Sunday, February 20, 2011

Total Survey Error – Past, Present and Future


 

Public Opinion Quarterly Vol 74, No 5, pp 849 – 879

Robert M Groves , Lars Lyberg

 

-          TSE – concept that purports to describe statistical properties of survey estimates, incorporating a variety of error sources

-          Among set of alternative designs, design that gives smallest total survey error (for given fixed cost) should be chosen

-          Theory laid out by Neyman apply only when non sampling errors are small

-          Sampling variance is measurable in most probability sample surveys, but other components of the notion (of TSE) cannot be measured directly without significant alteration of the typical survey designs

-          Survey design components

o   Identify population

o   Describe sample

o   Access responding units among the sample

o   Operationalise constructs that are targets of measurement

o   Obtain responses to the measurements

o   Summarise data

o   For estimating some stated population parameter

-          1950s – most textbooks treating surveys were sampling texts

-          Sampling errors are inherently errors of non-observation

-          Relate process quality with total survey error

-          Groves et al 2004

o   Two separate inferential steps required in surveys

§  First inference is from response to a question for a single respondent and the underlying construct of interest to the measurement

§  Second inference is from estimate based on set of respondents to target population

-          Development of typologies continues

-          TSE

o   Nested taxonomy of concepts of error

§  Variance & bias

§  Errors of observation and non-observation

-          Construct validity

o   If one were interested in a child's intelligence, we would not define the true value as the score a teacher would assign on a particular day, but rather a more permanent attribute separate from measurement itself.

-          MSE = sampling variance + response variance + covariance of the response and sampling deviations + the squared bias

-          Re interview approach

-          Interpenetration and re interviews are necessary for measurement of response variability components

-          Multiple indicators

-          TSE not the only way to think about information quality

-          Relevance and credibility

-          Survey designer faces problem of how much budget to spend on measuring quality versus on other things

-          Growing evidence that measurement error models are highly variable by type of measurement.

-          Value of decomposition of errors

o   Has worked

o   By isolating measurement error sources, have learnt how to construct better questionnaires

o   By separating types of non response, we are learning to attack non contact and refusal separately, given their different covariances with many survey variables

-          Interplay of different error sources

o   Respondents with lower response propensities tend to have higher likelihood of measurement biases.

-          How a multi mode, multi frame multi phase world may help us

o   Offer built in contrasts that can be informative about error properties of survey statistics

No comments: