There little advantages to do that by using pixels[]. It is interesting for one pixel, much less for an area, unless you want to bypass color conversions, perhaps.

Anyway, the solution is to use two nested loops, one on the width, the other on the height. The formula should look like: pixels[x + horizontalIterator + (y + verticalIterator) * width]
although I haven't tested it, so it can be wrong.

Honestly, you really are better off using get() to obtain the rectangle you're interested in as another, separate, PImage object.  Work with the pixels[] array of that object.

For one thing, doing it this way means you won't have to constantly account for the offsets from the source image's origin to the start of the rectangle you want.  You're thus much less likely to introduce bugs by getting the formula wrong somewhere.

For another thing, using get() insulates your code against any future changes in Processing's implementation of PImage; if Fry decides, for instance, that PImage.pixels should really have been a two-dimensional array of pixels, rather than a linear array, he can do that and update the implementation of get() so that it still does the right thing.  This minimizes the amount of code in your sketch that you'd have to change if anything like that happened.

Not that I'm saying it will, just that if somebody provides you an object type with a well defined interface, it is considered better programming practice to use that interface rather than essentially re-implementing parts of that interface yourself by means of accessing the object's underlying data directly.  I mean, why re-invent the wheel?