230 likes | 414 Views
Methodological Issues in Systematic Reviews in Education. Robert E. Slavin Institute for Effective Education University of York. Sample Size and Effect Size. Negative correlation noted in other fields Reasons: Underpowered studies with null results disappear
E N D
Methodological Issues in Systematic Reviews in Education Robert E. Slavin Institute for Effective Education University of York
Sample Size and Effect Size • Negative correlation noted in other fields • Reasons: • Underpowered studies with null results disappear • Small studies of lower methodological quality • Superrealization bias • Measures aligned with treatments
Present Studies: Best Evidence Encyclopedia • Elementary and secondary math • 185 qualifying studies • Studies with inherent measures, brief durations, big pretest differences excluded
Table 1 Total Sample Size RecodeRangeNumber of Studies 1 Up to 50 10 2 51-100 36 3 101-150 18 4 151-250 31 5 251-400 14 6 401-1000 41 7 1001-2000 12 8 2001 or more 23 TOTAL185
Notable Findings • Overall correlation: -.28, p<.001 • Sample sizes ≤100: ES= +0.40 • Sample sizes > 2000: ES= +0.09 • Random: ES=+0.24 • Randomized quasi-experiments: ES=+0.29 • Matched: ES= +0.17 • Difference disappears when sample size considered
Possible Solutions • Weight by sample size • Require minimum sample size for high ratings • BEE requires 500 students in 2+ studies
Implications • Results from large studies should be preferred, all else being equal • Such results tend to be modest. We should look for outcomes of +0.20 to +0.30, at best
Treatment-Inherent Measures • Experimenter-made • Assess outcomes emphasized in experimental but not control group
Treatment-Independent Measures • Usually standardized tests • May be experimenter-made if experimental and control groups received the same content
How Do Program Effectiveness Reviews Treat Inherent Measures? • What Works Clearinghouse includes • Best Evidence Encyclopedia excludes
Curriculum vs. Instruction • Legitimate need to measure and report outcomes emphasized in experimental group • But, potential bias introduced if inherent measures averaged with independent measures • How much bias?
Implications • Treatment-inherent measures must be excluded from reviews, or at least reported separately • Clear distinction between inherent and independent measures can be made
Implications in Light of Findings on Sample Size and Duration • Random assignment cannot be the only criterion of evaluation excellence • Effect sizes from large, extended studies of school and classroom interventions with independent measures are modest (+0.20 to +0.30 at best). These are the effects we should be looking for.