For our Ranked-Pairs analysis, we reject ballots that are entirely blank, or those for which any preference is non-numeric or is a number not between “1” and the number of candidates (inclusive).
The Ranked-Pairs analysis is not disturbed by ballots giving different candidates the same preference, or by not-specifying a preference for one or more candidates, so ballots ranking different candidates the same are accepted.
For FPTP, we take the “most-preferred” candidate as the ballot’s first-past-the-post choice. If there is more than one such candidate we cannot count the ballot for either candidate, and it is therefore rejected. This means that ballots that were accepted for Ranked Pairs can be rejected for FPTP. Since we only have one “round” of counting for FPTP, such a conflict is never resolved.
In the case of IRV, as for FPTP, we can only use a ballot if the given preference is unique to a given candidate. For IRV, however, if on any given counting round we have not yet determined a majority winner, we eliminate the lowest-count candidate, and re-allocate this candidate’s ballots according to their indicated next-preferences. If the given next-preference is not-unique, however, we cannot assign the ballot to a single candidate, and the ballot must be rejected.
But, since we’re eliminating candidates, we also re-evaluate earlier-rejected ballots in case ambiguities that caused them to be rejected have been resolved upon the elimination of the given candidates. This means that some previously rejected ballots might be returned to play, and the number of rejected ballots can be reduced in subsequent IRV rounds.
Whether rejected ballots would be handled thus in any given electoral system is up to the rules established by the particular enabling legislation; however, the rationale here is to attempt to use a ballot if it is meaningful to do so.
This also means that, when there is a significant difference in rejected ballots between one system or another, the comparisons are not quite apples-to-apples. For comparison purposes, then, it would be better to avoid data where such discrepancies occur.