Results of statistical calculations can vary depending on the sample size, comparability of subsamples, choice of statistical methods, coding of variables, use of imputation, and many other factors. It is up to the researcher whether the Bonferroni adjustment of p-values, or Bonferroni-Holm adjustment, or no adjustment at all is used, whether he or she utilizes parametrical or more conservative non-parametrical statistical methods, whether metrical data are z-transformed etc. This presentation shows some examples of statistical pitfalls in large-scale language assessment programs. First, data from German language screening programs are utilized to show why children with frequent otitis media seem to acquire German more quickly than children without frequent otitis media (N = 1,628). Another example shows why children born in January-July speak better German than those born in August-December (N = 4,316) and how one can reverse these results on the basis of another sample (the same age, the same region, the same language test; N = 6,144). One more example demonstrates how a link between stuttering and children’s German language competence can be described as very close or non-existent depending on chosen statistical methods (N = 746). A high variability of results, in spite of very large samples, validated and standardized language tests, high quality of data in terms of the inter-rater and intra-rater reliability, can be traced back either to errors in the study design or to unexpected confounding variables. For instance, children with frequent otitis media acquired German under comparatively favourable sociodemographic conditions.

Authors:
Eugen Zaretsky, Marburg University, Germany

About the Presenter(s)
Dr. Eugen Zaretsky is currently a researcher (post-doc) at the Department of Phoniatrics and Pediatric Audiology, University of Marburg

See this presentation on the full schedule – On Demand Schedule

Bookmark