0
$\begingroup$

I'm doing my Master in Occupational Safety and Health. My research topic is on the association between working condition, safety behaviour and work-related injuries among construction workers.

Safety behaviour was measured for individual workers by observation checklist "Safe - Unsafe and N/A". For some items like Personal Protective Equipment, some workers in specific situation don't have to wear some specific PPE; for example ear defender; when Noise is not exceeding the Threshold limit this PPE in checklist is N/A.

So, the issue is what value should I give to N/A to calculate the overall score for safety behaviour ? Or should I consider it as a missing value ? If so, the overall score will be calculated only for the valid cases and I can not determine the association..

$\endgroup$
0

1 Answer 1

0
$\begingroup$

In general, missing data comes back to the old thing of whether data is "missing at random" or not, or, alternatively and more broadly, whether the inclusion of data for these data points would impact the overall results.

For example, if your missing data was from one specific building site, where there were variables which were unique to that site (such as using an unusually noisy piece of equipment, say), AND if the data which were missing were all relevant - like whether or not workers who should be wearing ear defenders were doing so, I would say that that might be a problem, depending on what exactly it is you're trying to analyse.

If, however, data were randomly missing from 5% of the workers at all sites, that wouldn't be so much of a problem unless you had reason to believe that those 5% at each site shared some commonalities (sticking with the analogy, they were the 5% who all used an especially noisy drill, or whatever) which set them apart from the rest of the workers at those sites.

In your case, the reason for the omission is that the data are not relevant - they are not required to wear the PPE. This is not the same as "missing at random", but equally it has a simpler effect - it simply isn't relevant. Unless you're analysing whether or not the restrictions which are in place in terms of who has to wear what are valid (and it doesn't seem like you are), then this is data which is not missing-at-random, but is simply not relevant at all.

My instinct is therefore that it is legitimate for you to simply ignore these data - take them out. For example, if working out percentages (say, percentage of workers who got an injury despite wearing the PPE), just remove these NA values from the total when summing and dividing, etc. Possibly structure your model in such a way that it ignores these values, or assign them zero values (but obviously beware that depending on the calculation - for example, averages - zero values can have a distortive impact, though if you're doing an MA in a statistical subject I'm sure you're already aware of that).

I'm sure others on this site might have a different opinion. I'm a continuous improvement/data analysis professional in operations and that would be my take on it, assuming that the assumptions I've outlined above are valid.

$\endgroup$
2
  • $\begingroup$ Please could you review this and mark as the answer if it answers your question? $\endgroup$ Nov 30, 2016 at 16:09
  • $\begingroup$ Dear @Statsanalyst, I see your answer to this question, and I ask myself if the exploration of the missing to check whether these values are at random (mcar) or not is not an old thing. Actually, the recommendations from the literature relies mainly on checking the nature of missing data before any statistical procedure. Could you please provide a citation supporting your answer? Thank you. $\endgroup$
    – Luis
    Aug 21, 2019 at 14:34

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