Summary: | 博士 === 國立中山大學 === 教育研究所 === 107 === Differential Item Function (DIF) plays an important role in testing quality in terms
of fairness and validity. In the literature, methods of DIF assessment have been well
developed. The further researches would focus on the issues about the sources of DIF.
Recently, several qualitative and quantitative approaches for defining DIF sources were
proposed. In the qualitative approach, expert review is one of the most popular method.
In the quantitative approach, the potential DIF source is incorporated into the statistical
model, such as Logistic Regression model, Mixture model, or Mediated MIMIC model,
and is assessed in the statistical model simultaneously. However, this approach is limited
due to the requirement of pre-specified DIF sources which have to be collected in advance.
Therefore, a more flexible method for defining the structure of DIF sources without any
previous information is necessary. The purposes of this study are (a) developing a
procedure to define the sources of DIF (b) investigating independent variables which may
affect the efficacy of this procedure, and (c) demonstrating the procedure on an empirical
data and explaining the sources of DIF. Two simulation studies and an empirical data
analysis were conducted. The performance of the proposed procedure on each
independent variable in both simulation studies was evaluated with Hit Rate (HR) and
Per-Element Accurate (PEA). The results indicated that this procedure provided almost
perfect HR and PEA in most conditions. In addition, the performance of DIF assessment
would positively affect the HR and PEA of defining DIF source. In the analysis of 2011
TIMSS math, the procedure defined the population pyramid content as the DIF source in
2 DIF items. In summary, the procedure can define the DIF source without advance data
collection which previous quantitative approaches did. This study recommends
researchers to apply this procedure to define the DIF source when DIF items are identified
but the DIF source is not obvious.
|