This study is the first to consider the influence of dietary patterns
on the development of and death from cancer and CVD in
older adults. The measurement of chronic disease was subjective in
its basis on participant self-reports of diagnosis, which introduces
the potential for recall bias. Ideally medical diagnoses would be
used to identify cancer and CVD, however these clinical data were
not collected as part of the ALSA and hence present a limitation of
this study worthy of acknowledgement. Furthermore, cancer was
identified using the variable ‘cancer diagnosed’, which encompasses
all cancer types, not all of which are related to diet. Logistic
regression analysis was trialled using diet-related cancers, however
this resulted in an insufficient sample size and the goodness of fit
tests for the models produced did not reach statistical significance.
It is therefore possible that the lack of association observed for
patterns other than ‘red meat and protein alternatives’ may be
attributed to inclusion of non-diet-related cancers. There was,
however, a long follow up time between baseline and wave 6 which
allowed 8 years for cancer and CVD development, thereby optimising
the potential sample size for analyses and minimising potential
for type 2 error. The use of dietary pattern analysis, however,
can only be as good as the method of dietary assessment upon
which it is based. The FFQ is a retrospective measure of diet which
carries the potential for recall bias, which is of particular concern
given the demographics of this study sample. The FFQ used did not
allow differentiation between wholegrain and refined bread and
cereal products, and whole fat and low fat milk and dairy products,
which may have masked any associations with dietary patterns
which incorporated these food groups and disease outcomes.
However, the reproducibility and validity of dietary patterns using
data from FFQs has been demonstrated [36]. Standard protocols
were used for all measurements, and trained professionals conducted
all clinical assessments including anthropometric measures,
which reduces the potential for systematic error. Finally, the selection
of the ALSA cohort from the South Australian electoral roll
database was random, thereby minimising the potential for selection
bias and increasing confidence in the generalisability of the
findings.