How one measures the outcome of the experiment depends mainly on the available technology. However, there is often room for choices and it is important to keep an eye on the way in which data will be analyzed to optimize the results. In general, the way in which measurements are taken determines the nature of the error: because statistical analysis is all about recovering the true signal from the error, this is obviously very important.
Measuring the outcome of a microarray experiment is a complex matter; one step of the process involves transforming a two-dimensional picture in an array of numbers. The image segmentation technique used here, for example, can bias the results in different directions, so that it is important to discuss it with the statistician. It is important to correct for background intensities and the appropriate background values vary across the array. Attention should be paid that the background measurements are not excessively noisy: a robust and stable method for background estimation is needed (presumably using a smoothing procedure).
If the measurements are taken as ratio, it is often convenient to take the Log of these numbers making their distribution symmetric, so that a departure from the average in each direction has the same meaning. Excessively high or low expression values should be probably considered as outliers. Dye-renormalization is often needed to take into account differential incorporating power of the dyes. As this interact with the sensitivity of the scanner, the difference in luminosity between the dyes has been noted to follow a non-linear pattern. So that normalization should be non linear. Having a quadrant-dependent normalization is also a good idea. Care should also be taken in deciding which genes to use for normalization.
References The following groups have analyzed in detail the imaging, background correction and normalization of the signals.