1
$\begingroup$

I have some missing data for gender in a study (between 10-20%) and am wondering how to address it. There are methods that assign gender based on first name. However, this came up in discussions of them:

https://dapperstats.github.io/gendrendr/

I'm interested in:

  1. thoughts on the above link
  2. if one is not going to assign gender based on first name, what is the next least bad option? just dropping the missings?

UPDATE with further information: The gender for those with ‘known’ gender were provided by the respondents themselves, and I trust this as valid. There are too many subjects to ‘google’, and I fear that could lead to the same criticism as the name imputation: assigning a gender to someone without their consent. The results I get are that one gender has a higher rate of success than the other overall, and the Unknown as the lowest rate, which perhaps I might interpret as less conscientious in terms of proving complete demographic information, less conscientious overall :) Currently, in the descriptive analysis I am: 1) showing applicants/selections data with the unknown category included in the totals and 2) also dropped (so not included in the totals). The problem is determining to what extent those in the unknown could skew the overall results.

$\endgroup$
4
  • 2
    $\begingroup$ This is an opinion-based question. I think you should start by examining your objectives in seeking to use gender in your study. // Some names give strong evidence of gender-assignment at birth (e.g, Richard, Robert, Maria, Julia); one might assume people who are no longer comfortable with that assignment would have changed their names. Other names might give you almost no such information (e.g., Chris, Kim, and some names from non-Western cultures), // The link expresses a strong point of view; unfortunately, agreement or disagreement with it could lead to immediate controversy. $\endgroup$
    – BruceET
    Apr 6 '21 at 15:26
  • 1
    $\begingroup$ Thank you. I think that name list imputation will be frowned on in the particular climate in which I work (due to the above link, for example), and thus need another option. With the name list imputation, there are also individuals who do not identify as male or female. I am looking at differences between men and women in their success rates at getting a particular award. $\endgroup$
    – Laura
    Apr 6 '21 at 16:46
  • 1
    $\begingroup$ Not enough info for me to venture a recommendation. Questions to ponder: How did you get gender info for the subjects for whom you judge it as 'known', and do you trust the method as valid? // Are there too many subjects to 'google' for substantial clues? (News item: "First woman to win XYZ Award.") // What results do you get using M/F/U where U is for unknown/unspecified? How would you interpret results if U's are chosen least often? $\endgroup$
    – BruceET
    Apr 6 '21 at 18:16
  • $\begingroup$ Thank you. I updated the post in response since I couldn't fit all the responses here. $\endgroup$
    – Laura
    Apr 6 '21 at 19:54
2
$\begingroup$

Your first question is not really on-topic on this site, but what you decide to think about it might have consequences for the modeling. You should also ask yourself why gender is missing at such a large rate, might it have to do with people actively refusing to answer? So, depending on context and goals, missing might be informative, so just dropping might be bad.

Your question have little to do with imputation. To impute the missing data, you need to use the same conceptual variable definition as used with data set construction. So if a binary definition was used then, you must use the same def when imputing. Uncertainty in the imputation, whether by first names or otherwise, can be represented with multiple imputation.

$\endgroup$
6
  • $\begingroup$ Thank you. Yes, the missing might have to do with some people actively refusing to answer, and it could be that the rate at which they refuse differs according to men and women. However, without actually doing the name imputation and looking at who is in the missing, I'm not sure how you could even try to see if they refuse at different rates. $\endgroup$
    – Laura
    Apr 6 '21 at 19:56
  • 1
    $\begingroup$ Maybe, instead of worrying about whether the 'refusal' rate is different for men or women, it would be more useful to consider that there are (at least) three 'genders' M/W/U. // Nowadays it has become customary to allow more than two options in the 'gender question' on a survey. [Perhaps some researchers do that as a mandatory accommodation to 'political correctness', and some because they have learned they get more useful data that way.] $\endgroup$
    – BruceET
    Apr 6 '21 at 20:11
  • 1
    $\begingroup$ (We don't know which other data you have ...) you could try imputation different ways, and compare . $\endgroup$ Apr 6 '21 at 21:48
  • $\begingroup$ The other fields I have in the data are other demographic variables, area of work, and information on education. My concern with multiple imputation is that the data might be MNAR – the missingness might depend on whether someone is male or female. $\endgroup$
    – Laura
    Apr 7 '21 at 16:57
  • 1
    $\begingroup$ It might, but some of the demographic variables such as education and area of work might be informative ... worth a try $\endgroup$ Apr 7 '21 at 21:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.