5. メタ分析
• 例えば…
• Mizumoto, A., Urano, K., &
Maeda, H. (ママ) (2014). A
systematic review of published
articles in ARELE 1-24: Focusing
on their themes, methods, and
outcomes. ARELE, 25, 33-48.
• 全体の効果量はチュー程度!
(g = 0.76, 95% CI [0.59, 0.93])
• 検定力不足多!29人以上で♡
6. メタ分析
• Plonsky’s list: 187 (September, 2014)
• oak.ucc.nau.edu/ldp3/bib_metaanalysis.html
• もはやブームですやん; 粗製濫造が心配…
META-ANALYSIS IN L2 RESEARCH 87
1
2 2
10
14
16
14
12
10
8
6
4
2
0
1996-1998 1999-2001 2002-2004 2005-2007 2008-in press
Year of meta-analysis
Number of meta-analyses
Fig. 1. Growth of meta-analysis in L2 research.
Oswald & Plonsky (2010, p. 87)
magnitudes and patterns of relationships as well as the circumstances that
affect them.
7. メタ・メタ分析
• 例えばポリンキー
• Plonsky & Brown (2014)
• Corrective feedback
に関するメタ分析13
もあって結果バラバ
ラ(d = -0.16 to 1.16)
なのワロス
13. Norris & Ortega (2000)
• Effectiveness of L2 instruction: A research
synthesis and quantitative meta-analysis.
Language Learning, 50, 417–528.
META-ANALYSIS IN L2 RESEARCH 87
1
2 2
10
14
16
14
12
10
8
6
4
2
0
1996-1998 1999-2001 2002-2004 2005-2007 2008-in press
Year of meta-analysis
Number of meta-analyses
Fig. 1. Growth of meta-analysis in L2 research.
Oswald & Plonsky (2010, p. 87)
magnitudes and patterns of relationships as well as the circumstances that
affect them.
The third problem with narrative review concerns the limitations of the re-viewers
14. Norris & Ortega (2000)
• 40研究の明示的指導(k = 71)
• d = 1.13 95% CI [0.93, 1.33]
• 19研究の暗示的指導(k = 29)
• d = 0.54 95% CI [0.26, 0.82]
• 本当に?どういう意味で?
• Watari (2013; 2014)
15. Norris & Ortega (2000)
• 38(40)研究の明示的指導(Explicit, k = 71)
• + Rule explanation (deductive/metalinguistic,
Exp. FonFs) (k = 47)
• d = 1.08 95% CI [0.80, 1.36]
• + direction to attend to forms and arrive at
rules (explicit induction) (Exp. FonF, k =24)
• d = 1.22 95% CI [0.91, 1.53]
16. In Norris & Ortega (2000) …
DeKeyser (1997, p. 208)
Exp. FonFs vs. Control
Artificial
Number/person/gender
Leow (1998, p. 55) MLJ
Exp. FonF vs. Exp. FonFs
Spanish
Number/person/gender
18. メタ分析の結果の妥当性の問題
(山田・井上, 2012, p. 18)
• Apples and oranges problem:
• 多様な研究がごちゃ混ぜばんばん!
• Garbage in, garbage out:
• はっはっはまるでゴミのような研究だ
• File drawer problem:
• はいはい出版された研究えらいえらい
19. そもそも…
• “When the outcome is
reported on a meaningful
scale and all studies in the
analysis use the same
scale, the meta-analysis
can be performed directly
on the raw difference in
means.” (Borenstein, et al.,
2009, Chapter 4, Section
2, para. 1)
32. Watari (2014)
watariyoichi.net
research and
on
the L2
types of
specific
because
mechanisms
a series
point.
sample
to be
Table 2. Types of comparison & timing of metalinguistic intervention
Before B&W While After None g (95% CI)
Focus on FormS Exp.
vs Control 10 9 20 0 3 [0.60, 1.13]
vs Focus on FormS Exp. 2 0 7 0 3 [0.05, 1.05]
vs Focus on Form Imp. 2 0 2 0 0 [-0.84, 0.85]
vs Focus on FormS Imp. 0 2 0 0 0 [-1.29, 1.02]
FonF Exp.
vs Control 16 2 4 0 9 [0.75, 1.37]
vs Focus on Form Imp. 1 0 11 0 0 [0.28, 1.24]
vs Focus on FormS Exp. 2 4 0 0 0 [0.22, 1.58]
vs Focus on Form Exp. 2 0 0 0 0 [-0.25, 2.20]
g (95% CI) [0.53, 1.11] [0.21, 0.99] [0.72, 1.22] N/A [0.25, 1.11] [0.66, 0.99]
Biased toward giving rule description deductively, so we can’t get the whole picture
unless we construct and include rule discovery kind of instruction, with
38. In’nami Koizumi (2009)
• When multiple effect sizes were
available in a single study, a weighted
mean of the means was calculated,
along with a weighted mean of the
SDs, … Thus, each study contributed a
single effect size. This weighted-average
procedure for means and SDs
is widely used to reduce the bias
caused by dependency between the
effect sizes in one study. … However,
this procedure might mix up some
potentially important variables that
the primary researcher independently
investigated; therefore, information
on moderator variables could be lost.
(p. 229-230)
42. Norris Ortega (2006)
• (b) 各研究の全ての効果量を含めるが、慎重
な記述に努め、それ以上の(因果関係の)推
測的な分析は避ける
• “Given the limitations of (a), and the technical
and primary study reporting demands of (c), we
would suggest that option (b) generally
provides the most broadly applicable strategy
for the basic interpretive demands of meta-analysis
in LLLT research.” (p. 30)