Ideally, to exclude technical variation based on preprocessing you would do an identical upstream processing. But since this is no option as you say, I don't see any choice other than using existing data for your meta analysis. After all, meta-analysis asks whether your significant calls are consistent across datasets, so I tend to say that upstream processing does not matter too much. Of cource, for genes that are difficult to map you might get insignificant results simply because aligners ahndle them differently. That's a limitation.