Teaching and learning floating and sinking: A meta‐analysis

69 intervention studies, 191 effect sizes, 10,532 participants

Learners taught floating and sinking (density/buoyancy) between 1977 and 2021, preschool to undergraduate; 42% of samples (4,845 participants) had a mean age between 8 and 12. Studies in English, German and Greek.

A single-misconception meta-analysis, and therefore unusually informative - floating and sinking is the canonical naive-physics misconception (things float because they are light / have air inside) and this pools five decades of attempts to teach it away. Three-level hierarchical meta-regression handles multiple effect sizes per study. The critical design fact is that 61 of 84 design-codable studies are PRE-POST with no control group; only 23 studies (48 effect sizes) have a treatment-control comparison, and those give a smaller estimate. Read in full. Its unique contribution to the durability question is that it models the delay between intervention and test as a CONTINUOUS moderator, which nobody else in this literature does.

Mean effect of teaching floating and sinking g = 0.85, 95% CI [0.71, 0.99], across 191 effect sizes from 69 studies and 10,532 participants. THE DESIGN DISCOUNT: pre-post designs g = 0.89 (61 studies, 143 effect sizes) versus treatment-control designs g = 0.72 (23 studies, 48 effect sizes), F(1,189) = 6.46, p < .01, difference g = 0.17 - and in the multiple moderator model the design effect disappears entirely, leaving only duration and hands-on. THE DURABILITY RESULT: the delay between intervention and test, modelled continuously in days across 60 studies and 170 effect sizes, has a regression coefficient of 0.00 (95% CI 0.00 to 0.00), p = 0.11 - no measurable decay over the delays actually used. THE MEASURE RESULT IS THE MOST USEFUL THING HERE AND IT CUTS AGAINST THE STANDARD INFLATION STORY: test format did not moderate at all (multiple choice 0.95, running experiments 1.06, interview 0.87, open-ended 0.70, mixed 0.66; F(4,186) = 1.74, p = 0.14), and tests explicitly built around students preconceptions gave a SMALLER effect (g = 0.75) than tests that ignored preconceptions (g = 0.90), F(1,189) = 0.93, p = 0.34 - i.e. within a single well-defined topic, misconception-targeted instruments are not the ones producing the big numbers. What does moderate: hands-on experiments g = 0.96 versus virtual experiments g = 0.52, F(1,189) = 8.02, p < .001; and each additional hour of instruction adds 0.06 (95% CI 0.02-0.09). Age does NOT moderate (beta = 0.01, p = 0.61 across ages studied) - the concept is as teachable at 8 as at 18. No publication bias (Egger F(1,189) = 0.15, p = 0.70). Dose warning the authors make explicit: the well-known Hardy et al. 2006 study ran 11 periods of 45 minutes, "about half of all science lessons in a semester".

Low for the treatment-control subset; the pre-post majority has no arm structure at all and its estimate is a within-child gain, not a causal contrast.

First meta-analysis on this topic. Its treatment-control estimate (0.72) sits between Schroeder & Kucera 2022 (0.41, text only) and Pacaci et al. 2024 (1.10 raw, 0.64 randomised only), which is what a whole-unit intervention on one hard concept should look like.