Typically combining consecutive results from the same pseudorandom number generator is inferior to using a single result. Think of it this way: the random number generator generates a certain number of results before cycling back to the first number it provided. If you use up four results at a time, there are only one quarter as many results that you can get - assuming that the possible number of results is a multiple of four, which is likely on a binary machine - so a brute force attack would only take one quarter as long to break your data.
In addition, the behavior of the data itself may be problematic when you combine consecutive results. For example, many random number generators provide individual numbers which have good statistical distributions, but this doesn't necessarily apply to combinations of consecutive numbers, where the random number generator may be more or less likely to generate runs of similar results than would a truly random source. For example, I once used a random number generator to simulate rolls of a six sided die, and when I summed three consecutive results, the behavior of consecutive results was such that I never got a sum of 18, over many thousand attempts.
I discuss consecutive results, but these arguments apply to use of nonconsecutive results as well when they are selected in a systematic way. Your best bet is just to select a better random number generator and use the results one at a time. The only time when you need to combine multiple results is when a single result doesn't have enough bits to satisfy your needs - rare with off the shelf pseudorandom number generators.