Formatting error with visual analogue scales in Web surveys
Frederik Funke & Ulf-Dietrich Reips
Paper presented at the 11th annual
General Online Research Tagung (GOR) conference of the
German Society for Online Research (D.G.O.F.)
April 6-8, 2009 in Vienna (Austria)
Errors can't be avoided in the assessment of data in (Web) surveys. There are many sources of error on survey statistics resulting in either biased or unbiased estimates. One possible classification (see Groves, Fowler, Couper, Lepkowski, Singer, & Tourangeau, 2004) is to distinguish between representation related sources of error and measurement related errors. With the help of another approach (Sudman, Bradburn, & Schwarz, 1996) we focus on one component of the measurement error, namely formatting. This kind of error happens when there is no optimal option on the rating scale.
Visual analogue scales (VASs) - in our study plain horizontal lines with both ends anchored - are very well suited for Web-based research: They allow fine gradation and differentiation of ratings on a closed-ended continuum. Data from VASs reach - at least with a young and educated student sample - the level of an interval scale (Reips & Funke, 2008). By now it was questionable if certain respondent characteristics had an influence on the quality of data obtained with VASs.
In a Web experiment 1910 respondents from a heterogeneous US sample were asked to locate 15 target values (percentages presented in randomized order) on VASs. We decided to use numbers because they are mentally very well represented. Thus, deviation from target values should be owed to formatting error only. We were able to replicate that data from VASs reach the level of an interval scale (Reips & Funke, 2008) even with a non-student sample. We examined precision of ratings, i.e. the absolute difference between target value and actual rating. Overall, we found a very small formatting error. The mean overall difference was at 1 percentage point. 82% of the ratings were in the interval +/- 2 percentage points. The quality of formatting judgments on VASs was neither affected by sex, age or education nor by Internet experience. We found a statistically significant difference that very fast, spontaneous ratings were marginally less precise, but the effect size was very low.
As the examined respondent characteristics did not show any statistically significant influence on formatting error, we are encouraged to use VASs for surveying samples of the general population.