ROBUST STATISTICS IN MARKET RESEARCH AND VISION SCIENCE

‚Äč

The analyses and data-transformations were conducted on Excel and SPSS

  • Bootstrapping: sampling with and without replacement from normally distributed data in order to mitigate the effect of outliers and to add statistical power to better show existing effects.

    Bootstrapping either 5, 10 or 20 thousand times is done when analyzing the relationships between the variables in the mediator and moderator models in Marketing (see the Mediation and Moderation section, under the "Data Analysis" section).

    Bootstrapping was also used in the Vision Science collaboration with the colleagues from Ludwig Maximilian University of Munich

Puntiroli, Deubel & Szinte

  • Winsorizing 10% or 20%: Winsorizing the extreme tails of the distribution entails substituting all values on the left tail of the distribution (only the initial 10 percentiles) with the value that corresponds to the 10th percentile (i.e. 10% Winsorizing), and all values on the right tail of the distribution (from the 90th percentile onward) with the value corresponding to the 90th percentile. This technique mitigates the effect of outliers and is particularly useful in surveys when respondents are free to enter any value they like, or in reaction time studies.

    Winsorizing was highly advocated during my Masters in Quantitative Research Methods in Glasgow University, following the teachings of Rand Wilcox "Introduction to Robust Estimation and Hypothesis Testing"

    Winsorizing is often carried out on Expenditure values and Consumption Values, as part of analyses conducted for the "Competence Center for Research in Energy, Society and Transition"

    https://www.sccer-crest.ch/research/swiss-household-energy-demand-survey-sheds/

  • Arcsine Data Transformation: this is a rather standard technique when working with proportional data. In projects employing a range of percentages, which usually go from 50% (chance level) to 100% (the upper ceiling) arcsine transformation essentially makes these boundaries more fuzzy, simulating a scale or range that does not have these fixed limits.

    This type of transformation was performed on the data in the following Vision Science Project.

Puntiroli, Kerzel & Born