The SAS/STAT survey analysis procedures assume missing values are missing completely at random. Nonresponses are assumed to be ignorable. For this reason, observations with missing values are excluded or deleted prior to the analysis.

Missing values, or nonresponses, are common in sample survey data. Nonresponse can be viewed either as ignorable or as nonignorable for the mean estimates. For total estimates, nonresponse will introduce bias unless appropriate actions are taken, such as reweighting or imputation.

Ignorable nonresponse

The missing values are missing at random for reasons unrelated to the (unobserved) true values. Whether or not a unit responds is independent of the values of analysis and demographic variables. If you were to take another random sample of the same population that happened to include some of the same sample points, the values would not necessarily be missing again. This is a special case of the stochastic view of nonresponse (Lessler and Kalsbeek, 1992).

Nonignorable nonresponse

The missing values are not missing at random. That is, the reason the value is missing is related to the (unobserved) true value of the variables of interest and related demographic variables. If you were to take another random sample of the same population that happened to include some of the same sample points, the values would be missing again. This is also called the deterministic view of nonresponse.

When the nonresponse is ignorable, simply excluding observations containing such values from the analysis is the optimal procedure. The mean estimates, as well as the estimates of their variances, are unbiased and as accurate as possible based on the available data. When observations with missing values are deleted, the missing at random mechanism is implicitly adopted.

However, when the nonresponse is nonignorable, an analysis based only on the respondents in the sample can cause biased estimates for the study population. Under these circumstances, various techniques are available for adjusting for the undesirable effects of this nonignorable nonsampling error. Commonly used methods include weight adjustment, regression adjustment, and imputation. If your survey exhibits nonignorable nonresponse, many survey experts advise these or other remedies in order to address the problem before analyzing the data. See the references for more information.

Another way to handle nonignorable nonresponse is to change the objective of the study to focus only on the subset of the entire population comprised of the respondent domain. This leads to the domain analysis approach to handling nonresponse which is implemented by default in SUDAAN software and, beginning in SAS 9.2, with the NOMCAR option in the SAS survey procedures.

Reference

Brick, J.M. and Kalton, G. (1996), "Handling Missing Data in Survey Research," Statistical Methods in Medical Research, 5, 215-238.

Cochran, W. G. (1977), Sampling Techniques, Third Edition, New York: John Wiley & Sons, Inc., Chapter 5.

Kalton, G. and Kaspyzyk, D. (1986), "The Treatment of Missing Survey Data," Survey Methodology, 12, 1-16.

Kish, L. (1965), Survey Sampling, New York: John Wiley & Sons, Inc., Chapter 13.

Lessler and Kalsbeek, (1992), Nonsampling Error in Surveys, New York: John Wiley & Sons, Inc., Chapter 7.

Lohr, (1999), Sampling: Design and Analysis, Pacific Grove: Duxbury Press, Chapter 8.