Trimming quick responses from survey data

[Serena Menny]

Hello everyone,

I have the survey completion time for each individual (TIME) and would like to trim the quickest 5% of responses. Is there a direct way to perform this trimming, or should I use the drop command to remove these observations?

Thank you for you help,

[Rich Goldstein]

the user-contributed -winsor2- has a "trim" option; use -search winsor2- to find, download and install and then read the help file

you don't say why you want to do this but I generally believe this to be a bad idea

[Erik Ruzek]

I agree with Rich. What you might look at is the combination of timing and response usage. By that I mean cases where an individual completed a survey very quickly and always chose the same response option. That is probably a careless responder, and not one you want to keep in your data.

[Serena Menny]

"you don't say why you want to do this but I generally believe this to be a bad idea[

I plan to do this to check the robustness of my results. The goal is to ensure that my findings remain consistent even after removing respondents who completed the survey quickly, as they may not have been fully engaged.I’ve come across a paper that used a similar approach for this purpose.

[Nick Cox]

One way to tackle this is to treat completion time as another predictor, noting that you may need to scale or transform this in some way. Why not show us the distribution (making clear whether there was an upper time limit)?

The problem with 5% is why 5 and not 4 or 6 or 2 or 10 and why use a threshold any way? If you happily admit that 5% is arbitrary then that raises the question of whether any threshold is a good idea.

[Serena Menny]

One way to tackle this is to treat completion time as another predictor, noting that you may need to scale or transform this in some way. Why not show us the distribution (making clear whether there was an upper time limit)?

The problem with 5% is why 5 and not 4 or 6 or 2 or 10 and why use a threshold any way? If you happily admit that 5% is arbitrary then that raises the question of whether any threshold is a good idea.

The values were measured in seconds, and while there was no upper time limit imposed, respondents had the option to log out and continue at a later time, resulting in the high maximum values observed.
The decision to keep 5% was based on the article, where the same trimming threshold was applied. I do not have a solid reason for choosing 5%.

[Nick Cox]

You could try plotting

a histogram of log time or the reciprocal of time (the latter having an interpretation of speed)

any outcome of interest against that version of time, smoothed

That would give a solid sense of whether outcomes differ as you suspect. You would have a better case for segregating fast responses if they show up as a cluster.

[Serena Menny]

You could try plotting

a histogram of log time or the reciprocal of time (the latter having an interpretation of speed)

any outcome of interest against that version of time, smoothed

That would give a solid sense of whether outcomes differ as you suspect. You would have a better case for segregating fast responses if they show up as a cluster.

Thank you so much for the insight!