0
$\begingroup$

I have an ordinal variable describes self-assessments of IT/ICT literacy with 0-100 scale (only integers), where 0 – very bad, 100 – very good (sample size: 10000 respondents). I want to investigate the impact of some features (sex, age, education level etc.) on the value they declared. I recoded my variable to binary one based on their answers (0 to 50 – people with bad IT skills, 51-100 people with good IT skills) to be able to use logit\probit regression. For some reasons I want/need to use binary variable rather than categorical one with multiple levels\ gradations. The problem is that my supervisor told me that:

“A cut-off of 50 to decide whether or not respondent possess’ low IT skills seems far to crude”

I am looking for the best way of recoding my ordinal variable to binary one. May be it would be better to use clusterisation algorithms to divide respondents into 2 clusters based on self-assessments (if they exist). Would be also appreciated for references/scientific papers with similar problem.

Improve this question
$\endgroup$
3
  • 4
    $\begingroup$ Why do you want to dichotomize your data at all? Is this just to train using logit models? Dichotomizing loses a lot of information, usually for no benefit at all. I would recommend that rather than looking for better ways to lose information, you start modeling your data on a continuous scale (or at least an ordinal one). $\endgroup$
    – Stephan Kolassa
    Commented Dec 16, 2021 at 17:36
  • $\begingroup$ self-assessment equals 73 means to Me nothing, while if I use binary approach 73 would means that particular person roughly speaking has good IT skills. This is the only one reason I use binary data. $\endgroup$
    – Jeparov
    Commented Dec 16, 2021 at 18:48
  • 1
    $\begingroup$ That may well be. But if you use a threshold of (say) 57, then you treat people scoring 1, 23 and 56 exactly the same. Self-assessment does have its issues, true (a scale of 1-100 is probably far too fine-grained, Likert scales of 1-7 are more common), but people self-describing as 1, 23 or 56 probably do differ a bit. $\endgroup$
    – Stephan Kolassa
    Commented Dec 16, 2021 at 19:03

0

Reset to default