Page 62 - Spring2012

2002. 29-31 Oxford Journal, “A simple rule of thumb
In the 2002 CEBP study, Horn-Ross, Hog- is to use conditional logistic regression if
gatt and Lee attempted to determine how soy matching has been done, and unconditional
phytoestrogen intake relates to thyroid cancer if no matching has been done. A second
once other factors such as age, race and other rule of thumb is, when in doubt always use
known risk factors were taken into account. In conditional because it always gives unbiased
the results section they reported, “In general, a results.” 35
reduction in thyroid cancer risk of 35 percent
to 55 percent was associated with increased • Because the study used unconditional lo-
consumption of non-fermented traditional and gistic regression, the researchers did not
nontraditional soy-based foods and sprouts.” 32 include the matching information in the
36
An astonishing 35 to 55 percent reduction in analysis. This is most interesting in the
risk with clear cause and effect certainly seems light of research from the University Gradu-
to support the idea of consuming soy—includ- ate School of Public Health in Kyoto, Japan,
ing modern industrial soy products—for thyroid which examined 507 studies from 1991-2000
37
cancer prevention. But what seems to be too that used case control matched data sets. Of
good to be true is often just that. A long, hard these studies, conditional logistic regression
look at the study—and not just at the headlines was used in 90.5 percent, and unconditional
publicized by the soy industry—reveals serious logistic regression in only 9.5 percent of
flaws in design, methods, and analysis, including: them. Yet Horn-Ross and colleagues chose
to use the unconditional method.

• This paper describes an observational, case- • Unconditional logistic regression analy-
control, matched study. As J.M. Utts and sis seriously overestimates the odds ratio
R. Heckard write in their textbook, Mind when there are matching data—as was the
on Statistics, “The most common mistake case with Horn-Ross and colleagues—and
made in reporting research studies is to im- great caution should be taken in interpret-
38
ply that a cause and effect relationship can ing the results. In Statistical Methods in
be concluded from an observational study. Cancer Research, a classic text in disease
With an observational study, it is difficult, epidemiology, Breslow and Day state: “The
perhaps impossible, to separate the effects unconditional analysis of matched pair data
of confounding variables from the effects of results in an estimate of the odds ratio which
the main explanatory variables of interest.” 33 is the square of the correct, conditional one:
a relative risk of 2 will tend to be estimated
• The study was not a randomized, controlled as 4 by this approach…” (italic emphasis
trial, which is the gold standard for testing from Breslow and Day). 39
What seems an intervention. Cases were not randomized • The spotlighted phytoestrogens yet included
to be too to treatment groups but drawn from a cancer a large number of potentially interrelated
registry, which was a sample of convenience.
good to be As Utts and Heckard put it, “If the sample variables that could interact with one an-
true is often does not represent a larger population for other. In a high quality study, the researchers
should have addressed the possibility of col-
just that. A the question of interest, and randomization linearity and taken care to rule it out. Col-
to treatments was not used, no inferences
long, hard look can be drawn.” 34 linearity is a bias in statistical procedure due
at the study • The data were analyzed using unconditional to the correlation of multiple independent
reveals serious logistic regression. When the sample comes variables that influence a single dependent
variable. Collinearity can lead to unstable
flaws in design, from matched pairs—as was the case in and untrustworthy results. 40
methods and this study—conditional logistic regression • All the subjects came from the San Fran-
is the appropriate test, not unconditional
analysis. logistic regression. As summed up in the cisco Bay Area and many were of Asian

62 Wise Traditions SPRING 2012 SPRING 2012 Wise Traditions

89494_text.indd 62 3/13/12 1:47 AM