When I've asked a teacher (in the past) about how could I handle, in a survey with simple random sampling, the participants I couldn't reach for some reason, he told me that I could use a pre established "replacement procedure" (and by replacement, he doesn't mean the usual concept of selecting the same individual more than once).

According to him, I could (prior to the survey) state that, for instance, non-respondents (after applying some criteria), could be disregarded and replaced by some new randomly assigned population member. So, if I prior design some strict procedure for that, the sampling wouldn't be biased by nonresponse - or would be less biased, at least.

So here's my question: Is there any literature and/or reasonable foundations supporting this method? What are the main advantages and drawbacks of this approach? And what are the other alternatives?

Considering that I have rich prior information about the population members, does it change anything in the approach choice for mitigating nonresponse?

Edit Summary

  1. Reframed question after comment and further reading.
  2. "New randomly assigned respondent" was misexpressing what I've meant, giving the idea of replacement using another respondent's response! "New randomly assigned population member" better expresses my question, that is selecting a new random participant.
  • $\begingroup$ You're talking about using existing responses to impute the missing responses. There's absolutely no way to answer your question with YES or NO. It's one of the possibilities. $\endgroup$
    – SmallChess
    Sep 18, 2017 at 6:16
  • $\begingroup$ @SmallChess, so, should I edit and ask about the advantages and drawbacks of this approach? $\endgroup$
    – Bruno
    Sep 18, 2017 at 6:18
  • $\begingroup$ Done. Please, confirm if it's better now, @SmallChess. Thanks for your feedback!! $\endgroup$
    – Bruno
    Sep 18, 2017 at 6:32
  • $\begingroup$ I'm just wondering... If I do that, wouldn't I be creating another bias? $\endgroup$
    – Bruno
    Sep 19, 2017 at 11:52

1 Answer 1


Replacing the nonresponding units is indeed a common practice in survey statistics. (Mind you, the practice of survey statistics is about as remote from sampling books and classes as the practice of software development is from printf("Hello world!\n") programs.) Even though they seem like data cleaning procedures, the substitution procedures have to be a part of the design, as survey inference is with respect to randomness in who's in the sample, not with respect to the observed values (which is what frequentist statistics $x_i \sim \mbox{i.i.d.} f(x,\theta)$ teaches you).

Some references include:

The topic is more of an art than science; several other books I was honestly expecting to at least mention it (Longford's Missing Data and SAE; Kim and Shao's Methods for Handling Incomplete Data; Handbook of Missing Data Methodology) have zero coverage.

My gut feeling is that substitution works best by combining design procedures with some sort of modeling of nonresponse propensities, but probably Nishimura's dissertation gives a better reasoned set of recommendations than my guts do.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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