When conducting a survey the medium used for interacting with respondents and collecting data is called the survey mode. Modes are associated with survey mode effects such as measurement error due to social desirability bias in face-to-face interviews or non-response bias caused by low response rates in web surveys.

Groves & Lyberg (2010) write that a notable omission in total survey error models is parameters that describe the relationship between multiple error sources, for instance linking non-response and measurement error.

This presentation targets this issue by applying the principal stratification framework for causal inference (Frangakis & Rubin 2002). It will be shown that this framework is useful for relating different error sources caused by modes.

Also, it will be demonstrated how it can be used to design mixed-mode surveys that achieve the lowest bias given a budget constraint. Assumptions such as monotonicity, exclusion restriction and “constant pure mode effect” are explored. Large-sample bounds are derived which enables a sensitivity analysis of the assumptions.

Some references

Frangakis, C. and Rubin, D. (2002). Principal Stratification in Causal Inference. Biometrics. Vol. 58, No. 1, pp. 21-29.
Groves, R. and Lyberg, L. (2010). Total Survey Error. Past, Present, and Future. Public Opinion Quarterly, Vol. 74, No. 5, pp. 849–879.