo-𝑅2
e(ll) log likelihood
e(ll 0) log likelihood, constant-only model
e(N clust) number of clusters
e(chi2) 𝜒2
e(p) 𝑝-value for model test
e(rank) rank of e(V)
e(ic) number of iterations
e(rc) return code
e(converged) 1 if converged, 0 otherwise
Macros
e(cmd) clogit
e(cmdline) command as typed
e(depvar) name of dependent variable
e(group) name of group() variable
e(multiple) multiple if multiple positive outcomes within group
e(wtype) weight type
e(wexp) weight expression
e(title) title in estimation output
e(clustvar) name of cluster variable
e(offset) linear offset variable
e(chi2type) Wald or LR; type of model 𝜒2 test
e(vce) vcetype specified in vce()
e(vcetype) title used to label Std. err.
e(opt) type of optimization
e(which) max or min; whether optimizer is to perform maximization or minimization
e(ml method) type of ml method
e(user) name of likelihood-evaluator program
e(technique) maximization technique
e(properties) b V
e(predict) program used to implement predict
e(marginsok) predictions allowed by margins
e(marginsnotok) predictions disallowed by margins
e(marginsdefault) default predict() specification for margins
e(asbalanced) factor variables fvset as asbalanced
e(asobserved) factor variables fvset as asobserved
Matrices
e(b) coefficient vector
e(Cns) constraints matrix
e(ilog) iteration log (up to 20 iterations)
e(gradient) gradient vector
e(V) variance–covariance matrix of the estimators
e(V modelbased) model-based variance
Functions
e(sample) marks estimation sample
 clogit — Conditional (fixed-effects) logistic regression 15

In addition to the above, the following is stored in r():

Matrices
r(table) matrix containing the coefficients with their standard errors, test statistics, 𝑝-values, and
confidence intervals

Note that results stored in r() are updated when the command is replayed and will be replaced when any
r-class command is run after the estimation command.

Methods and formulas

Breslow and Day (1980, 247–279), Collett (2003, 251–267), and Hosmer, Lemeshow, and Sturdivant
(2013, 243–268) provide a biostatistical point of view on conditional logistic regression. Hamerle and
Ronning (1995) give a succinct and lucid review of fixed-effects logit; Chamberlain (1980) is a standard
reference for this model. Greene (2018, chap. 18) provides a straightforward textbook description of
conditional logistic regression from an economist’s point of view, as well as a brief description of choice
models.
Let 𝑖 = 1, 2, . . . , 𝑛 denote the groups and let 𝑡 = 1, 2, . . . , 𝑇𝑖 denote the observations for the 𝑖th group.
Let 𝑦𝑖𝑡 be the dependent variable taking on values 0 or 1. Let y𝑖 = (𝑦𝑖1 , . . . , 𝑦𝑖𝑇𝑖 ) be the outcomes for
the 𝑖th group as a whole. Let x𝑖𝑡 be a row vector of covariates. Let
𝑇𝑖
𝑘1𝑖 = ∑ 𝑦𝑖𝑡
𝑡=1

be the observed number of ones for the dependent variable in the 𝑖th group. Biostatisticians would say
that there are 𝑘1𝑖 cases matched to 𝑘2𝑖 = 𝑇𝑖 − 𝑘1𝑖 controls in the 𝑖th group.
We consider the probability of a possible value of y𝑖 conditional on ∑𝑇𝑡=1
𝑖
𝑦𝑖𝑡 = 𝑘1𝑖 (Hamerle and
Ronning 1995, eq. 8.33; Hosmer, Lemeshow, and Sturdivant 2013, eq. 7.4),
𝑇
𝑖
exp(∑𝑡=1 𝑦𝑖𝑡 x𝑖𝑡 β)
Pr(y𝑖 ∣ ∑𝑇𝑡=1
𝑖
𝑦𝑖𝑡 = 𝑘1𝑖 ) =
∑d exp(∑𝑇𝑡=1
𝑖
𝑑𝑖𝑡 x𝑖𝑡 β)
𝑖 ∈𝑆𝑖

where 𝑑𝑖𝑡 is equal to 0 or 1 with ∑𝑇𝑡=1

𝑖
𝑑𝑖𝑡 = 𝑘1𝑖 , and 𝑆𝑖 is the set of all possible combinations of 𝑘1𝑖 ones
𝑇𝑖
and 𝑘2𝑖 zeros. Clearly, there are (𝑘 ) such combinations, but we need not count all of these combinations
1𝑖
to compute the denominator of the above equation. It can be computed recursively.
Denote the denominator by
𝑇𝑖
𝑓𝑖 (𝑇𝑖 , 𝑘1𝑖 ) = ∑ exp(∑ 𝑑𝑖𝑡 x𝑖𝑡 β)
d𝑖 ∈𝑆𝑖 𝑡=1

Consider, computationally, how 𝑓𝑖 changes as we go from a total of 1 observation in the group to 2

observations to 3, etc. Doing this, we derive the recursive formula
𝑓𝑖 (𝑇 , 𝑘) = 𝑓𝑖 (𝑇 − 1, 𝑘) + 𝑓𝑖 (𝑇 − 1, 𝑘 − 1) exp(x𝑖𝑇 β)
where we define 𝑓𝑖 (𝑇 , 𝑘) = 0 if 𝑇 < 𝑘 and 𝑓𝑖 (𝑇 , 0) = 1.
The conditional log-likelihood is
𝑛 𝑇𝑖
ln𝐿 = ∑ {∑ 𝑦𝑖𝑡 x𝑖𝑡 β − log𝑓𝑖 (𝑇𝑖 , 𝑘1𝑖 )}
𝑖=1 𝑡=1
 16 clogit — Conditional (fixed-effects) logistic regression

The derivatives of the conditional log-likelihood can also be computed recursively by taking derivatives
of the recursive formula for 𝑓𝑖 .
Computation time is roughly proportional to
𝑛
𝑝2 ∑ 𝑇𝑖 min(𝑘1𝑖 , 𝑘2𝑖 )
𝑖=1

where 𝑝 is the number of independent variables in the model. If min(𝑘1𝑖 , 𝑘2𝑖 ) is small, computation time
is not an issue. But if it is large—say, 100 or more—patience may be required.
If 𝑇𝑖 is large for all groups, the bias of the unconditional fixed-effects estimator is not a concern, and
we can confidently use logit with an indicator variable for each group (provided, of course, that the
number of groups is held within a Stata matrix; see [R] Limits).
This command supports the clustered version of the Huber/White/sandwich estimator of the variance
using vce(robust) and vce(cluster clustvar). See [P] robust, particularly Maximum likelihood es-
timators and Methods and formulas. Specifying vce(robust) is equivalent to specifying vce(cluster
groupvar), where groupvar is the variable for the matched groups.
clogit also supports estimation with survey data. For details on VCEs with survey data, see
[SVY] Variance estimation.

References
Andersen, E. B. 1970. Asymptotic properties of conditional maximum-likelihood estimators. Journal of the Royal Statis-
tical Society, B ser., 32: 283–301. https://doi.org/10.1111/j.2517-6161.1970.tb00842.x.
Breslow, N. E., and N. E. Day. 1980. The Analysis of Case–Control Studies. Vol. 1 of Statistical Methods in Cancer Research.
Lyon: IARC.
Chamberlain, G. 1980. Analysis of covariance with qualitative data. Review of Economic Studies 47: 225–238. https:
//doi.org/10.2307/2297110.
Collett, D. 2003. Modelling Binary Data. 2nd ed. London: Chapman and Hall/CRC. https://doi.org/10.1201/b16654.
Greene, W. H. 2018. Econometric Analysis. 8th ed. New York: Pearson.
Hamerle, A., and G. Ronning. 1995. “Panel analysis for qualitative variables”. In Handbook of Statistical Modeling for the
Social and Behavioral Sciences, edited by G. Arminger, C. C. Clogg, and M. E. Sobel, 401–451. New York: Plenum.
https://doi.org/10.1007/978-1-4899-1292-3_8.
Hole, A. R. 2007. Fitting mixed logit models by using maximum simulated likelihood. Stata Journal 7: 388–401.
Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013. Applied Logistic Regression. 3rd ed. Hoboken, NJ:
Wiley.
Long, J. S., and J. Freese. 2014. Regression Models for Categorical Dependent Variables Using Stata. 3rd ed. College
Station, TX: Stata Press.
McFadden, D. L. 1974. “Conditional logit analysis of qualitative choice behavior”. In Frontiers in Econometrics, edited
by P. Zarembka, 105–142. New York: Academic Press.
 clogit — Conditional (fixed-effects) logistic regression 17

Also see
[R] clogit postestimation — Postestimation tools for clogit
[R] logistic — Logistic regression, reporting odds ratios
[R] mlogit — Multinomial (polytomous) logistic regression
[R] ologit — Ordered logistic regression
[R] scobit — Skewed logistic regression
[BAYES] bayes: clogit — Bayesian conditional logistic regression
[CM] cmclogit — Conditional logit (McFadden’s) choice model
[CM] nlogit — Nested logit regression
[MI] Estimation — Estimation commands for use with mi estimate
[SVY] svy estimation — Estimation commands for survey data
[XT] xtgee — GEE population-averaged panel-data models
[XT] xtlogit — Fixed-effects, random-effects, and population-averaged logit models
[XT] xtmlogit — Fixed-effects and random-effects multinomial logit models
[U] 20 Estimation and postestimation commands

Stata, Stata Press, and Mata are registered trademarks of StataCorp LLC. Stata and Stata
®
Press are registered trademarks with the World Intellectual Property Organization of the
United Nations. StataNow and NetCourseNow are trademarks of StataCorp LLC. Other
brand and product names are registered trademarks or trademarks of their respective com-
panies. Copyright © 1985–2023 StataCorp LLC, College Station, TX, USA. All rights
reserved.
For suggested citations, see the FAQ on citing Stata documentation.