How we research

Every figure this series draws from its own data comes from one source: a scored teacher competency assessment, published in two editions, 2022 and 2023.

• 26,544 teachers assessed across the two editions
• 69 countries represented among those teachers
• 10,562 schools represented among those teachers
• more than 1.2 million scored responses
• six CPAT domains, resolving into 25 sub-domains
• pooled mean competency 51%: 46.8% in the 2022 edition, 53.4% in the 2023 edition

Sub-domains were measured in the 2023 edition only, on a base of 754,971 attempted responses from 15,548 teachers. All 25 rows, with the correct-response rate and the base for each, are published as aggregate data on The dataset.

Results are reported as correct-response rates: the proportion of correct answers among attempted responses, at domain and sub-domain level. There is no index, no weighting and no composite score of our own construction sitting between the responses and the percentages we print. When a volume shows a figure, the arithmetic behind it is a division you could do yourself from the table on the dataset page.

The 2022 and 2023 editions share no common items. That single fact governs how their two numbers may be used. Mean competency was 46.8% in the 2022 edition and 53.4% in the 2023 edition; those are two separate measurements taken on two different item sets, not two points on a line.

The difference between them is not growth, not improvement and not a trend, and this site does not report it as one. Where a single figure across both editions is useful, we publish the pooled mean of 51% and say in the same sentence that it is pooled.

It measures knowledge-for-practice. A teacher sitting the assessment demonstrates what they know a teaching situation calls for. It does not measure enacted classroom behaviour. Nobody watched them teach.

So a sentence like “teachers scored 39.9% correct on learner behaviour management, on a base of 15,047 attempted responses in the 2023 edition” is a statement about knowledge. It is not a statement that four lessons in ten are well managed, and it must never be written as one. Read the other way round, a high score is not a guarantee that anything happens on Monday morning.

The distance between knowing and doing is a real research question. This dataset does not answer it and neither do we. Volume 4 is about that distance, and it argues from published literature (Joyce and Showers, 2002) rather than from our data, precisely because our data cannot reach it.

Participation was voluntary. Teachers chose to take the assessment; nobody was sampled into it.

The consequence is a population that skews motivated, graduate-qualified, toward India and toward women. It is not a probability sample of any national teaching workforce. No figure from it should be read as a national statistic, a country ranking or a system comparison, and this site publishes no country-level breakdowns for that reason. The 69 countries are where the assessed teachers came from; they are not 69 sampled frames.

This is the limitation we would most like to be able to remove and cannot. A voluntary assessment recruits the teachers most willing to sit an assessment. If self-selection moves these scores, the direction we would expect it to move them is upward, but we have not measured that, so it stays reasoning rather than a finding, and nothing in this series rests on it. What we can say is that the bias exists, that it is not quantified, and that it is a reason to read 51% as a description of the teachers who sat this assessment rather than a portrait of a workforce.

Two kinds of number appear in these volumes, and they are never merged into one.

Ours

Figures from the competency dataset described above. They are labelled as our own research and carry their base and their period in the same sentence as the number.

Somebody else’s

Figures from published literature. They are attributed to their original authors, on the page where they appear, in the form those authors published them. The external sources this series cites are:

• Nye, B., Konstantopoulos, S., & Hedges, L. V. (2004). How Large Are Teacher Effects? Educational Evaluation and Policy Analysis, 26(3), 237–257. Used in Volume 6.
• Joyce, B., & Showers, B. (2002). Student Achievement Through Staff Development, 3rd ed. ASCD. Used in Volume 4.
• Kraft, M. A., Blazar, D., & Hogan, D. (2018). The Effect of Teacher Coaching on Instruction and Achievement: A Meta-Analysis of the Causal Evidence. Review of Educational Research, 88(4), 547–588. 60 causal studies; 0.49 SD on instruction; 0.18 SD on achievement. We carry the authors’ caveat wherever we cite it: effects shrink as programmes scale, because coaching is capacity-bound.
• Stronge & Associates / TIE Online, citing Mancuso, Roberts & White (2010) and Desroches (2013). Used in Volume 3.
• ISC Research (2026). Global International Schools Snapshot 2026. Used in Volume 6.
• Hanushek, E. A., & Rivkin, S. G. (2012). The Distribution of Teacher Quality and Implications for Policy. Annual Review of Economics, 4, 131–157.
• Allen, J. P., Pianta, R. C., Gregory, A., Mikami, A. Y., & Lun, J. (2011). Science, 333(6045), 1034–1037.

We did not re-analyse any of these studies, pool them with our data, or restate their findings on a base other than the one their authors published. Where a volume sets an external number beside one of ours (as Volume 6 does with teacher effects and Volume 4 does with training transfer), both are labelled and the sentence says which is which.

The negative rule matters as much as the positive one. No external finding is ever presented as though it were ours, and no figure of ours is ever laundered through a citation to make it look external.

1. One statistic, one self-contained sentence, carrying the metric, the population, the base and the period. “Across 26,544 teachers assessed in 69 countries in 2022–2023, mean competency was 51%”, not “most teachers score poorly”.
2. One number, one phrasing, everywhere. If a figure is 51% in a summary it is 51% in the body, never “just over half”, and never rounded into a different number on a different page.
3. Absolute dates only. “July 2026”, never “recently” and never “three months ago”.
4. No base, no number. If we cannot show what a figure was computed on, we write the sentence without it or leave the section out.

There is a hard line on this site between the research and the product, and this is where it is drawn.

Everything we say about what our platform would do for a school, a group, a system or a training provider is an evidence-supported projection: reasoning outward from the published findings above to an expected outcome. It is not a measured result from a deployment. Where the reasoning leans on somebody else’s evidence (for instance Kraft, Blazar and Hogan’s 2018 meta-analysis of 60 causal studies, reporting 0.49 SD on instruction and 0.18 SD on achievement), that evidence is cited, and so is the caveat its authors attached to it.

We have published no product efficacy results, no adoption numbers and no student-achievement claim at any ESSA tier, because those studies have not been done. Nothing on this site is described as proven, and no case study is attributed to an unnamed school. When there is an efficacy study, it will publish with its method and its limits in the same place as its result, and the difference between projected and measured will be legible on the page rather than buried in a footnote.

That is stated here, once, in full, and it is not repeated elsewhere on this site. Every other page that needs it links back to this note rather than restating it.