Individual scores are determined by the underlying probability of choice distribution for each level (determined from overall responses) and also their individual responses, which dictate where they fit on the underlying distribution. This occurs for each level, subject to constraints, which veto possibly better fitting, but less logical solutions (in effect the analyst determines that non-logical responses are just random noise). I’m unclear if ‘similar respondents’, defined by an overlap of 'known' levels, are used to fill in the ‘missing data’ for each respondent, where the known information does not overlap, or if this is just inferred from the overall probability distribution.