If you use single years of education, then this recode will not distinguish between, say, the husband having one year more or having ten years more. I would construct the difference as "gen eddiff=mv133-v133". I am intentionally subtracting the woman's v133 from the man's mv133 because (like it or not) in most countries this difference will be positive, and it is easier (for me at least) to work with differences that are usually positive than with differences that are usually negative. I would then do "tab eddiff". If you need to categorize, construct the categories from this empirical distribution so you get 3 or 5 or whatever number of categories that have similar numbers of couples. Better yet, use all the information--do not categorize, but just use the difference as an interval-level covariate.