Guidelines:
Useconditionalif matching
Useunconditionalif no
matching and number of
variables not too large
Safe rule:
Useconditionalwhen in doubt.
Gives unbiased results always.
Unconditional may be biased
(may overestimate odds ratios).
The above examples indicate the following
guidelines regarding the choice between
unconditional and conditional ML methods or
programs: Useconditional ML estimationwhenever
matching has been done; this is because the
model will invariably be large due to the
number of dummy variables required to
reflect the matching strata.
Useunconditional ML estimationif
matching has not been done, provided the
total number of variables in the model is
not unduly large relative to the number
of subjects.Loosely speaking, this means that if the total
number of confounders and the total number
of interaction terms in the model are large, say
10–15 confounders and 10–15 product terms,
the number of parameters may be getting too
large for the unconditional approach to give
accurate answers.A safe rule is to use conditional ML estimation
whenever in doubt about which method to
use, because, theoretically, the conditional
approach has been shown by statisticians to
give unbiased results always. In contrast, the
unconditional approach, when unsuitable, can
give biased results and, in particular, can over-
estimate odds ratios of interest.As a simple example of the need to use condi-
tional ML estimation for matched data, con-
sider again a pair-matched case-control study
such as described above. For such a study
design, the measure of effect of interest is
an odds ratio for the exposure-disease rela-
tionship that adjusts for the variables being
controlled.EXAMPLE
Unconditional questionable if
10–15 confounders
10–15 product termsEXAMPLE: Conditional Required
Pair-matched case control study
measure of effect; OR110 4. Maximum Likelihood Techniques: An Overview