3. general considerations for the analysis of case-control ... - IARC
3. general considerations for the analysis of case-control ... - IARC
3. general considerations for the analysis of case-control ... - IARC
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
GENERAL CONSIDERATIONS 11 1<br />
confined to those first employed be<strong>for</strong>e 1930. Changes in <strong>the</strong> operation <strong>of</strong> <strong>the</strong> refinery<br />
at that time could, quite plausibly, have removed <strong>the</strong> carcinogenic agents, and <strong>the</strong><br />
change in risk assists in identifying what those agents may have been.<br />
<strong>3.</strong>6 Modelling risk<br />
The use <strong>of</strong> stratification and cross-tabulation to investigate <strong>the</strong> joint effect on risk<br />
<strong>of</strong> two variables, in terms <strong>of</strong> how <strong>the</strong> two factors mutually confound each o<strong>the</strong>r and<br />
interact, is reasonably straight<strong>for</strong>ward. However, even with two variables, as <strong>the</strong> number<br />
<strong>of</strong> values each variable can take increases, <strong>the</strong> <strong>control</strong> <strong>of</strong> confounding by means <strong>of</strong><br />
stratification can lead to substantial losses <strong>of</strong> in<strong>for</strong>mation, and tests <strong>for</strong> interaction will<br />
lack power. As <strong>the</strong> complexity <strong>of</strong> <strong>the</strong> problem increases, <strong>the</strong> approach via stratification<br />
becomes not only unwieldy but increasingly wasteful <strong>of</strong> in<strong>for</strong>mation. The different<br />
effects associated with different levels <strong>of</strong> a variable will not normally be unrelated, and<br />
can be expected to change smoothly. For example, <strong>for</strong> a quantitative variable, risk will<br />
usually vary in a manner which can be described by a simple family <strong>of</strong> curves. It would<br />
be rare to need more than second degree terms, after perhaps some initial trans<strong>for</strong>mation<br />
<strong>of</strong> <strong>the</strong> scale.<br />
Similarly, interactive effects between several variables will not normally vary in a<br />
structureless way, and <strong>general</strong> experience has been that most situations are well described<br />
by some simple structure <strong>of</strong> <strong>the</strong> interactions. These <strong>considerations</strong> lead to <strong>the</strong><br />
use <strong>of</strong> regression methods, in which <strong>the</strong> risk associated with each variable is expressed<br />
as some explicit function, and interaction effects are described in terms <strong>of</strong> <strong>the</strong> specific<br />
parameters <strong>of</strong> interest. Chapters 6 and 7 are devoted to <strong>the</strong> development <strong>of</strong> <strong>the</strong>se<br />
methods, which will not be fur<strong>the</strong>r discussed here except to outline briefly <strong>the</strong>ir advantages.<br />
These can be summarized as follows:<br />
1. One can study <strong>the</strong> joint effect <strong>of</strong> several exposures simultaneously. The stratification<br />
approach we considered earlier places <strong>the</strong> emphasis on one specific exposure.<br />
Study <strong>of</strong> <strong>the</strong> combined risk associated with several exposures is an important complement<br />
<strong>of</strong> <strong>the</strong> single exposure <strong>analysis</strong>.<br />
2. When <strong>the</strong> number <strong>of</strong> levels <strong>of</strong> <strong>the</strong> confounding variables increases, one can remove<br />
<strong>the</strong>ir effect as fully as by fine stratification but with less loss <strong>of</strong> in<strong>for</strong>mation.<br />
<strong>3.</strong> One can test <strong>for</strong> specific interaction effects <strong>of</strong> interest with <strong>the</strong> considerable increase<br />
in power this provides. One also obtains a parametric description <strong>of</strong> <strong>the</strong> interaction.<br />
4. The risk associated with different levels <strong>of</strong> a quantitative variable can be expressed<br />
in simple and descriptive terms.<br />
In studies where <strong>control</strong>s are individually matched to <strong>case</strong>s, <strong>the</strong>se advantages are<br />
accentuated, as Chapters 5 and 7 make apparent. But regression methods should not<br />
replace analyses based on cross-tabulation, ra<strong>the</strong>r <strong>the</strong>y should complement and extend<br />
<strong>the</strong>m, as we illustrate in Chapter 6.<br />
<strong>3.</strong>7 Comparisons between more than two groups<br />
So far, we have considered methods <strong>of</strong> <strong>analysis</strong> appropriate <strong>for</strong> comparisons between<br />
one <strong>case</strong> group and one <strong>control</strong> group. Situations occur, however, when comparisons<br />
among more than two groups are required. One may want to test whe<strong>the</strong>r <strong>the</strong> relative