Can small-scale experimental investigations of teaching carried-out in a couple of arbitrary classrooms really tells us anything about how to teach well?
Keith S. Taber
Undertaking valid educational experiments involves (often, insurmountable) challenges, but perhaps this grid (shown larger below) might be useful for researchers who do want to do genuinely informative experimental studies into teaching?
Applying experimental method to educational questions is a bit like trying to use a precision jeweller's screwdriver to open a tin of paint: you may get the tin open eventually, but you will probably have deformed the tool in the process whilst making something of a mess of the job.
In recent years I seem to have developed something of a religious fervour about educational research studies of the kind that claim to be experimental evaluations of pedagogies, classroom practices, teaching resources, and the like. I think this all started when, having previously largely undertaken interpretive studies (for example, interviewing learners to find out what they knew and understood about science topics) I became part of a team looking to develop, and experimentally evaluate, classroom pedagogy (i.e., the epiSTEMe project).
As a former school science teacher, I had taught learners about the basis of experimental method (e.g., control of variables) and I had read quite a number of educational research studies based on 'experiments', so I was pretty familiar with the challenges of doing experiments in education. But being part of a project which looked to actually carry out such a study made a real impact on me in this regard. Well, that should not be surprising: there is a difference between watching the European Cup Final on the TV, and actually playing in the match, just as reading a review of a concert in the music press is not going to impact you as much as being on stage performing.
Let me be quite clear: the experimental method is of supreme value in the natural sciences; and, even if not all natural science proceeds that way, it deserves to be an important focus of the science curriculum. Even in science, the experimental strategy has its limitations. 1 But experiment is without doubt a precious and powerful tool in physics and chemistry that has helped us learn a great deal about the natural world. (In biology, too, but even here there are additional complications due to the variations within populations of individuals of a single 'kind'.)
But transferring experimental method from the laboratory to the classroom to test hypotheses about teaching is far from straightforward. Most of the published experimental studies drawing conclusions about matters such as effective pedagogy, need to be read with substantive and sometimes extensive provisos and caveats; and many of them are simply invalid – they are bad experiments (Taber, 2019). 2
The experiment is a tool that has been designed, and refined, to help us answer questions when:
- we are dealing with non-sentient entities that are indifferent to outcomes;
- we are investigating samples or specimens of natural kinds;
- we can identify all the relevant variables;
- we can measure the variables of interest;
- we can control all other variables which could have an effect;
These points simply do not usually apply to classrooms and other learning contexts. 3 (This is clearly so, even if educational researchers often either do not appreciate these differences, or simply pretend they can ignore them.)
Applying experimental method to educational questions is a bit like trying to use a precision jeweller's screwdriver to open a tin of paint: you may get the tin open eventually, but you will probably have deformed the tool in the process whilst making something of a mess of the job.
The reason why experiments are to be preferred to interpretive ('qualitative') studies is that supposedly experiments can lead to definite conclusions (by testing hypotheses), whereas studies that rely on the interpretation of data (such as classroom observations, interviews, analysis of classroom talk, etc.) are at best suggestive. This would be a fair point when an experimental study genuinely met the control-of-variables requirements for being a true experiment – although often, even then, to draw generalisable conclusions that apply to a wide population one has to be confident one is working with a random or representatives sample, and use inferential statistics which can only offer a probabilistic conclusion.
My creed…researchers should prefer to undertake competent work
My proselytising about this issue, is based on having come to think that:
- most educational experiments do not fully control relevant variables, so are invalid;
- educational experiments are usually subject to expectancy effects that can influence outcomes;
- many (perhaps most) educational experiments have too few independent units of analysis to allow the valid use of inferential statistics;
- most large-scale educational experiments can not assure that samples are fully representative of populations, so strictly cannot be generalised;
- many experiments are rhetorical studies that deliberately compare a condition (supposedly being tested but actually) assumed to be effective with a teaching condition known to fall short of good teaching practice;
- an invalid experiment tells us nothing that we can rely upon;
- a detailed case study of a learning context which offers rich description of teaching and learning potentially offers useful insights;
- given a choice between undertaking a competent study of a kind that can offer useful insights, and undertaking a bad experiment which cannot provide valid conclusions, researchers should prefer to undertake competent work;
- what makes work scientific is not the choice of methodology per se, but the adoption of a design that fits the research constraints and offers a genuine opportunity for useful learning.
However, experiments seem very popular in education, and often seem to be the methodology of choice for researchers into pedagogy in science education.
Read: Why do natural scientists tend to make poor social scientists?
This fondness of experiments will no doubt continue, so here are some thoughts on how to best draw useful implications from them.
A guide to using experiments to inform education
It seems there are two very important dimensions that can be used to characterise experimental research into teaching – relating to the scale and focus of the research.
Two dimensions used to characterise experimental studies of teaching
Scale of studies
A large-scale study has a large number 'units of analysis'. So, for example, if the research was testing out the value of using, say, augmented reality in teaching about predator-prey relationships, then in such a study there would need to be a large number of teaching-learning 'units' in the augmented learning condition and a similarly large number of teaching-learning 'units' in the comparison condition. What a unit actually is would vary from study to study. Here a unit might be a sequence of three lessons where a teacher teaches the topic to a class of 15-16 year-old learners (either with, or without, the use of augmented reality).
For units of analysis to be analysed statistically they need to be independent from each other – so different students learning together from the same teacher in the same classroom at the same time are clearly not learning independently of each other. (This seems obvious – but in many published studies this inconvenient fact is ignored as it is 'unhelpful' if researchers wish to use inferential statistics but are only working with a small number of classes. 4)
Read about units of analysis in research
So, a study which compared teaching and learning in two intact classes can usually only be considered to have one unit of analysis in each condition (making statistical tests completely irrelevant 5, thought this does not stop them often being applied anyway). There are a great many small scale studies in the literature where there are only one or a few units in each condition.
Focus of study
The other dimension shown in the figure concerns the focus of a study. By the focus, I mean whether the researchers are interested in teaching and learning in some specific local context, or want to find out about some general population.
Read about what is meant by population in research
Studies may be carried out in a very specific context (e.g., one school; one university programme) or across a wide range of contexts. That seems to simply relate to the scale of the study, just discussed. But by focus I mean whether the research question of interest concerns just a particular teaching and learning context (which may be quite appropriate when practitioner-researchers explore their own professional contexts, for exmample), or is meant to help us learn about a more general situation.
| local focus | general focus |
| Why does school X get such outstanding science examination scores? | Is there a relationship between teaching pedagogy employed and science examination results in English schools? |
| Will jig-saw learning be a productive way to teach my A level class about the properties of the transition elements? | Is jig-saw learning an effective pedagogy for use in A level chemistry classes? |
If that seems a subtle distinction between two quite similar dimensions then it is worth noting that the research literature contains a great many studies that take place in one context (small-scale studies) but which claim (implicitly or explicitly) to be of general relevance. So, many authors, peer reviewers, and editors clearly seem think one can generalise from such small scale studies.
Generalisation
Generalisation is the ability to draw general conclusions from specific instances. Natural science does this all the time. If this sample of table salt has the formula NaCl, then all samples of table salt do; if the resistance of this copper wire goes up when the wire is heated the same will be found with other specimens as well. This usually works well when dealing with things we think are 'natural kinds' – that is where all the examples (all samples of NaCl, all pure copper wires) have the same essence.
Read about generalisation in research
Education deals with teachers, classes, lessons, schools…social kinds that lack that kind of equivalence across examples. You can swap any two electrons in a structure and it will make absolutely no difference. Does any one think you can swap the teachers between two classes and safely assume it will not have an effect?
So, by focus I mean whether the point of the research is to find out about the research context in its own right (context-directed research) or to learn something that applies to a general category of phenomena (theory-directed research).
These two dimensions, then, lead to a model with four quadrants.
Large-scale research to learn about the general case
In the top-right quadrant is research which focuses on the general situation and is larger-scale. In principle 6 this type of research can address a question such as 'is this pedagogy (teaching resource, etc.) generally effective in this population', as long as
- the samples are representative of the wider population of interest, and
- those sampled are randomly assigned to conditions, and
- the number of units supports statistical analysis.
The slight of hand employed in many studies is to select a convenience sample (two classes of thirteen years old students at my local school) yet to claim the research is about, and so offers conclusions about, a wider population (thirteen year learners).
Read about some examples of samples used to investigate populations
Even when a population is properly sampled, it is important not to assume that something which has been found to be generally effective in a population will be effective throughout the population. Schools, classes, courses, learners, topics, etc. vary. If it has been found that, say, teaching the reactivity series through enquiry generally works in the population of English classes of 13-14 year students, then a teacher of an English class of 13-14 year students might sensibly think this is an approach to adopt, but cannot assume it will be effective in her classroom, with a particular group of students.
To implement something that has been shown to generally work might be considered research-based teaching, as long as the approach is dropped or modified if indications are it is not proving effective in this particular context. That is, there is nothing (please note, UK Department for Education, and Ofsted) 'research-based' about continuing with a recommended approach in the face of direct empirical evidence that it is not working in your classroom.
Large-scale research to learn about the range of effectiveness
However, even large-scale studies where there are genuinely sufficient units of analysis for statistical analysis may not logically support the kinds of generalisation in the top-right quadrant. For that, researchers needs either a random sampling of the full population (seldom viable given people and institutions must have a choice to participate or not 7), or a sample which is known to be representative of the population in terms of the relevant characteristics – which means knowing a lot about
- (i) the population,
- (ii) the sample, and
- (ii) which variables might be relevant!
Imagine you wanted to undertake a survey of physics teachers in some national context, and you knew you could not reach all that population so you needed to survey a sample. How could you possibly know that the teachers in your sample were representative of the wider population on whatever variables might potentially be pertinent to the survey (level of qualification?; years of experience?; degree subject?; type of school/college taught in?; gender?…)
But perhaps a large scale study that attracts a diverse enough sample may still be very useful if it collects sufficient data about the individual units of analysis, and so can begin to look at patterns in how specific local conditions relate to teaching effectiveness. That is, even if the sample cannot be considered representative enough for statistical generalisation to the population, such a study might be a be to offer some insights into whether an approach seems to work well in mixed-ability classes, or top sets, or girls' schools, or in areas of high social deprivation, or…
In practice, there are very few experimental research studies which are large-scale, in the sense of having enough different teachers/classes as units of analysis to sit in either of these quadrants of the chart. Educational research is rarely funded at a level that makes this possible. Most researchers are constrained by the available resources to only work with a small number of accessible classes or schools.
So, what use are such studies for producing generalisable results?
Small-scale research to incrementally extend the range of effectiveness
A single small-scale study can contribute to a research programme to explore the range of application of an innovation as if it was part of a large-scale study with a diverse sample. But this means such studies need to be explicitly conceptualised and planned as part of such a programme.
At the moment it is common for research papers to say something like
"…lots of research studies, from all over the place, report that asking students to
(i) first copy science texts omitting all the vowels, and then
(ii) re-constituting them in full by working from the reduced text, by writing it out adding vowels that produce viable words and sentences,
is an effective way of supporting the learning of science concepts; but no one has yet reported testing this pedagogic method when twelve year old students are studying the topic of acids in South Cambridgeshire in a teaching laboratory with moveable stools and West-facing windows.
In this ground-breaking study, we report an experiment to see if this constructivist, active-learning, teaching approach leads to greater science learning among twelve year old students studying the topic of acids in South Cambridgeshire in a teaching laboratory with moveable stools and West-facing windows…"
Over time, the research literature becomes populated with studies of enquiry-based science education, jig-saw learning, use of virtual reality, etc., etc., and these tend to refer to a range of national contexts, variously aged students, diverse science topics, etc., this all tends to be piecemeal. A coordinated programme of research could lead to researchers both (a) giving rich description of the context used, and (b) selecting contexts strategically to build up a picture across ranges of contexts,
"When there is a series of studies testing the same innovation, it is most useful if collectively they sample in a way that offers maximum information about the potential range of effectiveness of the innovation.There are clearly many factors that may be relevant. It may be useful for replication studies of effective innovations to take place with groups of different socio-economic status, or in different countries with different curriculum contexts, or indeed in countries with different cultural norms (and perhaps very different class sizes; different access to laboratory facilities) and languages of instruction …. It may be useful to test the range of effectiveness of some innovations in terms of the ages of students, or across a range of quite different science topics. Such decisions should be based on theoretical considerations.
Given the large number of potentially relevant variables, there will be a great many combinations of possible sets of replication conditions. A large number of replications giving similar results within a small region of this 'phase space' means each new study adds little to the field. If all existing studies report positive outcomes, then it is most useful to select new samples that are as different as possible from those already tested. …
When existing studies suggest the innovation is effective in some contexts but not others, then the characteristics of samples/context of published studies can be used to guide the selection of new samples/contexts (perhaps those judged as offering intermediate cases) that can help illuminate the boundaries of the range of effectiveness of the innovation."
Taber, 2019
Not that the research programme would be co-ordinated by a central agency or authority, but by each contributing researcher/research team (i) taking into account the 'state of play' at the start of their research; (ii) making strategic decisions accordingly when selecting contexts for their own work; (iii) reporting the context in enough detail to allow later researchers to see how that study fits into the ongoing programme.
This has to be a more scientific approach than simply picking a convenient context where researchers expect something to work well; undertake a small-scale local experiment (perhaps setting up a substandard control condition to be sure of a positive outcome); and then report along the lines "this widely demonstrated effective pedagogy works here too", or, if it does not, perhaps putting the study aside without publication. As the philosopher of science, Karl Popper, reminded us, science proceeds through the testing of bold conjectures: an 'experiment' where you already know the outcome is actually a demonstration. Demonstrations are useful in teaching, but do not contribute to research. What can contribute is an experiment in a context where there is reason to be unsure if an innovation will be an improvement or not, and where the comparison reflects good teaching practice to offer a meaningful test.
Small-scale research to inform local practice
Now, I would be the first to admit that I am not optimistic that such an approach will be developed by researchers; and even if it is, it will take time for useful patterns to arise that offer genuine insights into the range of convenience of different pedagogies.
Does this mean that small-scale studies in single context are really a waste of research resource and an unmerited inconvenient for those working in such contexts?
Well, I have time for studies in my final (bottom left) quadrant. Given that schools and classrooms and teachers and classes all vary considerably, and that what works well in a highly selective boys-only fee-paying school with a class size of 16 may not be as effective in a co-educational class of 32 mixed ability students in an under-resourced school in an area of social deprivation – and vice versa, of course!, there is often value in testing out ideas (even recommended 'research-based' ones) in specific contexts to inform practice in that context. These are likely to be genuine experiments, as the investigators are really motived to find out what can improve practice in that context.
Often such experiments will not get published,
- perhaps because the researchers are teachers with higher priorities than writing for publication;
- perhaps because it is assumed such local studies are not generalisable (but they could sometimes be moved into the previous category if suitably conceptualised and reported);
- perhaps because the investigators have not sought permissions for publication (part of the ethics of research), usually not necessary for teachers seeking innovations to improve practice as part of their professional work;
- perhaps because it has been decided inappropriate to set up control conditions which are not expected to be of benefit to those being asked to participate;
- but also because when trying out something new in a classroom, one needs to be open to make ad hoc modifications to, or even abandon, an innovation if it seems to be having a deleterious effect.
Evaluation of effectiveness here usually comes down to professional judgement (rather than statistical testing – which assumes a large random sample of a population – being used to invalidly generalise small, non-random, local results to that population) which might, in part, rely on the researcher's close (and partially tacit) familiarity with the research context.
I am here describing 'action research', which is highly useful for informing local practice, but which is not ideally suited for formal reporting in academic journals.
So, I suspect there may be an irony here.
There may be a great many small-scale experiments undertaken in schools and colleges which inform good teaching practice in their contexts, without ever being widely reported; whilst there are a great many similar scale, often 'forced' experiments, carried out by visiting researchers with little personal stake in the research context, reporting the general effectiveness of teaching approaches, based on misuse of statistics. I wonder which approach best reflects the true spirit of science?
Source cited:
- Taber, K. S. (2019). Experimental research into teaching innovations: responding to methodological and ethical challenges. Studies in Science Education, 55(1), 69-119. doi:10.1080/03057267.2019.1658058 [Download article]
Notes:
1 For example:
Even in the natural sciences, we can never be absolutely sure that we have controlled all relevant variables (after all, if we already knew for sure which variables were relevant, we would not need to do the research). But usually existing theory gives us a pretty good idea what we need to control.
Experiments are never a simple test of the specified hypothesis, as the experiment is likely to depends upon the theory of instrumentation and the quality of instruments. Consider an extreme case such as the discovery of the Higgs boson at CERN: the conclusions relied on complex theory that informed the design of the apparatus, and very challenging precision engineering, as well as complex mathematical models for interpreting data, and corresponding computer software specifically programmed to carry out that analysis.
The experimental results are a test of a hypothesis (e.g., that a certain particle would be found at events below some calculated energy level) subject to the provisos that
- the theory of the the instrument and its design is correct; and
- the materials of the apparatus (an apparatus as complex and extensive as a small city) have no serious flaws; and
- the construction of the instrumentation precisely matches the specifications;
- and the modelling of how the detectors will function (including their decay in performance over time) is accurate; and
- the analytical techniques designed to interpret the signals are valid;
- the programming of the computers carries out the analysis as intended.
It almost requires an act of faith to have confidence in all this (and I am confident there is no one scientist anywhere in the world who has a good enough understanding and familiarity will all these aspects of the experiment to be able to give assurances on all these areas!)
I would hope at least that as professional scientists and engineers they might be a little more aware of this complex chain of argumentation needed to support robust conclusions than many students – for students often seem to be overconfident in the overall value of research conclusions given any doubts they may have about aspects of the work reported.
Read about the Critical Reading of Empirical Studies Tool
Galileo Galilei was one of the first people to apply the telescope to study the night sky (image by Dorothe from Pixabay)
A historical example is Galileo's observations of astronomical phenomena such as Jovian moons (he spotted the four largest: Io, Europa, Ganymede and Callisto) and the irregular surface of the moon. Some of his contemporaries rejected these findings on the basis that they were made using an apparatus, the newly fanged telescope, that they did not trust. Whilst this is now widely seen as being arrogant and/or ignorant, arguably if you did not understand how a telescope could magnify, and you did not trust the quality of the lenses not to produce distortions, then it was quite reasonable to be sceptical of findings which were counter to a theory of the 'heavens' that had been generally accepted for many centuries.
2 I have discussed a number of examples on this site. For example:
Shock result: more study time leads to higher test scores (But 'all other things' are seldom equal)
Experimental pot calls the research kettle black: Do not enquire as I do, enquire as I tell you
Lack of control in educational research: Getting that sinking feeling on reading published studies
3 For a detailed discussion of these and other challenges of doing educational experiments, see Taber, 2019.
4 Consider these two situations.
A researcher wants to find out if a new textbook 'Science for the modern age' leads to more learning among the Grade 10 students she teaches than the traditional book 'Principles of the natural world'. Imagine there are fifty grade 10 students divided already into two classes. The teacher flips a coin and randomly assigns one of the classes to the innovative book, the other being assigned by default the traditional book. We will assume she has a suitable test to assess each students' learning at the end of the experiment.
The teacher teaches the two classes the same curriculum by the same scheme of work. She presents a mini-lecture to a class, then sets them some questions to discuss using the text book. At the end of the (three part!) lesson, she leads a class disucsison drawing on students' suggested answers.
Being a science teacher, who believes in replication, she decides to repeat the exercise the following year. Unfortunately there is a pandemic, and all the students are sent into lock-down at home. So, the teacher assigns the fifty students by lot into two groups, and emails one group the traditional book, and the other the innovative text. She teaches all the students on line as one cohort: each lesson giving them a mini-lecture, then setting them some reading from their (assigned) book, and a set of questions to work through using the text, asking them to upload their individual answers for her to see.
With regard to experimental method, in the first cohort she has only two independent units of analysis – so she may note that the average outcome scores are higher in one group, but cannot read too much into that. However, in the second year, the fifty students can be considered to be learning independently, and as they have been randomly assigned to conditions, she can treat the assessment scores as being from 25 units of analysis in each condition (and so may sensibly apply statistics to see if there is a statistically significant different in outcomes).
5 Inferential statistical tests are usually used to see if the difference in outcomes across conditions is 'significant'. Perhaps the average score in a class with an innovation is 5.6, compared with an average score in the control class of 5.1. The average score is higher in the experimental condition, but is the difference enough to matter?
Well, actually, if the question is whether the difference is big enough to likely to make a difference in practice then researchers should calculate the 'effect size' which will suggest whether the difference found should be considered small, moderate or large. This should ideally be calculated regardless of whether inferential statistics are being used or not.
Inferential statistical tests are often used to see if the result is generalisable to the wider population – but, as suggested above, this is strictly only valid if the population of interest have been randomly sampled – which virtually never happens in educational studies as it is usually not feasible.
Often researchers will still do the calculation, based on the sets of outcome scores in the two conditions, to see if they can claim a statistically significant difference – but the test will only suggest how likely or unlikely the difference between the outcomes is, if the units of analysis have been randomly assigned to the conditions. So, if there are 50 learners each randomly assigned to experimental or control condition this makes sense. That is sometimes the case, but nearly always the researchers work with existing classes and do not have the option of randomly mixing the students up. [See the example in the previous note 4.] In such a situation, the stats. are not informative. (That does not stop them often being reported in published accounts as if they are useful.)
6 That is, if it possible to address such complications as participant expectations, and equitable teacher-familiarity with the different conditions they are assigned to (Taber, 2019).
7 A usual ethical expectation is that participants voluntarily (without duress) offer informed consent to participate.
Read about voluntary informed consent
