1
$\begingroup$

We did customer survey in last fall from which we know how different stratas answer to survey questions. We did notice that amongst many strata having younger people who move often the non-response rate was much higher than amongst more established of our customers.

Original survey was based on simple random sampling without replacement amongst strate and sample sizes strata-wise were based on proportional allocation.

Should we use this information next time when we do customer survey? Should we allocate more for those strata which have more non-response. We could perhaps use some re-weighting to correct marginal distributions afterwards.

Could this be reasonable approach?

We have not done customer surveys earlier ourselves and previous survey was financed together with our industry association and we got only summary statistics. And survey was done by outside consultant firm. So we do not have any technical reports what they had actually done in that survey.

$\endgroup$
1
$\begingroup$

The problem induced by non-response is bias due to differences in responders and non-responders. Adding extra observations to strata with higher non-response will not reduce the bias, so that approach cannot be recommended. However, adding extra resources, such as more personalized follow-up, is a reasonable strategy. I would not apply the resources to everyone in the sample of the low-response strata, but to a sample of the non-responders, no matter what strata they come from. This technique is known as two-phase, or "double", sampling.

The overall choice of a sampling scheme, including the effort devoted to each phase of a double-sampling plan is governed by an equation which relates desired precision to cost. See any good sampling text, e.g. Lohr, 2009.

There are also weighting solutions for non-response, for example response-cell weighting and inverse-probability-of response weighting (IPW). I'm traveling, but I believe that you will find good examples in Groves et al. (2009) as well as in Lohr (2009). Also, google "inverse probability weighting response" or related phrases.

References, both highly recommended:

Groves, Robert M., Floyd J. Fowler, Mick P. Couper, James M. Lepkowski, Eleanor Singer, and Roger Tourangeau. 2009. Survey methodology. Hoboken, N.J.: Wiley.

Lohr, Sharon L. 2009. Sampling: Design and Analysis. Boston, MA: Cengage Brooks/Cole.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Steve Samuels thanks for your answer. Adding some extra resources for trying to increase response rate seems good advice. We sent questionnaires and did not allocate people from our customer service for the added initiatives. Reason was of course cost saving. But perhaps this is something where costs should not be too easily cut... :) $\endgroup$ – Analyst Jan 19 '14 at 10:37
  • $\begingroup$ You are welcome, Analyst. If this answer is helpful, please "Accept" it. In general, high precision from a "big" sample is wasted in face of too much non-response. $\endgroup$ – Steve Samuels Jan 19 '14 at 17:14

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.