Perhaps someday we will learn how to ask perfect survey questions that yield perfect answers. Until such times arrive, however, respondents' answers to survey question will typically reflect the opinion or fact we as researchers are after only partly. The other part will be misremembering, differences in interpretation, differences in how arbitrary choices in the answering process are made, mistakes, and so on - in short: measurement error.
Measurement error can seriously disturb analyses of substantive interest. Means, totals, and proportions will be off if the average answer people give is inaccurate. However, measurement error disturbs not just estimates of means but can also severely bias apparent relationships, conditional probabilities, means differences, and other regression-type analyses. To remove such biases it is therefore essential to estimate the extent of measurement error in survey variables.
The most obvious way to estimate the extent of measurement error is to know the true value we are after. For example, survey methodologists often use “gold-standard” data from administrative registers to validate respondents’ survey answers. But not all that is administrative data is gold: often such records contain measurement error themselves, or do not fully reflect the concept of actual interest. Moreover, there are many survey variables for which true values are unavailable or impossible to get. Opinions are a good example, but facts such as the party a respondent votes for in elections may also be unknown outside of the survey answer.
This 1-credit/2 ECTS course introduces you to the main alternative solution to measurement error in surveys: statistical modeling. You will be introduced to the three main competencies in this field:
We will have four sessions in which you will watch online lectures, do homework exercises using R, and answer online quizzes.
By the end of the course, students will be able to…