More information about text formats
I read with interest the article by Joshy et al on prevalence of
diabetes. The authors aimed at assessing the influence of deprivation on
the prevalence of diabetes and have used cross-sectional study design. The
authors estimated odds ratios using logistic regression. There are two
fundamental interpretative issues in using odds ratio as a measure of
association in cross sectional studies.
The first issue is...
The first issue is that the odds ratios tend to overestimate the risk
ratio when the outcome events are common (more than 10%). For instance, if
80 out of 100 exposed persons have a particular disease and 40 out of 100 non
-exposed persons have the disease, the OR is 6, but the exposed persons
are only 2 times more likely to have the disease as the non-exposed. The
second issue relates to the common tendency to interpret it as relative
risk, which is misleading both in theoretical and practical terms. In the
previous example, it is misleading if the exposure is considered to be
related to six-fold increase in the chances of getting disease. So an
appropriate measure of association for cross-sectional study designs is
the prevalence ratios.
I suppose the common use of ORs as effect measures in cross-sectional studies was due to greater availability of computer programs
which produce OR as standard output of the fitting of logistic regression
models. Though the use of OR is not intrinsically wrong, with recent
advancements in computing methods, alternative statistical models like
log-binomial regression can be used to directly estimate the prevalence
1.Barros AJ, Hirakata VN. Alternatives for logistic regression in cross-
sectional studies: an empirical comparison of models that directly
estimate the prevalence ratio. BMC Medical Research Methodology.