Downsampling is essentially taking an image and averaging neighboring pixels together to produce an image with fewer pixels than before. By averaging together neighboring pixels, image noise can be reduced at the expense of resolution. For example, if a 6000x4000 image is re-sampled to 3000x2000, then each pixel in the new image is effectively derived from the signal of 4 pixels in the original. (In science this is called "binning"--we do this all the time with high resolution CCD sensors to increase weak signals. Four photosites collect 4 times the signal as one photosite, and signal-to-noise increased by the square root of 4, or 2-fold. Of course we get 1/4 of the image resolution this way, but S/N is often more desirable than resolution.
To re-iterate an earlier comment, "silicon is cheap." Shoot at full resolution, and downsample later if you must. You can always down-sample more pixels into fewer pixels later. You can't go the other way around, because the information is already gone if you downsample in-camera.