08.02.2013 Views

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

70 CHAPTER 4. (BIO-)MEDICAL APPLICATIONS<br />

be subjected to calculations known as linear regression. This works as a foundation<br />

<strong>for</strong> correlation analysis and to various o<strong>the</strong>r tests <strong>of</strong> statistical significance.<br />

Despite <strong>the</strong>ir ma<strong>the</strong>matical preciseness and undoubtedly correctness,<br />

inferences drawn from <strong>the</strong>se results may be incorrect, thoroughly misleading,<br />

or failing to call attention to <strong>the</strong> basic insufficiency <strong>of</strong> <strong>the</strong> experiment. This<br />

usually has two main reasons:<br />

� assumptions underlying <strong>the</strong> statistical procedure are not fulfilled<br />

� problems connected to <strong>the</strong> data were <strong>of</strong> a completely different type from<br />

those <strong>for</strong> which <strong>the</strong> particular statistical methods provide useful in<strong>for</strong>mation.<br />

Indeed, most data sets provide some useful in<strong>for</strong>mation, but this is no guarantee<br />

that <strong>the</strong> in<strong>for</strong>mation actually desired has been obtained.<br />

In most cases <strong>the</strong> goal <strong>of</strong> statistical analyses is to draw inferences from <strong>the</strong><br />

particular to <strong>the</strong> general and <strong>of</strong>ten people are not familiar with <strong>the</strong> problems<br />

<strong>of</strong> inductive inference, which is closely tied to this. R. A. Fisher has pointed<br />

to a basic and most important difference between <strong>the</strong> results <strong>of</strong> induction and<br />

deduction (Fisher, 1959) which we recall briefly and illustrate by two small<br />

examples. By using deduction, conclusion based on correct partial in<strong>for</strong>mation<br />

are always correct, despite <strong>the</strong> incompleteness <strong>of</strong> <strong>the</strong> premises. Let us use a<br />

well-known <strong>the</strong>orem from geometry as an example: <strong>the</strong> sum <strong>of</strong> <strong>the</strong> angles<br />

<strong>of</strong> a plane triangle always equals to 180 degrees. This does not necessitate<br />

in<strong>for</strong>mation as to whe<strong>the</strong>r it is isosceles or not. If any in<strong>for</strong>mation <strong>of</strong> this<br />

type is subsequently added, it cannot possibly alter <strong>the</strong> fact expressed by <strong>the</strong><br />

<strong>the</strong>orem.<br />

As an counterexample, inferences drawn by induction from incomplete<br />

in<strong>for</strong>mation may be entirely wrong, even when <strong>the</strong> in<strong>for</strong>mation given is unquestionably<br />

correct. Let us use a simple example from physics. Suppose we<br />

were given <strong>the</strong> data <strong>of</strong> Table 4.2.3 on <strong>the</strong> pressure and volume <strong>of</strong> a fixed mass<br />

<strong>of</strong> gas. Analyzing <strong>the</strong> data one might infer (by induction) that <strong>the</strong> pressure<br />

Molar volume (liters) Pressure (atmospheres)<br />

0.182 54.5<br />

0.201 60.0<br />

0.216 64.5<br />

0.232 68.5<br />

0.243 72.5<br />

Table 4.2.1: Volume-Pressure Relation <strong>for</strong> a gas, an apparently proportional relationship<br />

<strong>of</strong> a gas is proportional to its volume (<strong>of</strong> course a completely erroneous statement).<br />

What went wrong ? The answer is simply that ano<strong>the</strong>r important<br />

item <strong>of</strong> in<strong>for</strong>mation was omitted, namely that each pair <strong>of</strong> measurements was<br />

obtained at a different temperature, as indicated in Table 4.2.3. Of course,<br />

this example is artificially constructed and extreme but it emphasizes <strong>the</strong> basic<br />

problem in inductive reasoning: <strong>the</strong> data not only has to be correct but also<br />

complete to enable correct inference. In this simple example <strong>the</strong> missing piece<br />

<strong>of</strong> in<strong>for</strong>mation was easily identifiable because we have a good understanding<br />

<strong>of</strong> <strong>the</strong> physical background. What researchers need to be aware <strong>of</strong> is that

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!