BIOST 536 Lecture 12 1 Lecture 12 – Introduction to Matching Matching/Stratification G roup subjectsinto subsetson the basisofassum ed confounders(e.g. age, gender, clinic, etc) M atching variablesshould notbe ofdirectscientific interest M atching D one in advance, often in a specified ratio:1 case to 1 control;or 1 case to m controls M ay be m atched on underlying relationship thatcannotbe quantified (tw ins, siblings, neighbors) Stratification U sually done after data collection on the basisofrecorded covariates Little controlover the num ber ofcasesand controlsin each stratum n j casesm atched to m j controlsin stratum j Strata w ith either no casesor no controlsgetelim inated from the analysis M atched analyzed asstratified Suppose w e m atch a 40-44 year old case w ith four 40-44 year old controls, then m atch another 40-44 year old case w ith four different40-44 year old controls C om bine alltogether and analyze astw o casesm atched to eightcontrols(“frequency m atching”)
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
BIOST 536 Lecture 12 1
Lecture 12 – Introduction to MatchingMatching/Stratification
Group subjects into subsets on the basis of assumed confounders (e.g. age, gender, clinic, etc)
Matching variables should not be of direct scientific interest
Matching
Done in advance, often in a specified ratio: 1 case to 1 control ; or 1 case to m controls
May be matched on underlying relationship that cannot be quantified (twins, siblings, neighbors)
Stratification
Usually done after data collection on the basis of recorded covariates
Little control over the number of cases and controls in each stratum
n j cases matched to m j controls in stratum j
Strata with either no cases or no controls get eliminated from the analysis
Matched analyzed as stratified
Suppose we match a 40-44 year old case with four 40-44 year old controls, then match another 40-44
year old case with four different 40-44 year old controls
Combine all together and analyze as two cases matched to eight controls (“frequency matching”)
BIOST 536 Lecture 12 2
Conditional logistic regressionStratified or matched data are often analyzed with conditional logistic regression
Matched data can be analyzed with unconditional logistic regression in some cases
Unmatched data can be analyzed with conditional logistic regression in some cases (post-hoc
stratification)
Conditional logistic regression
For finely matched or stratified data with many nuisance parameters
Simplest case
1 to 1 matching for J pairs of observations
In the unconditional model ( j = 1, 2, … , J )
ikkij XX ...)(plogit 11i
Each matched pair has its own j
Number of parameters is J + K
As the number of pairs goes to , then the number of parameters does as well
Asymptotic statistical properties fail if this happens
BIOST 536 Lecture 12 3
Conditional logistic regression
Solutions
1. Do not try to estimate the 's directly, but rather estimate some hidden distribution generating
the 's, e.g. assume they are from a standard normal distribution N( 0, 2 ) and estimate 2
instead of the 's
Approach is called a "random effects model"
2. Get rid of the 's by conditioning them out of the likelihood and estimating the remaining parameters
(conditional logistic regression)
Consider the likelihood contribution of the j th matched pair (one case-one control)
controlcase
case
controlcase
case
j
j
controljcasej
casej
XX
X
XX
X
XX
X
ee
e
ee
e
e
e
ee
e
which depends only on the covariate and , but not j
BIOST 536 Lecture 12 4
Conditional logistic regressionSituation where a conditional likelihood is necessary or at least useful
Let the first stratum be the referent group
2 stratumin not 0
2 stratumin 1Z2
3 stratumin not 0
3 stratumin 1Z3 …
J stratumin not 0
J stratumin 1ZJ
kkJJ XXZZZ ......(p)logit 1133221
has J matched sets (strata)
Suppose each set is 1 case : m controls, then there are n = J (m+1) observations
If J is small relative to n, then we can use standard logistic regression and estimate the 's
If J as n then we do need to use the conditional likelihood
If the unconditional likelihood is used, the estimates of the 's are biased anti-conservatively
(too high for > 0)
For example, in 1-1 matching with a single exposure and many matched sets
nalunconditio estimates 2 not and nalunconditio estimates 2 not
Breslow& Day (page 252) shows the effects of this bias under different sample sizes and true OR’s
BIOST 536 Lecture 12 5
Conditional logistic regressionFew strata relative to sample size:
Age group categorized into six groups (30-39), (40-49), (50-59), (60-69),
(70-79), 80+
Only six strata and the number of strata does not increase as n
Can use unconditional logistic regression :
kk XX
AGEGRPAGEGRP
...
...(p)logit
11
66221
Can also use conditional logistic regression and condition out the six
parameters
Often compare both methods to make sure they agree
My personal rule: I tend to use conditional logistic regression and would
always use it when the number of nuisance parameters ('s) is
greater than 10
BIOST 536 Lecture 12 6
Example Oxford data with a parameter for each year of birth 1944-1964
Cases and controls were frequency matched on year of birth Ungrouped data: n = 11,852 observations
. tabulate xray y | y xray | 0 1 | Total -----------+----------------------+---------- 0 | 5,324 4,994 | 10,318 1 | 602 932 | 1,534 -----------+----------------------+---------- Total | 5,926 5,926 | 11,852
Example Can use matched case-control command (mcc)
Can get the OR easily and get confidence intervals and exact p-values based on the exact binomial distribution with null hypothesis p=0.50 and n = number discordant on exposure status
Easier to just use conditional logistic regression
ExamplePrevious gall bladder disease; univariate odds ratio . clogit case gall, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 LR chi2(1) = 3.68 Prob > chi2 = 0.0550 Log likelihood = -41.826784 Pseudo R2 = 0.0422 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 2.6 1.368211 1.82 0.069 .9269183 7.292984 ------------------------------------------------------------------------------
Hypertension; univariate odds ratio . clogit case hyperten, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 chi2(1) = 0.81 Prob > chi2 = 0.3681 Log Likelihood = -43.26328 Pseudo R2 = 0.0093 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ---------+-------------------------------------------------------------------- hyperten | 1.384615 .5039669 0.894 0.371 .678441 2.825831 ------------------------------------------------------------------------------
Obesity; univariate odds ratio . clogit case obes, group(set) or Note: 17 groups (17 obs) dropped due to all positive or negative outcomes. Conditional (fixed-effects) logistic regression Number of obs = 90 chi2(1) = 0.25 Prob > chi2 = 0.6166 Log Likelihood = -31.066296 Pseudo R2 = 0.0040 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ---------+-------------------------------------------------------------------- obes | 1.285714 .6479391 0.499 0.618 .4788276 3.45231 ------------------------------------------------------------------------------
BIOST 536 Lecture 12 21
ExampleEstrogen use; univariate odds ratio . clogit case estrogen, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 chi2(1) = 24.45 Prob > chi2 = 0.0000 Log Likelihood = -31.443696 Pseudo R2 = 0.2799 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ---------+-------------------------------------------------------------------- estrogen | 9.666667 5.862608 3.741 0.000 2.944712 31.73296 ------------------------------------------------------------------------------
Gall bladder and estrogen use; multivariate odds ratios . clogit case gall estrogen, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 chi2(2) = 25.79 Prob > chi2 = 0.0000 Log Likelihood = -30.77245 Pseudo R2 = 0.2953 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ---------+-------------------------------------------------------------------- gall | 2.003172 1.233219 1.128 0.259 .5993636 6.694932 estrogen | 9.107077 5.552653 3.623 0.000 2.756735 30.08589 ------------------------------------------------------------------------------
Estrogen use odds ratio reduced from 9.67 to 9.11
Gall bladder odds ratio reduced from 2.60 to 2.00
BIOST 536 Lecture 12 22
ExampleWhat can we do in conditional logistic regression ?
Model building proceeds in a fashion similar to unconditional logistic regression
Can do Wald test and LR test
Fit gall bladder disease . clogit case gall, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 LR chi2(1) = 3.68 Prob > chi2 = 0.0550 Log likelihood = -41.826784 Pseudo R2 = 0.0422 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 2.6 1.368211 1.82 0.069 .9269183 7.292984 ------------------------------------------------------------------------------
Gall bladder disease is marginally important
Now add estrogen use to the model . est store A . clogit case gall estrogen, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 LR chi2(2) = 25.79 Prob > chi2 = 0.0000 Log likelihood = -30.77245 Pseudo R2 = 0.2953 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 2.003172 1.233221 1.13 0.259 .5993628 6.694941 estrogen | 9.107077 5.552675 3.62 0.000 2.756722 30.08604 ------------------------------------------------------------------------------
BIOST 536 Lecture 12 23
Example. est store B . lrtest A B Likelihood-ratio test LR chi2(1) = 22.11 (Assumption: A nested in B) Prob > chi2 = 0.0000
Consider the interaction of estrogen and gall bladder disease . xi: clogit case i.gall*estrogen, group(set) or i.gall _Igall_0-1 (naturally coded; _Igall_0 omitted) i.gall*estrogen _IgalXestro_# (coded as above) Conditional (fixed-effects) logistic regression Number of obs = 126 LR chi2(3) = 28.35 Prob > chi2 = 0.0000 Log likelihood = -29.492281 Pseudo R2 = 0.3246 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- _Igall_1 | 9.898637 12.10023 1.88 0.061 .9016854 108.6665 estrogen | 14.45528 10.88972 3.55 0.000 3.302049 63.28045 _IgalXestr~1 | .1174832 .1609567 -1.56 0.118 .008013 1.722488 ------------------------------------------------------------------------------ . est store C . lrtest B C Likelihood-ratio test LR chi2(1) = 2.56 (Assumption: B nested in C) Prob > chi2 = 0.1096 . lincom _Igall_1 + estrogen + _IgalXestro_1 , or ( 1) [case]_Igall_1 + [case]estrogen + [case]_IgalXestro_1 = 0 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- (1) | 16.81038 15.2039 3.12 0.002 2.855758 98.95412 ------------------------------------------------------------------------------
Remove the interaction and consider whether gall bladder is needed
BIOST 536 Lecture 12 24
Example. clogit case estrogen, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 126 LR chi2(1) = 24.45 Prob > chi2 = 0.0000 Log likelihood = -31.443696 Pseudo R2 = 0.2799 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- estrogen | 9.666667 5.862625 3.74 0.000 2.944702 31.73307 ------------------------------------------------------------------------------ . est store A . lrtest A B Likelihood-ratio test LR chi2(1) = 1.34 (Assumption: A nested in B) Prob > chi2 = 0.2466
Estrogen alone is a strong predictor and gall bladder does not add significantly
Compute the conditional probabilities of being a case given 1-1 matching and the covariates . clogitp condprob, group(set) . list set case condprob estrogen set case condprob estrogen 1. 1 1 .90625 1 2. 1 0 .09375 0 etc 123. 62 0 .09375 0 124. 62 1 .90625 1 125. 63 0 .5 1 126. 63 1 .5 1
The sum of the conditional probabilities for each set is always 1 (for 1-m matching)
Sets where the covariate is the same for the case and control (condprob=0.5)
do not contribute to the conditional likelihood
BIOST 536 Lecture 12 25
1-m matching 1 case matched to m controls
For simplicity assume, m = 3
For the j th matched set, let
X j0 = covariate for the case
X j1 = covariate for the first control
X j2 = covariate for the second control
X j3 = covariate for the third control
and Pij
X
X
ji
ji
e
e
j
j
1 where i = 0, 1, 2, 3
Then Pij is the absolute probability of becoming a case
However, what we know is that one and only one person became a case so what we want is the
conditional probability of being a case given exactly one person is a case
p00 1 2 3
0 1 2 3 1 0 2 3 2 0 1 3 3 0 1 2
P (1- P (1- P (1- P
P (1- P (1- P (1- P P (1- P (1- P (1- P P (1- P (1- P (1- P P (1- P (1- P (1- P
) ) )
) ) ) ) ) ) ) ) ) ) ) )
e
e e e e
j
j j j j
X
X X X X
j0
j0 j1 j2 j3
e
e e e e
X
X X X X
j0
j0 j1 j2 j3
BIOST 536 Lecture 12 26
1-m matching Since j is not estimated, we can not get the unconditional probability Pij
We can estimate the conditional probabilities, i.e. p0, p1, p2, p3 , instead
These actually are multinomial probabilities
j3j2j1j0
j1
XXXX
X
1
eeee
ep
j3j2j1j0
j2
XXXX
X
2
eeee
ep
j3j2j1j0
j3
XXXX
X
3
eeee
ep
are the
conditional probabilities that each control was the one and only case in the set
Note: 13210 pppp by the way they are defined
If the covariate is the same for all members of the set we would have the likelihood contribution
4
1xxxx
x
XXXX
X
0 j3j2j1j0
j0
eeee
e
eeee
ep
so would not tell us anything about
If there are several covariates then the conditional probability for 1 to m matching is
e
e e
k
k k
1 2
1 2 1 2
X X X
X X X X X X
controls i=1 to m
j01 j02 j0k
j01 j02 j0k ji1 ji2 jik
...
... ...
BIOST 536 Lecture 12 27
1-m matching
Note there is no 0 in this model ( no const )
Interactions can be included in the same framework
If all members in the set have exactly the same covariate values for all covariates, then the set
drops out
If all members in the set have exactly the same covariate values for some of the covariates , then
that set does not help estimation for those covariates
If the case has a missing value for one covariate, the case drops out and takes the entire set with it
BIOST 536 Lecture 12 28
ExampleEndometrial cancer with 4 controls per case
n = 315 63 sets of five observations matched on age and neighborhood
First three sets have all the same value for gall bladder - no contribution from these sets when
estimating for gall bladder
First set has three missing values for obesity (code 9) - would end up as a 1-1 match since the case
value is not missing
BIOST 536 Lecture 12 29
ExampleChange missing value codes to Stata missing values . replace obes=. if obes==9 (51 real changes made, 51 to missing) . replace estdose=. if estdose==9 (8 real changes made, 8 to missing) . replace premarin=. if premarin==99 (17 real changes made, 17 to missing)
Fit gall bladder . clogit case gall , group(set) or Conditional (fixed-effects) logistic regression Number of obs = 315 LR chi2(1) = 11.98 Prob > chi2 = 0.0005 Log likelihood = -95.404465 Pseudo R2 = 0.0591 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 3.691907 1.372794 3.51 0.000 1.781317 7.651743 ------------------------------------------------------------------------------ . est store A
Gall bladder disease is highly significant - odds ratio is 3.69 compared to the 2.60 we found before
with 1-1 matching
Now add estrogen use
BIOST 536 Lecture 12 30
Example. clogit case gall estrogen, group(set) or Conditional (fixed-effects) logistic regression Number of obs = 315 LR chi2(2) = 45.05 Prob > chi2 = 0.0000 Log likelihood = -78.871308 Pseudo R2 = 0.2221 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 3.577465 1.469865 3.10 0.002 1.598984 8.003994 estrogen | 8.287802 3.644927 4.81 0.000 3.500144 19.62424 ------------------------------------------------------------------------------ . est store B . lrtest A B Likelihood-ratio test LR chi2(1) = 33.07 (Assumption: A nested in B) Prob > chi2 = 0.0000
Estrogen use also is highly significant - odds ratio is 8.29 (adjusted for gall bladder) - gall bladder
remains significant
Consider an interaction term between gall bladder and estrogen . xi: clogit case i.gall*estrogen, group(set) or i.gall _Igall_0-1 (naturally coded; _Igall_0 omitted) i.gall*estrogen _IgalXestro_# (coded as above) Conditional (fixed-effects) logistic regression Number of obs = 315 LR chi2(3) = 49.33 Prob > chi2 = 0.0000 Log likelihood = -76.730576 Pseudo R2 = 0.2432 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- _Igall_1 | 18.07166 15.95823 3.28 0.001 3.201415 102.0127 estrogen | 14.88179 9.104216 4.41 0.000 4.486594 49.36211 _IgalXestr~1 | .1283818 .1277365 -2.06 0.039 .0182633 .902457 ------------------------------------------------------------------------------
BIOST 536 Lecture 12 31
Example. est store C . lrtest B C Likelihood-ratio test LR chi2(1) = 4.28 (Assumption: B nested in C) Prob > chi2 = 0.0385
Statistically significant interaction between gall bladder and estrogen, consider the joint effects . lincom _Igall_1 + estrogen + _IgalXestro_1 , or ( 1) [case]_Igall_1 + [case]estrogen + [case]_IgalXestro_1 = 0 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- (1) | 34.52683 24.97059 4.90 0.000 8.366586 142.4837 ------------------------------------------------------------------------------
With an interaction term between gall bladder and estrogen
Estimated OR No estrogen Estrogen
No gall bladder 1.0 14.88
Gall bladder 18.07 18.07*14.88*0.1284 = 34.52
Without an interaction term between gall bladder and estrogen
Estimated OR No estrogen Estrogen
No gall bladder 1.0 8.29
Gall bladder 3.58 8.29*3.58 = 29.65
BIOST 536 Lecture 12 32
Consider a larger set of main effects to find significant predictors (ignore interactions here)
Need to be careful of missing values (obesity, estdose) . clogit case gall estrogen hyperten obes thyroid barbit tranq other, group(set) or note: 6 groups (15 obs) dropped because of all positive or all negative outcomes.
If the case has a missing value, the entire set is lost Conditional (fixed-effects) logistic regression Number of obs = 249 LR chi2(8) = 35.63 Prob > chi2 = 0.0000 Log likelihood = -65.021923 Pseudo R2 = 0.2150 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 3.564345 1.569089 2.89 0.004 1.504061 8.446836 estrogen | 6.743934 3.497201 3.68 0.000 2.440652 18.63463 hyperten | .7363322 .2787066 -0.81 0.419 .3506618 1.546177 obes | 1.50281 .6185417 0.99 0.322 .6707458 3.367055 thyroid | 1.039858 .414751 0.10 0.922 .4758495 2.272369 barbit | 1.126494 .4418284 0.30 0.761 .5222451 2.429871 tranq | .8861228 .3554218 -0.30 0.763 .4037187 1.944952 other | .5143312 .326884 -1.05 0.295 .1479995 1.787415 ------------------------------------------------------------------------------ . est store D . clogit case gall estrogen if obes~=. , group(set) or note: 6 groups (15 obs) dropped because of all positive or all negative outcomes. Conditional (fixed-effects) logistic regression Number of obs = 249 LR chi2(2) = 32.60 Prob > chi2 = 0.0000 Log likelihood = -66.535452 Pseudo R2 = 0.1968 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 3.514066 1.533611 2.88 0.004 1.493922 8.265933 estrogen | 7.210559 3.63667 3.92 0.000 2.683258 19.3765 ------------------------------------------------------------------------------
BIOST 536 Lecture 12 33
Example. lrtest D . Likelihood-ratio test LR chi2(6) = 3.03 (Assumption: . nested in D) Prob > chi2 = 0.8054
Only main effects of estrogen and gall bladder disease are risk factors
Look at estrogen dose coded 0-6 where 0 = lowest dose
Consider whether a dose-repsonse model is an improvement . table estrogen estdose ---------------------------------------------------- | estdose estrogen | 0 1 2 3 4 5 6 ----------+----------------------------------------- 0 | 132 1 | 62 22 16 40 16 12 7 ---------------------------------------------------- . clogit case gall estrogen estdose , group(set) or note: 4 groups (16 obs) dropped because of all positive or all negative outcomes. Conditional (fixed-effects) logistic regression Number of obs = 291 LR chi2(3) = 51.12 Prob > chi2 = 0.0000 Log likelihood = -68.502974 Pseudo R2 = 0.2717 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 4.883757 2.265871 3.42 0.001 1.967117 12.1249 estrogen | 6.030169 2.975007 3.64 0.000 2.292904 15.8589 estdose | 1.229028 .1265513 2.00 0.045 1.004419 1.503865 ------------------------------------------------------------------------------ . est store C
BIOST 536 Lecture 12 34
Example. clogit case gall estrogen if estdose~=. , group(set) or note: 4 groups (16 obs) dropped because of all positive or all negative outcomes. Conditional (fixed-effects) logistic regression Number of obs = 291 LR chi2(2) = 47.05 Prob > chi2 = 0.0000 Log likelihood = -70.539476 Pseudo R2 = 0.2501 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 4.84514 2.186536 3.50 0.000 2.000655 11.73385 estrogen | 8.676606 3.935269 4.76 0.000 3.566876 21.10628 ------------------------------------------------------------------------------ . lrtest . C Likelihood-ratio test LR chi2(1) = 4.07 (Assumption: . nested in C) Prob > chi2 = 0.0436
Significant difference so keep estrogen dose in the model . clogit case gall estrogen estdose , group(set) or Conditional (fixed-effects) logistic regression Number of obs = 291 LR chi2(3) = 51.12 Prob > chi2 = 0.0000 Log likelihood = -68.502974 Pseudo R2 = 0.2717 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- gall | 4.883757 2.265871 3.42 0.001 1.967117 12.1249 estrogen | 6.030169 2.975007 3.64 0.000 2.292904 15.8589 estdose | 1.229028 .1265513 2.00 0.045 1.004419 1.503865
BIOST 536 Lecture 12 35
ExampleCompute the OR’s if gall=0, est=1, for different choices of dose . lincom estrogen+0*estdose, or ( 1) [case]estrogen = 0 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- (1) | 6.030169 2.975007 3.64 0.000 2.292904 15.8589 ------------------------------------------------------------------------------ . lincom estrogen+1*estdose, or ( 1) [case]estrogen + [case]estdose = 0 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- (1) | 7.411248 3.463135 4.29 0.000 2.965795 18.52003 ------------------------------------------------------------------------------ . lincom estrogen+2*estdose, or ( 1) [case]estrogen + 2 [case]estdose = 0 ------------------------------------------------------------------------------ case | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- (1) | 9.108635 4.218685 4.77 0.000 3.674662 22.57819
Compute the conditional probabilities under this model . clogitp condprob, group(set) . list set case condprob gall estrogen estdose