Bridging the gap: Turning classical statistics experience directly into a working knowledge of survey data analysis
Sunday, May 29, 2016 from 9 am to 4 pm, lunch included — Thistle 255
Claude Girard, Methodological Researcher – Data Analysis Resource Center, Statistics Canada
Abstract
Researchers, like most statisticians, are trained in classical statistics, the realm of independent and identically distributed (i.i.d.) data. To fulfill their research proposals, they often turn to survey data collected by Statistics Canada as it is a rich source of quality information about various aspects of the Canadian society. However, such survey data is not i.i.d. so that commonly used classical analysis methods must be adjusted.
The goal of this one-day workshop is neither to teach nor to preach survey statistics, but rather to help users bridge the gap between classical and survey statistics in a meaningful and lasting way. The intent is for researchers who are already comfortable with classical statistics to carry out their statistical research programs using survey data with the same level of confidence (!).
To achieve this, we convert key notions of classical statistics into their survey counterparts hereby providing researchers with a working knowledge of survey principles. Key issues related to survey data analysis will be addressed by drawing parallels with what researchers already know from classical statistics. For example, an analyst who is already familiar with the classical Bootstrap procedure by Bradley Efron will gain the insights required to understand and make proper use of the Rao-Wu Rescaling Bootstrap, a survey-adapted variant used by Statistics Canada. In the same way, researchers accustomed to building confidence intervals using the Normality assumption relying on the Central Limit Theorem to justify its use will better understand their applicability and limitations in the context of finite population surveys. And, in the context of hypothesis testing, we will see that the Rao-Scott corrected Chi-square statistic is obtained from the usual Chi-square statistic to account for survey data not meeting the i.i.d. assumption, much like the t-statistic is derived from the z-statistic to address the issue of small sample sizes.