The "pedestal effect" is a DC offset or bias that is leftover in an image after it has had the dark reference file subtracted from it. Normally this reference file would be a perfect representation of the actual instrumental bias and dark current signature of the array, but some or all of these components are not stable on both short and long timescales. It is believed that these changes may be thermally driven and are either in the electronics or the detectors themselves or a combination of both. Typical pedestal amplitudes are a few DN up to ~15 DN, but occasional reports of much larger values have been received.
The effect of leaving a residual bias in the image during the calibration process is that the flatfield reference file is then multiplied by this constant. The flatfield is used to normalize the relative sensitivities of all the pixels in the array to each other. If the mean value in an image with no sky signal is 0, then multiplying the entire array by a normalizing image whose mean is 1 is also 0. If you then add a bias to that image, you multiply the flatfield response by that value. If the flatfield image were perfectly flat, the result would be a constant added to the image. Since the NICMOS sensitivity has quite a bit of spatial structure in it, the effect is to add a multiple of that pattern to the data. Since the flatfield reference files are inverted flats, the pattern that is added to an image is a multiple of the inverse sensitivity of the array. (that is - it looks like someone has stamped your data with a negative flatfield). Of course the sense depends on the sign of the residual bias, but typically it is a positive residual.
The image shown here has been calibrated through the calnica pipeline and shows the effect of pedestal in an image. Note that each quadrant can have its own pedestal value.
A number of groups have developed code for detecting and removing the pedestal from data. The simplest form involves just measuring and subtracting the median value of each quadrant before the flatfielding step. Care must be taken to account for background signal, which must be left in the image through the flatfielding step, as it does indeed modulate with the detector sensitivity structure. This method is very sensitive to the amount of real signal in the images, and so works best with sparse data. More sophisticated methods use iterative algorithms to measure the pedestal by minimizing the amount of falsely imprinted flatfield pattern on both large and pixel-to-pixel scales. These tend to work much better on large-scale low surface-brightness objects such as galaxies, and certainly will work with sparse data as well.
Another issue that must be contended with is the spatial structure of the pedestal. Most algorithms currently assume a constant pedestal per quadrant. To first order this works quite well, but it has recently been discovered that the detector shading is time (or more likely, temperature) dependent (see the section on "shading" in this document). The change in the shape of the shading will result in a leftover bias after the dark subtraction that is *NOT* a constant across a quadrant, but rather has a spatial structure like a ramp or smooth curve. This change in the shading function is probably the major constituent of the pedestal as we know it.