score for f__numeric_response

[andreaskutka]

numeric_response

make it an issue, also take out special values for some of the numeric based features

Calculate score for feature f__numeric_response .

Feature

f__numeric_response contains the response to all questions of type NumericQuestion in the microdata.

Rationale

1. At a minimum automatically detect single variate outliers. Check distribution, if goes over orders of magnitude, take ln(). Work with iqrs or mad better than SD from mean as less outlier prone.

2. Ideally, it would be great to also look into multivariate outlier detection, i.e. how weird is a response, given the other responses. Not clear how this could be automated, and how memory heavy this would be.

3. Instead of only looking for outliers (just at the extremes), it would be great to also normalise the in a meaningful and outlier independent way, to get a measure of how extreme/non-extreme they are. The hypothesis is that cheaters attempt to avoid extreme values.

Note, numeric responses may contain "special values", often -99,-98 or 99, 999, 9999. They vary from survey to survey, but ideally are outside of the valid range. There might be multiple special values per question. They should be the same for the question and ideally for the entire questionnaire, but often people are sloppy. We can try to identify them automatically (outliers, that are the same number), or if not possible, ask this as input from the user. If they used the special value feature from Survey Solutions, we might be able to extract it.