Title stata.

com
lrtest — Likelihood-ratio test after estimation

Description Quick start Menu Syntax

Options Remarks and examples Stored results Methods and formulas
References Also see

Description
lrtest performs a likelihood-ratio test of the null hypothesis that the parameter vector of a
statistical model satisfies some smooth constraint. To conduct the test, both the unrestricted and the
restricted models must be fit using the maximum likelihood method (or some equivalent method),
and the results of at least one must be stored using estimates store.
lrtest also supports composite models. In a composite model, we assume that the log likelihood
and dimension (number of free parameters) of the full model are obtained as the sum of the
log-likelihood values and dimensions of the constituting models.

Quick start
Likelihood-ratio test that the coefficients for x2 and x3 are equal to 0
logit y x1 x2 x3
estimates store full
logit y x1 if e(sample)
estimates store restricted
lrtest full restricted
Display additional information, including AIC and BIC
lrtest full restricted, stats
Likelihood-ratio test that the coefficients for x1 and x3 are equal
constraint 1 x1=x3
logit y x1 x2 x3, constraints(1)
estimates store constrained
lrtest full constrained
Compare stored estimates full with the last model run
lrtest full

Menu
Statistics > Postestimation

1
 2 lrtest — Likelihood-ratio test after estimation

Syntax
  
lrtest modelspec1 modelspec2 , options

modelspec1 and modelspec2 specify the restricted and unrestricted model in any order. modelspec# is
name | . | (namelist)
name is the name under which estimation results were stored using estimates store (see
[R] estimates store), and “.” refers to the last estimation results, whether or not these were
already stored. If modelspec2 is not specified, the last estimation result is used; this is equivalent
to specifying modelspec2 as “.”.
If namelist is specified for a composite model, modelspec1 and modelspec2 cannot have names in
common; for example, lrtest (A B C) (C D E) is not allowed because both model specifications
include C.

options Description
stats display statistical information about the two models
dir display descriptive information about the two models
df(#) override the automatic degrees-of-freedom calculation; seldom used
force force testing even when apparently invalid
collect is allowed; see [U] 11.1.10 Prefix commands.

Options
stats displays statistical information about the unrestricted and restricted models, including the
information indices of Akaike and Schwarz.
dir displays descriptive information about the unrestricted and restricted models; see estimates
dir in [R] estimates store.
df(#) is seldom specified; it overrides the automatic degrees-of-freedom calculation.
force forces the likelihood-ratio test calculations to take place in situations where lrtest would
normally refuse to do so and issue an error. Such situations arise when one or more assumptions
of the test are violated, for example, if the models were fit with vce(robust), vce(cluster
clustvar), or pweights; when the dependent variables in the two models differ; when the null log
likelihoods differ; when the samples differ; or when the estimation commands differ. If you use
the force option, there is no guarantee as to the validity or interpretability of the resulting test.

Remarks and examples stata.com

The standard way to use lrtest is to do the following:
1. Fit either the restricted model or the unrestricted model by using one of Stata’s estimation commands
and then store the results using estimates store name.
2. Fit the alternative model (the unrestricted or restricted model) and then type ‘lrtest name .’.
lrtest determines for itself which of the two models is the restricted model by comparing the
degrees of freedom.
Often, you may want to store the alternative model with estimates store name2 , for instance,
if you plan additional tests against models yet to be fit. The likelihood-ratio test is then obtained as
lrtest name name2 .
 lrtest — Likelihood-ratio test after estimation 3

Remarks are presented under the following headings:

Nested models
Composite models

Nested models
lrtest may be used with any estimation command that reports a log likelihood, including heckman,
logit, poisson, stcox, and streg. You must check that one of the model specifications implies a
statistical model that is nested within the model implied by the other specification. Usually, this means
that both models are fit with the same estimation command (for example, both are fit by logit, with
the same dependent variables) and that the set of covariates of one model is a subset of the covariates
of the other model. Second, lrtest is valid only for models that are fit by maximum likelihood or
by some equivalent method, so it does not apply to models that were fit with probability weights or
clusters. Specifying the vce(robust) option similarly would indicate that you are worried about the
valid specification of the model, so you would not use lrtest. Third, lrtest assumes that under
the null hypothesis, the test statistic is (approximately) distributed as χ2 . This assumption is not true
for likelihood-ratio tests of “boundary conditions”, such as tests for the presence of overdispersion or
random effects (Gutierrez, Carter, and Drukker 2001).

Example 1
We have data on infants born with low birthweights along with the characteristics of the mother
(Hosmer, Lemeshow, and Sturdivant 2013; see also [R] logistic). We fit the following model:
. use https://www.stata-press.com/data/r18/lbw
(Hosmer & Lemeshow data)
. logistic low age lwt i.race smoke ptl ht ui
Logistic regression Number of obs = 189
LR chi2(8) = 33.22
Prob > chi2 = 0.0001
Log likelihood = -100.724 Pseudo R2 = 0.1416

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

age .9732636 .0354759 -0.74 0.457 .9061578 1.045339

lwt .9849634 .0068217 -2.19 0.029 .9716834 .9984249

race
Black 3.534767 1.860737 2.40 0.016 1.259736 9.918406
Other 2.368079 1.039949 1.96 0.050 1.001356 5.600207

smoke 2.517698 1.00916 2.30 0.021 1.147676 5.523162

ptl 1.719161 .5952579 1.56 0.118 .8721455 3.388787
ht 6.249602 4.322408 2.65 0.008 1.611152 24.24199
ui 2.1351 .9808153 1.65 0.099 .8677528 5.2534
_cons 1.586014 1.910496 0.38 0.702 .1496092 16.8134

Note: _cons estimates baseline odds.

We now wish to test the constraint that the coefficients on age, lwt, ptl, and ht are all zero or,
equivalently here, that the odds ratios are all 1. One solution is to type
4 lrtest — Likelihood-ratio test after estimation

. test age lwt ptl ht

( 1) [low]age = 0
( 2) [low]lwt = 0
( 3) [low]ptl = 0
( 4) [low]ht = 0
chi2( 4) = 12.38
Prob > chi2 = 0.0147

This test is based on the inverse of the information matrix and is therefore based on a quadratic
approximation to the likelihood function; see [R] test. A more precise test would be to refit the model,
applying the proposed constraints, and then calculate the likelihood-ratio test.
We first save the current model:
. estimates store full

We then fit the constrained model, which here is the model omitting age, lwt, ptl, and ht:
. logistic low i.race smoke ui
Logistic regression Number of obs = 189
LR chi2(4) = 18.80
Prob > chi2 = 0.0009
Log likelihood = -107.93404 Pseudo R2 = 0.0801

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

race
Black 3.052746 1.498087 2.27 0.023 1.166747 7.987382
Other 2.922593 1.189229 2.64 0.008 1.316457 6.488285

smoke 2.945742 1.101838 2.89 0.004 1.415167 6.131715

ui 2.419131 1.047359 2.04 0.041 1.035459 5.651788
_cons .1402209 .0512295 -5.38 0.000 .0685216 .2869447

Note: _cons estimates baseline odds.

That done, lrtest compares this model with the model we previously stored:
. lrtest full .
Likelihood-ratio test
Assumption: . nested within full
LR chi2(4) = 14.42
Prob > chi2 = 0.0061

Let’s compare results. test reported that age, lwt, ptl, and ht were jointly significant at the 1.5%
level; lrtest reports that they are significant at the 0.6% level. Given the quadratic approximation
made by test, we could argue that lrtest’s results are more accurate.
lrtest explicates the assumption that, from a comparison of the degrees of freedom, it has assessed
that the last fit model (.) is nested within the model stored as full. In other words, full is the
unconstrained model and . is the constrained model.
The names in “(Assumption: . nested in full)” are actually links. Click on a name, and the
results for that model are replayed.

Aside: The nestreg command provides a simple syntax for performing likelihood-ratio tests for
nested model specifications; see [R] nestreg. In the previous example, we fit a full logistic model,
used estimates store to store the full model, fit a constrained logistic model, and used lrtest
to report a likelihood-ratio test between two models. To do this with one call to nestreg, use the
lrtable option.
 lrtest — Likelihood-ratio test after estimation 5

Technical note
lrtest determines the degrees of freedom of a model as the rank of the (co)variance matrix
e(V). There are two issues here. First, the numerical determination of the rank of a matrix is a subtle
problem that can, for instance, be affected by the scaling of the variables in the model. The rank of a
matrix depends on the number of (independent) linear combinations of coefficients that sum exactly
to zero. In the world of numerical mathematics, it is hard to tell whether a very small number is
really nonzero or is a real zero that happens to be slightly off because of roundoff error from the finite
precision with which computers make floating-point calculations. Whether a small number is being
classified as one or the other, typically on the basis of a threshold, affects the determined degrees of
freedom. Although Stata generally makes sensible choices, it is bound to make mistakes occasionally.
The moral of this story is to make sure that the calculated degrees of freedom is as you expect before
interpreting the results.

Technical note
A second issue involves regress and related commands such as anova. Mainly for historical
reasons, regress does not treat the residual variance, σ 2 , the same way that it treats the regression
coefficients. Type estat vce after regress, and you will see the regression coefficients, not σ b2 .
Most estimation commands for models with ancillary parameters (for example, streg and heckman)
treat all parameters as equals. There is nothing technically wrong with regress here; we are usually
focused on the regression coefficients, and their estimators are uncorrelated with σb2 . But, formally,
2
σ adds a degree of freedom to the model, which does not matter if you are comparing two regression
models by a likelihood-ratio test. This test depends on the difference in the degrees of freedom,
and hence being “off by 1” in each does not matter. But, if you are comparing a regression model
with a larger model—for example, a heteroskedastic regression model fit by arch—the automatic
determination of the degrees of freedom is incorrect, and you must specify the df(#) option.

Example 2
Returning to the low-birthweight data in example 1, we now wish to test that the coefficient on
2.race (black) is equal to that on 3.race (other). The base model is still stored under the name
full, so we need only fit the constrained model and perform the test. With z as the index of the
logit model, the base model is

z = β0 + β1 age + β2 lwt + β3 2.race + β4 3.race + · · ·

If β3 = β4 , this can be written as

z = β0 + β1 age + β2 lwt + β3 (2.race + 3.race) + · · ·

6 lrtest — Likelihood-ratio test after estimation

We can fit the constrained model as follows:

. constraint 1 2.race = 3.race
. logistic low age lwt i.race smoke ptl ht ui, constraints(1)
Logistic regression Number of obs = 189
Wald chi2(7) = 25.17
Log likelihood = -100.9997 Prob > chi2 = 0.0007
( 1) [low]2.race - [low]3.race = 0

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

age .9716799 .0352638 -0.79 0.429 .9049649 1.043313

lwt .9864971 .0064627 -2.08 0.038 .9739114 .9992453

race
Black 2.728186 1.080207 2.53 0.011 1.255586 5.927907
Other 2.728186 1.080207 2.53 0.011 1.255586 5.927907

smoke 2.664498 1.052379 2.48 0.013 1.228633 5.778414

ptl 1.709129 .5924776 1.55 0.122 .8663666 3.371691
ht 6.116391 4.215585 2.63 0.009 1.58425 23.61385
ui 2.09936 .9699702 1.61 0.108 .8487997 5.192407
_cons 1.309371 1.527398 0.23 0.817 .1330839 12.8825

Note: _cons estimates baseline odds.

Comparing this model with our original model, we obtain

. lrtest full .
Likelihood-ratio test
Assumption: . nested within full
LR chi2(1) = 0.55
Prob > chi2 = 0.4577

By comparison, typing test 2.race=3.race after fitting our base model results in a significance
level of 0.4572. Alternatively, we can first store the restricted model, here using the name equal.
Next, lrtest is invoked specifying the names of the restricted and unrestricted models (we do not
care about the order). This time, we also add the option stats requesting a table of model statistics,
including the model selection indices AIC and BIC.
. estimates store equal
. lrtest equal full, stats
Likelihood-ratio test
Assumption: equal nested within full
LR chi2(1) = 0.55
Prob > chi2 = 0.4577
Akaike’s information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

equal 189 . -100.9997 8 217.9994 243.9334

full 189 -117.336 -100.724 9 219.448 248.6237

Note: BIC uses N = number of observations. See [R] IC note.

 lrtest — Likelihood-ratio test after estimation 7

Composite models
lrtest supports composite models; that is, models that can be fit by fitting a series of simpler
models or by fitting models on subsets of the data. Theoretically, a composite model is one in which
the likelihood function, L(θ), of the parameter vector, θ, can be written as the product

L(θ) = L1 (θ1 ) × L2 (θ2 ) × · · · × Lk (θk )

of likelihood terms with θ = (θ1 , . . . , θk ) a partitioning of the full parameter vector. In such a
case, the full-model likelihood L(θ) is maximized by maximizing the likelihood terms Lj (θj ) in
turn. Obviously, logL(θ)b = Pk logLj (θbj ). The degrees of freedom for the composite model is
j=1
obtained as the sum of the degrees of freedom of the constituting models.

Example 3
As an example of the application of composite models, we consider a test of the hypothesis that the
coefficients of a statistical model do not differ between different portions (“regimes”) of the covariate
space. Economists call a test for such a hypothesis a Chow test.
We continue the analysis of the data on children of low birthweight by using logistic regression
modeling and study whether the regression coefficients are the same among the three races: white,
black, and other. A likelihood-ratio Chow test can be obtained by fitting the logistic regression model
for each of the races and then comparing the combined results with those of the model previously
stored as full. Because the full model included dummies for the three races, this version of the
Chow test allows the intercept of the logistic regression model to vary between the regimes (races).
. logistic low age lwt smoke ptl ht ui if 1.race, nolog
Logistic regression Number of obs = 96
LR chi2(6) = 13.86
Prob > chi2 = 0.0312
Log likelihood = -45.927061 Pseudo R2 = 0.1311

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

age .9869674 .0527757 -0.25 0.806 .8887649 1.096021

lwt .9900874 .0106101 -0.93 0.353 .9695089 1.011103
smoke 4.208697 2.680133 2.26 0.024 1.20808 14.66222
ptl 1.592145 .7474264 0.99 0.322 .6344379 3.995544
ht 2.900166 3.193537 0.97 0.334 .3350554 25.1032
ui 1.229523 .9474768 0.27 0.789 .2715165 5.567715
_cons .4891008 .993785 -0.35 0.725 .0091175 26.23746

Note: _cons estimates baseline odds.

. estimates store white
8 lrtest — Likelihood-ratio test after estimation

. logistic low age lwt smoke ptl ht ui if 2.race, nolog

Logistic regression Number of obs = 26
LR chi2(6) = 10.12
Prob > chi2 = 0.1198
Log likelihood = -12.654157 Pseudo R2 = 0.2856

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

age .8735313 .1377846 -0.86 0.391 .6412332 1.189983

lwt .9747736 .016689 -1.49 0.136 .9426065 1.008038
smoke 16.50373 24.37044 1.90 0.058 .9133647 298.2083
ptl 4.866916 9.33151 0.83 0.409 .1135573 208.5895
ht 85.05605 214.6382 1.76 0.078 .6049308 11959.27
ui 67.61338 133.3313 2.14 0.033 1.417399 3225.322
_cons 48.7249 169.9216 1.11 0.265 .0523961 45310.94

Note: _cons estimates baseline odds.

. estimates store black
. logistic low age lwt smoke ptl ht ui if 3.race, nolog
Logistic regression Number of obs = 67
LR chi2(6) = 14.06
Prob > chi2 = 0.0289
Log likelihood = -37.228444 Pseudo R2 = 0.1589

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

age .9263905 .0665386 -1.06 0.287 .8047407 1.06643

lwt .9724499 .015762 -1.72 0.085 .9420424 1.003839
smoke .7979034 .6340585 -0.28 0.776 .1680885 3.787586
ptl 2.845675 1.777944 1.67 0.094 .8363053 9.682908
ht 7.767503 10.00537 1.59 0.112 .6220764 96.98826
ui 2.925006 2.046473 1.53 0.125 .7423107 11.52571
_cons 49.09444 113.9165 1.68 0.093 .5199275 4635.769

Note: _cons estimates baseline odds.

. estimates store other

We are now ready to perform the likelihood-ratio Chow test:

. lrtest (full) (white black other), stats
Likelihood-ratio test
Assumption: full nested within (white, black, other)
LR chi2(12) = 9.83
Prob > chi2 = 0.6310
Akaike’s information criterion and Bayesian information criterion

Model N ll(null) ll(model) df AIC BIC

full 189 -117.336 -100.724 9 219.448 248.6237

white 96 -52.85752 -45.92706 7 105.8541 123.8046
black 26 -17.71291 -12.65416 7 39.30831 48.11499
other 67 -44.26039 -37.22844 7 88.45689 103.8897

Note: BIC uses N = number of observations. See [R] IC note.

We cannot reject the hypothesis that the logistic regression model applies to each of the races at any
reasonable significance level. By specifying the stats option, we can verify the degrees of freedom
of the test: 12 = 7 + 7 + 7 − 9. We can obtain the same test by fitting an expanded model with
interactions between all covariates and race.
 lrtest — Likelihood-ratio test after estimation 9

. logistic low race##c.(age lwt smoke ptl ht ui)

Logistic regression Number of obs = 189
LR chi2(20) = 43.05
Prob > chi2 = 0.0020
Log likelihood = -95.809661 Pseudo R2 = 0.1835

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

race
Black 99.62137 402.0829 1.14 0.254 .0365434 271578.9
Other 100.3769 309.586 1.49 0.135 .2378638 42358.38

age .9869674 .0527757 -0.25 0.806 .8887649 1.096021

lwt .9900874 .0106101 -0.93 0.353 .9695089 1.011103
smoke 4.208697 2.680133 2.26 0.024 1.20808 14.66222
ptl 1.592145 .7474264 0.99 0.322 .6344379 3.995544
ht 2.900166 3.193537 0.97 0.334 .3350554 25.1032
ui 1.229523 .9474768 0.27 0.789 .2715165 5.567715

race#c.age
Black .885066 .1474079 -0.73 0.464 .638569 1.226714
Other .9386232 .0840486 -0.71 0.479 .7875366 1.118695

race#c.lwt
Black .9845329 .0198857 -0.77 0.440 .9463191 1.02429
Other .9821859 .0190847 -0.93 0.355 .9454839 1.020313

race#c.smoke
Black 3.921338 6.305992 0.85 0.395 .167725 91.67917
Other .1895844 .1930601 -1.63 0.102 .025763 1.395113

race#c.ptl
Black 3.05683 6.034089 0.57 0.571 .0638301 146.3918
Other 1.787322 1.396789 0.74 0.457 .3863582 8.268285

race#c.ht
Black 29.328 80.7482 1.23 0.220 .1329492 6469.623
Other 2.678295 4.538712 0.58 0.561 .0966916 74.18702

race#c.ui
Black 54.99155 116.4274 1.89 0.058 .8672471 3486.977
Other 2.378976 2.476124 0.83 0.405 .309335 18.29579

_cons .4891008 .993785 -0.35 0.725 .0091175 26.23746

Note: _cons estimates baseline odds.

. lrtest full .
Likelihood-ratio test
Assumption: full nested within .
LR chi2(12) = 9.83
Prob > chi2 = 0.6310

Applying lrtest for the full model against the model with all interactions yields the same test
statistic and p-value as for the full model against the composite model for the three regimes. Here
the specification of the model with interactions was convenient, and logistic had no problem
computing the estimates for the expanded model. In models with more complicated likelihoods, such
as Heckman’s selection model (see [R] heckman) or complicated survival-time models (see [ST] streg),
fitting the models with all interactions may be numerically demanding and may be much more time
consuming than fitting a series of models separately for each regime.
 10 lrtest — Likelihood-ratio test after estimation

Given the model with all interactions, we could also test the hypothesis of no differences among
the regions (races) by a Wald version of the Chow test by using the testparm command; see [R] test.
. testparm race#c.(age lwt smoke ptl ht ui)
( 1) [low]2.race#c.age = 0
( 2) [low]3.race#c.age = 0
( 3) [low]2.race#c.lwt = 0
( 4) [low]3.race#c.lwt = 0
( 5) [low]2.race#c.smoke = 0
( 6) [low]3.race#c.smoke = 0
( 7) [low]2.race#c.ptl = 0
( 8) [low]3.race#c.ptl = 0
( 9) [low]2.race#c.ht = 0
(10) [low]3.race#c.ht = 0
(11) [low]2.race#c.ui = 0
(12) [low]3.race#c.ui = 0
chi2( 12) = 8.24
Prob > chi2 = 0.7663

We conclude that, here, the Wald version of the Chow test is similar to the likelihood-ratio version
of the Chow test.

Stored results
lrtest stores the following in r():
Scalars
r(p) p-value for likelihood-ratio test
r(df) degrees of freedom
r(chi2) LR test statistic

Programmers wishing their estimation commands to be compatible with lrtest should note that
lrtest requires that the following results be returned:
e(cmd) name of estimation command
e(ll) log likelihood
e(V) variance–covariance matrix of the estimators
e(N) number of observations

lrtest also verifies that e(N), e(ll 0), and e(depvar) are consistent between two noncomposite
models.

Methods and formulas

Let L0 and L1 be the log-likelihood values associated with the full and constrained models,
respectively. The test statistic of the likelihood-ratio test is LR = −2(L1 − L0 ). If the constrained
model is true, LR is approximately χ2 distributed with d0 − d1 degrees of freedom, where d0 and
d1 are the model degrees of freedom associated with the full and constrained models, respectively
(Greene 2018, 554 – 555).
lrtest determines the degrees of freedom of a model as the rank of e(V), computed as the
number of nonzero diagonal elements of invsym(e(V)).
 lrtest — Likelihood-ratio test after estimation 11

References
Greene, W. H. 2018. Econometric Analysis. 8th ed. New York: Pearson.
Gutierrez, R. G., S. L. Carter, and D. M. Drukker. 2001. sg160: On boundary-value likelihood-ratio tests. Stata
Technical Bulletin 60: 15–18. Reprinted in Stata Technical Bulletin Reprints, vol. 10, pp. 269–273. College Station,
TX: Stata Press.
Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013. Applied Logistic Regression. 3rd ed. Hoboken,
NJ: Wiley.
Raciborski, R. 2015. Spotlight on irt. The Stata Blog: Not Elsewhere Classified.
http://blog.stata.com/2015/07/31/spotlight-on-irt/.
Tauchmann, H. 2023. lgrgtest: Lagrange multiplier test after constrained maximum-likelihood estimation. Stata Journal
23: 386–401.

Also see
[R] test — Test linear hypotheses after estimation
[R] testnl — Test nonlinear hypotheses after estimation
[R] nestreg — Nested model statistics

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. Other brand and product names are registered trademarks or
trademarks of their respective companies. Copyright c 1985–2023 StataCorp LLC,
College Station, TX, USA. All rights reserved.