Statistician for education research: multilevel models for students nested in classrooms, quasi-experimental designs, Likert scales and rubric reliability. Theses and Q1 journals.
Statistics for education research that knows your pupils sit inside classrooms
Theses, dissertations and journal articles in education, subject didactics and teacher training. Classroom interventions with few groups, multilevel models for pupils nested in classes and schools, attitude and motivation scales, rubric reliability and effect sizes that mean something to a school.
Nearly every problem in the field comes from the same two facts: you cannot randomize inside a school, and pupils are not independent observations. The earlier that is looked at, the cheaper it is.
You taught your approach in one class and used another as control. No randomization, groups differing at baseline, and a reviewer telling you the comparison does not hold. It is fixable, but not with a t-test.
Your n looks large and your real power is not: the unit you assigned was the classroom, not the pupil. Treating them as independent is the single most common reason education papers get rejected on method.
Motivation, attitude toward math, classroom climate, teacher satisfaction: what can be treated as continuous, what needs non-parametric methods, and how to defend that choice to a reviewer.
Intraclass correlation to quantify how much the classroom matters, two- and three-level models, and random slopes when your intervention does not work equally well in every class.
ANCOVA with the pretest as covariate, propensity score matching and interrupted time series: what you do when the school will not let you randomize pupils, which is always.
Cohen's d with 95% CI, translated into what a staff meeting actually asks about: months of additional progress and percentile gains, and why d = 0.40 is already meaningful in education.
Exploratory and confirmatory factor analysis, reliability (alpha and omega), measurement invariance across year groups, stages or sexes, and norms if your instrument will be used to classify pupils.
How closely two or more teachers agree when scoring with the same rubric: ICC for continuous scores, kappa and Krippendorff's alpha for categories, and what to do when agreement comes out low.
How many classes you need, not how many pupils, given the design effect that clustering imposes. It is the calculation that saves a study before it starts.
No clock running at your expense. We agree the scope, the deadlines and a fixed price up front. How long it takes me is my problem: no inflated hours, no invoice larger than the one agreed. Before delivery every analysis goes through a professional review: I re-run the code, check assumptions and go through tables and figures line by line. A dissertation with an intervention in two classrooms and a multilevel thesis across twenty schools both fit this model: the scope and the price change, not the way I work.
Illustrative example of the commonest design in education: two intact groups measured before and after. The honest comparison is not between the final means, but adjusting for the starting level and accounting for the classroom being the unit assigned.
Free resources on the analyses that come up most in education theses and papers.
An intervention across two classrooms, a motivation scale, a rubric with several markers or a thesis spanning twenty schools. Fill in the form and I will reply within 24h with an indicative proposal.
Biostatistics and statistical consulting for health research · All services
You can, but not as though it were an experiment. With two intact groups, two things are expected of you: controlling for the starting level (ANCOVA with the pretest as covariate, not raw gain scores) and acknowledging that the unit assigned was the classroom, not the pupil. And there is an honest limit: with one class per condition, the effect of the intervention and the effect of the teacher cannot be separated. That does not invalidate the study, but it belongs in your limitations, and saying it before the reviewer does works in your favor.
Two pupils in the same class resemble each other more than two pupils from different classes: same teacher, same timetable, same peers, same neighborhood. That resemblance means each pupil carries less information than an independent observation would, and if the analysis ignores it, standard errors come out too small and significant effects appear that are not there. Multilevel models handle it. The first thing to look at is the intraclass correlation: if the classroom explains an appreciable share of the variance, this is not optional.
More than most people expect, and it is the question to ask BEFORE collecting data. When you assign classrooms, what drives power is the number of classrooms, not the number of pupils: adding more children to classes you already have barely helps. You can get a first estimate yourself with the clinical sample size calculator on this site, which includes the cluster design effect.
It depends on what you are averaging. A single 1-to-5 item is ordinal and is best treated as such. A total score from a validated scale, with several items and a checked factor structure, is routinely treated as continuous in the education literature. The choice gets justified in the manuscript rather than assumed, and I leave that justification written for you.
Yes, and it is worth flagging that they are not ordinary spreadsheets: they carry sampling weights and plausible values, and ignoring either gives you both the wrong means and the wrong standard errors. They need the analysis tools built for that design. It is different work from a questionnaire of your own, so tell me at the start.
Increasingly, yes. Journals and evidence standards such as the What Works Clearinghouse expect clustering to be accounted for, effect sizes reported with confidence intervals, and baseline equivalence addressed when assignment was not random. Meeting those three is most of the difference between a desk rejection and a review.