causes and conditionsThe search for causation |
CHAPTER 4 |
This chapter introduces key issues in understanding causation in educational research. These include:
causes and conditions
causal inference and probabilistic causation
causation, explanation, prediction and correlation
causal over-determination
the timing and scope of the cause and the effect
causal direction, directness and indirectness
establishing causation
the role of action narratives in causation
researching causes and effects
researching the effects of causes
researching the causes of effects
Educationists and social scientists are concerned not only for ‘what works’ but ‘why’, ‘how’, ‘for whom’ and ‘under what conditions and circumstances’. They want to predict what will happen if such-and-such an intervention is introduced, and how and why it will produce a particular effect. This points us to an important feature of educational research, which is to look for causation: what are the effects of causes and what are the causes of effects? This is not a straightforward enterprise, not least because causation is not often observable but can only be inferred, and it is highly unlikely that indisputable causality is ever completely discoverable in the social sciences. At best probabilistic causation offers a more fitting characterization of causation in educational research. Causation is often considered to be the ‘holy grail’ of educational research, and this chapter introduces some key considerations in investigating causation.
Novice researchers are faced with many decisions concerning causation in their research, for example:
Whether the research is seeking to establish causation, and if so, why.
Deciding when causation is demonstrated, recognizing that causation is never 100 per cent certain.
Deciding what constitutes a cause and what constitutes an effect.
Deciding what constitutes evidence of the cause and evidence of the effect.
Deciding the kind of research and the methodology of research that is necessary if causation is to be investigated.
Deciding whether the research is investigating the cause of an effect, the effect of a cause, or both.
To infer simple, deterministic or regular causation may be to misread many situations excepting, perhaps, those where massive single causation is clear. It may be more useful for the researcher to consider causal processes than single events (Salmon, 1998), not least because there is often more than a single cause at work in any effect and there may be more than one effect from a single cause. Indeed, the researcher has to distinguish between causes, reasons, motives, determination and entailment, and whilst these might all exert causal force in some circumstances or enable us to make causal explanations or predictions, in other circumstances they do not.
What then makes a cause a cause, and an effect an effect? How do we know? Though we can say that causation takes places in a temporal sequence – the cause precedes the effect, and with temporal succession (Hume’s criterion of ‘priority’ (Hume, 1955; Norton and Norton, 2000)), this does not help the researcher very much.
One distinguishing indication that causation is taking place or has taken place is the presence of counterfactuals (Mackie, 1993), i.e. the determination that the absence of X (the supposed cause) would have led to the absence of Y (the effect). If we are seeking to establish that such-and-such is a contributing cause (X) of an effect (Y) we ask ourselves whether, if that supposed cause had not been present, then would the effect have occurred or been what it actually was; if the answer is ‘no’ then we can suppose that X is a true cause. For example, if there had been no ice on a path then I would not have fallen over and broken my arm. So the presence of ice must have been a contributing cause of the effect – one of many causes (e.g. my poor sense of balance, my poor eyesight in not seeing the ice, the ambient darkness, wearing slippery soled shoes, my brittle bones because of my age, etc.).
The counterfactual argument is persuasive, but problematical: how do we know, for example, what the outcome would have been if there had been no patch of ice on the path where I was walking? Can we predict with sufficient certainty to attribute the counterfactual causality here? How can we prove that the effect would not have happened if a particular cause had not been present? How do we know that I would or would not have slipped and fallen if the ice had not been present? In true experiments this is addressed by having a control group: the control group is supposed to indicate what would have happened if the intervention had not occurred. The problem is that much research is not experimental.
If it were only the presence of ice that caused me to fall and break my arm, then this would be a very simple indication of causality; the problem is that the presence of ice in this instance is perhaps not a sufficient cause – had my balance been good, my eyesight good, the ambient light good, my shoes had good grips on their soles and my bones been less brittle then I would not have fallen and broken my arm.
The difficulty here is to establish the relative strength of the causes in a multi-causal situation, i.e. the several conditions that, themselves, contribute to the accident. The presence of those included causes affects their relative strengths in a specific context, and the absence of some of these causes in the same context may raise or lower the relative strengths of others.
The example of falling on the ice also indicates an important feature of causation: causes cannot be taken in isolation, they may need to be taken together (compound causes, i.e. they only exert causative force when acting in concert), and there may be interaction effects between them. On its own, the patch of ice might not have caused my fall and broken arm; it was perhaps neither sufficient nor necessary, as I could have fallen and broken my arm anyway because of my slippery shoes and poor balance. On its own, my poor balance did not cause me to fall and break my arm. On its own, the darkness did not cause me to fall and break my arm. On their own, my slippery soles did not cause me to fall and break my arm. On their own, my brittle bones did not cause me to fall and break my arm. But put all these together and we have sufficient conditions to cause the accident. For the researcher, looking for individual causes in a contextualized situation may be futile.
In understanding the causes of effects one has to understand the circumstances and conditions in which the two independent factors – the cause and the effect are located and linked (the link is contingent rather than analytic). Discovering the circumstances – conditions – in which one variable causes an effect on another is vital in understanding causation, for it is the specific combination of necessary and/or sufficient conditions that may produce an effect. Causes of effects work in specific circumstances and situations, and account has to be taken of these circumstances and conditions.
For the researcher, the difficulty in unravelling the effects of causes and the causes of effects is heightened by the fact that causes may be indirect rather than direct (cause A causes effect B, and effect B causes effect C) or that they may only become a cause in the presence of other factors (I may fall over on ice and not break my arm if I am young and land well, but as an older person with more brittle bones I may land awkwardly and break my arm – the fall is not a sufficient condition or cause of my broken arm).
Identifying and understanding causation may be problematic for researchers, as effects may not be direct, linear functions of causes, and because there may be few, many, increasing, reducing, unpredictable, i.e. non-linear effects of causes. A small cause can bring about a large or irregular effect; a large cause may bring about a small or irregular effect. Causation is often an inductive and empirical, rather than a logical, deductive matter, and, indeed, it is often unclear what constitutes a cause and what constitutes an effect as these are often umbrella terms, under which are sub-causes and sub-effects, causal processes and causal links bringing several factors together both at a particular point in time (the moment of falling, in the example above) and in a temporal sequence (e.g. taking and passing a public examination).
Further, there is an asymmetry at work in causation effect: a cause can produce an effect but not vice versa: being young, good looking and female may help me to pass my driving test if I am in the presence of a leering male examiner, but passing my driving test does not cause me to be young, good looking and female.
It is often dangerous to say that such-and-such is definitely the cause of something, or that such-and-such is definitely the effect of something. Causation in the human sciences is much more tentative, and may be probabilistic rather than deterministic. Hume’s (2000) own rules for causation are:
contiguity (of space and time) (the cause is contiguous with the effect);
priority/succession (the cause precedes the effect);
constant conjunction (the coupling of one event and its successor are found to recur repeatedly);
necessary connection (which is learned from experience, habit and custom rather than from deductive, logical, necessary proof ).
One can detect correlation in Hume’s ideas rather than actual causation. He argues that causation is inferred, inductively, by humans rather than being an objective matter.
The inferential, conjectural and probabilistic nature of much causation in educational research (rather than being absolute, deductive and deterministic), coupled with the fact that causation is frequently unobserved or unobservable, renders the study of causation problematic for educational researchers. Indeed there is a danger in isolating and focusing on singular causes separately from other contributing causes, contexts and conditions, and it is perhaps more fitting to regard causes as processes over time rather than single events. Further, in an interconnected world of multiple causes and causal nets, conditions and interactions may provide better accounts of causation than linear determinism.
In unravelling causes and effects, the researcher is faced with the task of identifying what actually constitutes a cause and what constitutes an effect. The contexts and conditions of an event are as important as the trigger of an effect, and may be contributing causes. In the example earlier, my falling and breaking my arm was precipitated – triggered – by the ice on the path, but, without the presence of other contributing factors I might not have fallen and I might not have broken my arm. The trigger of the effect may not be its sole cause but only the last cause in a causal chain, sequence of events, series or network of conditions before the effect occurs, even though causes often raise the likelihood of their effects rather than guaranteeing them (Mellor, 1995: 69–70). Indeed, whilst probability often concerns identifying likelihood, the strongest probability is not always the same as the strongest causation. I might think that putting pressure on a child to succeed has the strongest possibility of causing her success, but the actual cause might lie elsewhere, e.g. the teacher might be very effective, the students might be highly motivated or the examination might be very easy.
The demonstration of causation is difficult. Causation is neither the same as explanation (Clogg and Haritou, 1997: 106; Salmon, 1998: 5–8) (e.g. an explanation may be wrong, or it may be giving the meaning of something, or it may be indicating how to do something), nor is causation the same as giving a reason. For example, I might take a day off work, giving the reason that I am sick, but the real reason may be simply that I am lazy or want to go shopping instead.
Nor is causation the same as prediction. Just because I observe something happening once does not mean I can predict that it will happen again (the problem of induction), as the conditions could be different, or indeed, even if the conditions were very similar (as chaos theory tells us).
I might be able to predict something even though my prediction is based on the wrong identification of causes – I can predict that there will be a storm because I have observed the barometric reading falling, but the fall in the barometric reading does not cause the storm. Formally put, the two variables – the barometric reading and the storm – are ‘screened off ’, separated and kept apart from each other (Reichenbach, 1956; Salmon, 1998). They have correlational but no causal relationships to each other, and are both caused by a third factor – the drop in air pressure, see Figure 4.1.
I might predict that a person’s hands might be large if she has large feet, but having large hands does not cause her to have large feet – the cause might lie in a genetic predisposition to both. It is one thing to say that a change in one variable (A) is associated with a change in another variable (B); it is an entirely different thing to say that a change in one variable (A) brings about a change in another variable (B); and it is an entirely different thing again to say that a change in one variable (B) is brought about by a change in another variable (A), i.e. that it is caused by that change in variable A.
In attributing genuine causation, it is useful to ‘screen of ’ unrelated dependent variables from the variables that are directly relevant to the situation being researched, in order to ensure that the effect of one variable is removed from the equation, e.g. to discount that variable or to control for the effects of other variables, such as the presence of a third variable or several variables, by partial correlations and structural equation modelling. It is also important in screening off to ensure that one variable is not deemed to have an influence on another when, in fact, this is not the case. This presumes that it is actually possible to identify which factors to screen off from which (Pearl (2009: 423–7) indicates how this can be approached), and, in the case of multiple causality (or in the cases of over-determination, discussed below), this may not be possible.
Screening off requires the ability to separate out causes, and this may be difficult to the point of impossibility. However, in seeking to establish genuine causation, the researcher must consider controlling for the effects of additional variables, be they prior/exogenous variables or intervening/endogenous variables, as these will exert a non-causal influence on the dependent variables.
In conducting research that seeks to attribute causation, it is important to control for the effects of variables, i.e. to hold them constant (matched) so that fair attribution of causality and the weight of causal variables can be assessed (though relative weights of causes are, strictly speaking, superfluous in discussing causation, they are questionable indicators of causation). Further, identifying the relative strengths of causes depends on the presence or absence of other causes. For example, in looking at examination success (the effect), if my research confines itself to looking at the relative strength of causes A (hours of study), B (IQ) and C (motivation) in producing the effect, I might find that C (motivation) is the strongest of the three causes. However, if I were to add a new variable (D) (an outstanding teacher helping the student), then it may be that D is the overriding cause and that A, B and C are of equally low strength, or that A becomes the second strongest factor.
Statistical tools such as crosstabulations, correlation and partial correlation, regression and multiple regression, and structural equation models (see Chapters 34–36) can be used to assist here in the analysis of causation, though it is often difficult to control direct, indirect, antecedent, intervening and combined influences of variables on outcomes (though statistical tools and graphical methods can assist here (Pearl, 2009: 423–7)).
In considering the control of variables, let us examine, for example, the subject choices of secondary school male and female students (Table 4.1). Here we can see overwhelmingly that males choose physics far more than females, and females choose biology far more than males. The researcher wishes to know if the allocation of certain teachers to teach the secondary school science subjects affects the students’ choice (i.e. whether it is the subject of the teacher, or some combination of these) that is causing the students to choose the subjects that they choose. The researcher introduces the third variable of the ‘teacher’ as a control variable, with two values: Teacher A and Teacher B, and then partitions the data for males and females according to either Teacher A or Teacher B, see Table 4.2.
When the data are partitioned by teacher (Teacher A and Teacher B) the researcher notes that the percentages in each of the partial tables (one part of the table for Teacher A and the other part of the table for Teacher B) in Table 4.2 are very similar to the original percentages of the root table (Table 4.1). She concludes that whether Teacher A or Teacher B is teaching the class makes no appreciable difference to the choices made by the students. The percentages in the new table (4.2) replicate very closely those in the original table (4.1). The researcher concludes that the teacher involved is exerting no causal influence on the choice of subjects by the secondary school males and females.
However, let us imagine that the partial tables had yielded different data (Table 4.3). This time the results of the choices made by males and females who are with Teacher A and Teacher B are very different. The percentages in the new table (4.3) are very different from those in the original table (4.1). This suggests to the researcher that, in this instance, the teacher of the class in question is making a causal difference to the choices of science subject made by the students.
TABLE 4.1 SCIENCE CHOICES OF SECONDARY SCHOOL MALES AND FEMALES
|
Male |
Female |
Total |
Preference for physics |
175 (55.1%) |
87 (27.9%) |
262 (41.6% of total) |
Preference for biology |
143 (44.9%) |
225 (72.1%) |
368 (58.4% of total) |
Column total |
318 (100%) |
312 (100%) |
630 |
Percentage of total |
50.5% |
49.5% |
100% |
However, this only tells us the ‘what’ of causation, or, to be more precise, it only gives us an indication of association and possible causation: it appears that the teacher makes no difference in Table 4.2 but does make a difference in Table 4.3. How this becomes a causal matter is another question altogether: how does the teacher actually affect the males’ or females’ choices of which science subject to follow. For example is it that: (a) Teacher A is male and Teacher B is female, and students tend to prefer to be with teachers of their own sex; (b) Teacher A has a better reputation than Teacher B for helping students to pass public examinations with high grades, and students are anxious to do well; (c) Teacher A is more sympathetic than Teacher B, so that students can relate more easily to Teacher A, and so they choose Teacher A; (d) Teacher A has a better sense of humour than Teacher B, and students prefer a good humoured teacher; (e) Teacher A explains matters more clearly than Teacher B, and students prefer clear explanations, and so on. The point here is that, though one can deduce certain points from contingency tables and partial tables, they may not actually indicate causality. The same principle for holding variables constant, this time in correlational research, is discussed in Chapter 35.
In establishing causation, it is important to separate covariance and correlation between two unrelated and non-interacting dependent variables due to a common cause from the interaction of dependent variables due to the presence of a common cause (as in the examples of the barometer and the storm earlier).
It is rare to find a single cause of a single effect. It is more often the case that there are several causes at work in a single situation and that these produce a multiplicity of effects. For example, why are so many young children well behaved at school, when nobody has explicitly taught them the hidden curriculum (Jackson, 1968) of rules, regulations, taking turns, sharing, being quiet, knowing that the teacher is in charge and has all the power, putting up with delay, denial and only being one out of many children who has to gain the teacher’s attention? The answer is over-determination: many events, both separately and in combination, lead to the same outcome: the young child must do as she is told and that having a nice time at school depends on how effectively she learns these rules and abides by them. Many causes; same effect: good behaviour.
Causal over-determination ís ‘where a particular effect is the outcome of more than one cause, each of which, in itself, would have been sufficient to have produced the effect’ (Morrison, 2009: 51). A familiar example is the issue of which bullet can be said to have killed a man, which causes his death (Horwich, 1993), if two bullets simultaneously strike a man’s head. Either one bullet or the other caused the death (cf. Mellor, 1995: 102). ‘The man would have died, even if one or the other of the bullets was not fired, and if bullet A did not cause the death, it would be causally true to say that the man would have died’ (Morrison, 2009: 51). Let us say that, in a study of homework and its effect on mathematics performance, a rise in homework might produce a rise in students’ mathematics performance. However, this is not all: there may have been tremendous parental pressure on the child to do well in mathematics, or the student might have been promised a vast sum of money if her mathematics performance increased, or the school might have exerted huge pressure on the student to succeed, or the offer of a university place was contingent on a high mathematics score. The rise in mathematics performance may not have required all of the factors to have been present in order to bring about the effect; any one of them could have produced the effect. The effect is ‘over-determined’. One effect may have one or several causes. Whilst this is commonplace, it is important to note this in order to refute the claims frequently made by protagonists of such-and-such an intervention in education that it alone improves performance; if only it were that simple!
Turn back to the earlier example of my falling on the ice and breaking my arm. Maybe I had a weakness in my arm from an injury many years before, and maybe when I injured my arm years before, I could not have predicted that, many years later, I would fall on ice and break my arm. The issue is not idle for researchers, for it requires them to consider, in terms of temporality, what are relevant causes and what to include and exclude from studies of causation, how far back in time to go in establishing causes and how far forward in time to go in establishing effects.
Just as the timing of causes may be unclear, so the timing of the effects of a cause may be unclear. Effects may be short term only, delayed, instantaneous, immediate, cumulative and long term; indeed the full effects of a cause may not be revealed in a single instance, as an effect may be a covering term for many effects that emerge over time (e.g. the onset and presenting of cancer has several stages; cancer is not a single event at one point in time). Temporality and causation are intimately connected but separate.
The examples above also indicate that terms such as ‘cause’ and ‘effect’ are, in many cases, a shorthand for many sub-causes, sub-processes and sub-effects. Further, causes and effects may only reveal themselves over time, and, indeed, it may be difficult to indicate when a cause begins (which cigarette brought about the onset of cancer, or when did smoking first bring about the onset of the cancer) or ends, and when an effect begins (e.g. I may continue smoking even after the early onset of lung cancer). I might hate studying mathematics at school but find it very attractive twenty years later; had my interest in mathematics been post-tested immediately I left school, the result would have been lower than if I had been tested twenty years later.
Where a cause begins and ends, where an effect begins and ends, when and how causes and effects should be measured, evaluated, ascertained and assessed, are often open questions, requiring educational researchers to clarify and justify their decisions on timings in isolating and investigating causes and effects. Quantitative data may be useful for identifying the ‘what’ of causation – what causes an effect – but qualitative data are pre-eminently useful for identifying the ‘how’ of causation – how causation actually works, the causal processes at work.
Consider, too, the reason for the ice patch being present on the path in the earlier example, and my being on the ice on the day in question. Maybe the local government services had not properly cleared the path of ice on that day, or maybe, as an ailing pensioner, I would normally be accompanied by a carer or with an assistant whenever I went out, but on that day the service provider failed to turn up, so I was forced to go out on my own. Again, the issue is not idle for researchers, for it requires them to consider how widely or narrowly to cast their net in terms of looking for causes (how far out and how far in). In determining what are relevant causes the researcher has to decide what to include and exclude from studies of causation, e.g. from the psychological to the social, from the micro to the macro, and to decide the direction and combination of such causes.
The determination of a cause involves decisions on how far to go back in a temporal causal chain or network of events, and how wide or narrow to go in the causal space (how many conditions and circumstances contribute to the causation at work in a given situation). It may be difficult, if not impossible, to identify and include all the causal antecedents in a piece of research. Here the concept of necessary and sufficient conditions is raised, as is the importance of identifying the causal trigger in a situation (the last cause in a causal chain or a linkage of several conditions). The striking of a match might cause it to flare, but that is not the only factor to be taken into account. Whether it flares depends on the abrasiveness of the striking surface, the dryness of the materials, the strength of the strike, the duration of the strike, the presence of sufficient oxygen in the atmosphere, and so on.
There may be an infinite number of causes and effects, depending on how far back one goes in time and how wide one goes in terms of contexts. This presents a problem of where to establish the ‘cut of ’ point in identifying causes of an effect. Whilst this may be addressed through the identification of necessary and sufficient conditions (Mackie, 1993), in fact this does little to attenuate the problem in social sciences, as not only is it problematic to identify what qualify as necessary or sufficient conditions, but these will vary from context to context, and even though there may be regularities of cause and effect from context to context, there are also differences from context to context.
The issue to be faced by researchers here is one of ‘boundary conditions’ and ‘circumscription’ (Pearl, 2009: 420) – which factors we include or exclude can affect our judgements of causality. If, in a study of student performance, I only look at teacher behaviour and its influence on student performance then I might be led to believe that teacher behaviour is the cause of student performance, whereas if I only look at student motivation and its influence on student performance then I might be led to believe that student motivation is the cause of student performance. Researchers rarely, if ever, include the universe of conditions, only a selection from that universe, and this might distort the judgements made about causation or where to look for causation. Whilst it may not be possible to identify the universe of conditions, the researcher has to be aware of the dangers of circularity, i.e. I am only interested in effect Y, so I only look at possible causes X, and then I find, wonder upon wonders, that X is the cause of Y, simply because I have not considered alternatives. It is important to identify and justify the inclusion and exclusion of variables in researching causation, that is to select the field of focus sufficiently widely and to consider possible alternative explanations of cause and effect.
The problem of identifying causes and effects is further compounded by consideration of direct and indirect causes and effects. It is also useful to describe causation in terms of recursive (single lines of direction) rather than non-recursive (mutual lines of direction) models, as causation can take more than one direction at a time. This justifies the use of non-recursive models of causation and of causal nets: clusters of causes that act together in multiple directions. In a recursive model the causality is unidirectional (which may oversimplify the direction of causation), whereas in a non-recursive model causation is in one or more directions. Many structural equation models are non-recursive.
Whilst the research may wish to identify a cause A that brings about the direct effect B, in practice this is seldom the case, as between A and B might be a huge number of intervening variables and processes operating, both exogenous and endogenous:
An exogenous variable is one whose values are determined outside the model (e.g. a structural equation model or a causal model) in which that variable is being used, or which is considered not to be caused by another variable in the model, or which is extraneous to the model.
An endogenous variable is one whose values or variations are explained by other variables within the model, or which is caused by one or more variables within the model. It is important to identify which causes mediate, and are mediated by, other causes. How the researcher does this takes many forms, from theoretical modelling and testing of the model with data, to eliciting from participants what are the causes.
Whilst causation is not straightforward to demonstrate, this is not to suggest that establishing causation should not be attempted. There are regularities, there are likelihoods based on experience, there are similarities between situations and people, indeed the similarities may be stronger than the differences. This suggests that establishing probabilistic causation or inferring causation, whilst complex and daunting, may be possible for the researcher.
The problem for the researcher is to decide which variables to include, as the identification and inclusion/exclusion of relevant variables in determining causation is a major difficulty in research. Causes, like effects, might often be better regarded in conjunction with other causes, circumstances and conditions rather than in isolation. Contextuality – the conditions in which the cause and effect take place – and the careful identification and inclusion of all relevant causes are key factors in identifying causation.
It is not an easy task to establish causation. For example, causation may be present but unobserved and indeed unobservable, particularly in the presence of stronger causes or impeding factors. I might take medication for a headache but the headache becomes worse; this is not to say that the medication has not worked, as the headache might have become even stronger without the medication. The effects of some causes may be masked by the presence of others, but nonetheless causation may be occurring.
Morrison (2009: 45) gives an example where,
in the case of the causal relationship between smoking (A), heart disease (B) and exercise (C), smoking (A) is highly correlated with exercise (C): smokers exercise much more than non-smokers. Though smoking causes heart disease, exercise actually is an even stronger preventative measure that one can take against heart disease. The corollary of this is that smoking prevents heart disease.
(Hitchcock, 2002: 9)
The way in which the cause operates may also be unclear. There are many examples one can give, Morrison (2009: 45), for instance, gives the example of small class teaching. In one class operating with small class teaching, the teacher in that class uses highly didactic, formal teaching with marked social distance between the teacher and the student (Factor A), and this is deemed to be an inhibitor of the beneficial effects of small class teaching on students’ attainment in mathematics: didactic teaching reduces mathematics performance. However the same highly didactic, formal class teaching (Factor A) significantly raises the amount of pressure placed on the students to achieve highly (Factor B), and this (Factor B) is known to be the overriding cause of any rise in students’ performance in mathematics, e.g. in small classes the teacher can monitor very closely the work of each child: high pressure raises mathematics performance. Now, it could be argued that Factor A – an ostensibly inhibiting factor for the benefits of small class teaching – actually causes improvements in mathematics performance in the small class teaching situation.
Another example is where greater examination pressure on students (A) increases their lack of self-confidence (C), but it also increases the student’s hard work (B), and hard work reduces the student’s lack of self-confidence (C). In other words, the likelihood of the effect of A on C may be lower than the effect of B on C, given A. Let us say that A increases the likelihood of C by 20 per cent, and A increases the likelihood of B by 35 per cent, whilst B reduces the likelihood of C by 75 per cent. In this instance increasing the examination pressure increases the student’s self-confidence rather than reduces it (see Figure 4.2).
The point here that a cause might raise the likelihood of an effect, but it may also lower that likelihood, and the presence of other conditions or causes affects the likelihood of an effect of a cause. A diagrammatic representation of these examples is in Figure 4.2 (note that the length of the lines indicates the relative strength of the influence). A cause might lower the likelihood of an effect rather than increase it.
Many cause-and-effect models are premised on linear relationships between cause and effect (i.e. a regular relationship, e.g. a small cause has a regular small effect and a large cause has a regular large effect, or a small cause has a regular large effect and a large cause has a regular small effect). However, seeking linear relations between cause and effect might be misguided, as the effects of causes might be non-linear (e.g. a small or large cause may produce a large, small, irregular or no effect), and it might be to deal with singular or a few causes and singular or a few effects, overlooking the interrelatedness and interactions of multiple causes with each other, with multiple effects and indeed with the multiple interactions of multiple effects. Relationships and their analysis may be probabilistic, conditional and subjunctive rather than linear. Indeed nets and conditions of causation might be more fitting descriptions of causation than causal lines or chains of events or factors.
One way of focusing a causal explanation is to examine regularities and then to consider rival explanations of causes and rival hypotheses of these regularities. The observation of regularities, however, is not essential to an understanding of causation, as all cases may be different but no less causative. Further, the best causal explanation is that which is founded on, and draws from, the most comprehensive theory (e.g. that theory which embraces intentionality, agency, interaction as well as structure, i.e. micro-and macro-factors), that explain all the elements of the phenomenon, that fit the explanandum (that which must be explained) and data more fully than rival theories, and which are tested in contexts and with data other than those that have given rise to the theory and causal explanation.
Given the complexity of probabilistic causation, it would be invidious to suppose that a particular intervention will necessarily bring about the intended effect. Any cause or intervention is embedded in a web of other causes, contexts, conditions, circumstances and effects, and these can exert a mediating and altering influence between the cause and its effect.
Statistics, both inferential and descriptive, can indicate powerful relationships. However, these do not necessarily establish unequivocal, direct causation; they may establish the ‘what’ of causation but not the ‘how’. I might assume that A and B cause C, and that C causes D; it is a causal model, and I might measure the effects of A and B on C and the effect of C on D. However, here causation lies in the assumptions behind the model rather than in the statistical tests of the model, and the causal assumptions that lie behind the model may derive from theory rather than the model itself. Statistics alone do not prove causation. To believe that they do is to engage in circular thinking. Rather, causation is embodied in the theoretical underpinnings and assumptions that support the model, and the role of statistics is to confirm, challenge, extend and refine these underpinnings and assumptions. Behind statistics that may illuminate causation lie theories and models, and it is in the construct validity of these that causation lies. It is the mechanisms of causation that should concern researchers rather than solely numbers and statistical explanations.
Many statistics rely on correlational analysis or on assumptions that pre-exist the statistics, i.e. the statistics might only reinforce existing assumptions and models rather than identify actual causation. Even more sophisticated statistics such as structural equation modelling, multiple regression and multivariate analysis succumb to the charge of being no more powerful than the assumptions of causation underpinning them, and, indeed, they often grossly simplify the number or range of causes in a situation, in the pursuit of a simple, clear and easily identifiable model.
How is it that X causes Y; what is happening in X to cause Y? In short, what are the processes of causation? In order to understand this involves regarding causation as dynamic rather than static, as a process rather than a single event, and as involving motives, volitions, reasons, understandings, perceptions, individuality, conditions and context, and the dynamic and emerging interplay of factors, more often than not over time. It is here that qualitative data come into their own, for they ‘get inside the head’ of the actors in a situation.
A neat example of this is what has come to be known as the ‘Rashomon effect’ in social sciences (e.g. Roth and Mehta, 2002). It is over 60 years since Kurosawa’s film Rashomon stunned audiences at the Venice film festival. It provides four discrepant witness accounts of the same event – an encounter between a samurai, his wife and a bandit, that led to the effect of the samurai’s death – in which the causes could have been murder or suicide, consensual sex or rape, fidelity or infidelity. The causal accounts are given by a woodcutter, the bandit, the wife and the spirit of the dead samurai speaking through a medium. Each self-serving account protects the honour of the teller and tries to exonerate each. At the end, there is no clear statement of whose version is correct; truth flounders in the quagmire of epistemology, perception and motives.
Anthropologists, lawyers and social scientists (Roth and Mehta, 2002) seized on the film as an example of the multilayered, contested truth of any situation or its interpretation, coining the term the ‘Rashomon effect’ to describe an event or truth which is reported or explained in contradictory terms, that gives differing and incompatible causal accounts of an effect: a death. There is more than one causal explanation at work in a situation, and it is the task of the researcher to uncover these, and to examine the causation through the eyes of those imputing the causation.
Action narratives and agency are important in accounting for causation and effects, and, because of a multiplicity of action narratives and individual motivations in a situation, there are multiple pathways of causation rather than simple input–output models. In understanding the processes of causation, the power of qualitative data is immense and, indeed, argues for mixed methods in establishing causation: numerical data to identify the variables at work, and qualitative data to indicate how they are working in specific situations.
Causal explanations that dwell at the level of aggregate variables are incomplete, as behind them, and feeding into them, lie individuals’ motives, values, goals and circumstances, and it is these that could be exerting the causal influence; hence a theory of individual motives might be required in understanding and explaining the causation here.
For example, it is commonplace for a survey to ask respondents to indicate their sex, but it is an entirely different matter – even if different responses are given by males and females to rating scales in a survey – to say that sex causes the differences in response. How, actually, is sex a causal factor?
Further, between aggregate independent and dependent variables of cause and effect respectively lie a whole range of causal processes, and these could be influencing the effect and, therefore, have to be taken into account in any causal explanation. The argument supports micro-to macro-analysis and explanation rather than macro-to micro-analysis and explanation. How macro-structural features from society actually enter into individuals’ actions and interactions, and how individuals’ actions and interactions determine social structures – the causal processes involved – need cautious elucidation, their current status often being opaque processes in a black box, input–output model of causation.
The researcher investigating the effects of a cause or the causes of an effect has many questions to answer, for example:
What is the causal connection between the cause and the effect (how does the cause bring about the effect and how has the effect been brought about by the cause)?
What are the causal processes at work in the situation being investigated?
What constitutes the evidence of the causal connection?
On what basis will the inference of causality be made?
What constitutes the evidence that a cause is a cause and that an effect is an effect?
What constitutes the evidence that a cause is the cause (and that there is not another cause) and that an effect is the effect (and that there is not another effect)?
Is the research investigating the effects of a cause (an interventionist strategy) or the cause of an effect (a post hoc investigation)?
How will the research separate out a range of possible causes and effects, and how will decisions be made to include and/or exclude possible causes and effects?
What methodology will be chosen to examine the effects of causes?
What methodology will be chosen to examine the causes of effects?
What kind of data will establish probabilistic causation?
When will the data be collected from which causation will be inferred?
As mentioned earlier, the timing of data collection is a critical feature in establishing causation and the effects of causes. Here the greater the need to establish causal processes, the closer and more frequent should be the data collection points. Moreover, qualitative data could hold pre-eminence over quantitative methods in establishing causation and causal processes. Indeed longitudinal studies might yield accounts of causation that are more robust than cross-sectional studies.
It is not enough to say that such-and-such a cause brings about such-and-such an effect, for, whilst it might establish the likelihood that the cause brings about an effect or that an effect has been brought about by a cause, this does not tell the researcher how the cause brings about the effect or how the effect has been brought about by the cause, i.e. what are the causal processes at work in connecting the cause with the effect and vice versa. If the research really wishes to investigate the processes of causation then this requires detailed, in-depth analysis of the connections between causes and effects.
For example, it is not enough to say that smoking can cause cancer; what is required is to know how smoking can cause cancer – what happens between the inhalation of smoke and the presentation of cancer cells. I might say that turning on a light switch causes the light bulb to shine, but this is inaccurate, as turning on the switch completes a circuit of electricity and the electricity causes a filament to heat up such that, when white hot, it emits light.
In education, it is not enough to say that increasing the time spent on reading causes students’ reading to improve; that is naive. What might be required is to know how and why the increase in time devoted to reading improves reading. This opens up many possible causes: motivation; concentration levels and spans; interest level of the materials; empathy between the reader and the material; level of difficulty of the text; purposes of the reading (e.g. for pleasure, for information, for learning, for a test); reading abilities and skills in the reader; subject matter of the text; ambient noise; where, when and for how long the reading is done; prior discussion of, and preparation for, the reading material; follow-up to the reading; choice of reading materials; whether the reading is done individually or in groups; teacher help and support in the reading time; relatedness of the reading to other activities; the nature, contents and timing of the pre-test and post-test; the evidence of improvement (and improvement in which aspects of reading); and so on.
It can be seen immediately in this example of reading that the simple input variable – increasing time for reading – may bring about an improvement in reading, but that may only be one of several causes of the improvement, or an umbrella term, or may liberate a range of other causes to come into play, both direct and indirect causes. Identifying the true cause(s) of an effect is extremely difficult to pin down.
Take for example the introduction of total quality management into schools. Here several interventions are introduced into a school for school improvement, and, at the next school inspection, the school is found to have improved. The problem is trying to decide which intervention(s) has/have brought about the improvement, or which combinations of interventions have worked, or which interventions were counterproductive, and so on. It is akin to one going to the doctor about a digestion problem, where the doctor prescribes six medicines and the digestion problem goes but the patient then contracts a stomach ulcer. Which medicine(s) were responsible for the cure and the ulcer, and in what combinations, or is it actually the medicines that have brought about the cure and/or the ulcer; were there other factors that brought about the cure or the ulcer, or would the digestion problem have been cured naturally?
The researcher has to identify which cause (A) or combination of causes have brought about which effect (B), both intended and unintended, or whether the supposed cause (A) brought about another effect (C), which in turn became the cause of the effect (B) in question, and whether the effect (B) is really the consequence of the supposed cause(s) (A), and not the consequence of something else. What looks like being a simple cause-and-effect actually explodes into a multiplicity of causes and effects (Figure 4.3) (cf. Morrison, 2009: 124).
How, then, can the researcher proceed in trying to uncover causes and effects? A main principle underpinning how some researchers operate here is through control, isolating and controlling all the variables deemed to be at work in the situation. By such isolation and control, one can then manipulate one or more variables and see the difference that they make to the effect. If all the variables in a situation are controlled, and one of these is manipulated, and that changes the effect, then the researcher concludes that the effect is caused by the variable that has been manipulated. Moreover, if the research (e.g. an experiment) can be repeated, or if further data (e.g. survey data) are added, and the same findings are discovered, then this might give added weight to the inferred cause-and-effect connection (though regularity – Hume’s (2000) ‘constant conjunction’ – is no requirement for causation to be demonstrated). This assumes that one has identified, isolated and controlled all the relevant variables, but, as the earlier part of this chapter has suggested, this may be an impossibility.
One way in which the problem of isolation and control of variables is addressed is through randomization – a key feature of the ‘true’ experiment (see Chapter 16). For example, random allocation of individuals to a control group or an experimental group is one widely used means of allowing for the many uncontrolled variables that are part of the make-up of the groups in question (Schneider et al., 2007). It adopts the ceteris paribus condition (all other things being equal) that assumes that the distribution of these many other variables is evenly distributed across the groups, such that there is no need to control for them. This is a bold and perhaps dangerous assumption to make, not least as chaos and complexity theory tell us that small changes and differences can bring about major differences in outcome.
Whilst control is one prime means of trying to establish causation, it does raise several problems of the possibility, acceptability or manageability of isolating and controlling variables, of disturbing and distorting the real work of the participants, and of operating an undesirable – even unethical – control and manipulation of people. This is the world in which the researcher is king or queen and the participants are subjects – subjected to control and manipulation. On the one hand the claim is made that the research is ‘objective’, ‘clean’ (i.e. not affected by the particular factors within each participant), laboratory based and not prone to bias; on the other hand it is a manipulative and perhaps unrealistic attempt to control a world that cannot in truth be controlled. Are there alternatives?
A major alternative is one that keeps the ‘real’ world of participants as undisturbed as possible, avoids the researcher controlling the situation and uses qualitative data to investigate causation. Here observational, interview and ethnographic methods come to the fore, and these are very powerful in addressing the processes of causation and in establishing the causes of an effect as recounted by the participants or the observers themselves. These methods deliberately ‘get inside the heads’ of individuals and groups, as well as including the researcher’s own views, identifying and reporting causation in their terms. They yield considerable authenticity to the causal accounts given or which compile a sufficiently detailed account of a situation for the researcher to make informed comments on the workings of causation in the situation under investigation. Further, it is often the participants themselves who identify what are the causes of effects in the situations being investigated (though the researcher would need to be assured that these are genuine, as participants may have reasons for not disclosing the real causes or motives in a situation or, indeed, may be mistaken).
These two approaches are not mutually exclusive in a piece of research, and, as Chapter 1 has indicated, there is an advantage in adopting a mixed methods approach, or, indeed, in a mixed methodology approach, in which positivist and experimental approaches might yield accounts of the ‘what’ of causation – which variables are operating to produce an effect; whilst an interpretive approach might be used to yield data on the ‘how’ of causation – how the causal processes are actually working.
Researchers examining causes and effects have to decide whether they are researching the effects of causes (e.g. in which they introduce an intervention and see what happens as a consequence) or the causes of effects (e.g. backtracking from an observed situation to try to discover its causes). These are discussed below.
In trying to investigate the effects of one or more causes, the researcher can commence with a theory of causality operating in a situation (e.g. bringing pressure to bear on students causes them to work harder, or dropping out of school reduces income at age 50 by a factor of five, or improving self-esteem improves creativity), operationalize it, and then test it, eliminate rival theories and explanations using data other than those that gave rise to the explanation, and then proceed to the drawing and delimiting of conclusions. The use of continuous rather than categorical variables might be more effective in establishing the nature and extent of causation.
On the other hand, the researcher can proceed along an entirely different track, using qualitative research to really understand the causal processes at work in a situation and in the minds of the participants in that situation – the ‘how’ of causation.
Determining the effects of causes is often undertaken using an interventionist strategy in educational research, installing an intervention either to test a hypothesized causal influence or a causal model, or because it is already known that it may exert a causal influence on effects, i.e. manipulating variables in order to produce effects. (Of course, a non-intervention may also be a cause, for example, I may cause a plant to die by not watering it, i.e. by doing nothing.)
Manipulation takes many forms, including:
action research (discussed in Chapter 18), but which has the problems of rigour brought about by a lack of controls and a lack of external checks such that the attribution of causation may be misplaced;
a range of experimental approaches (discussed in Chapter 16), which assume, perhaps correctly or incorrectly, acceptably or unacceptably, that variables and people can be isolated, controlled and manipulated; and
participant observation in qualitative research. In addressing these approaches, however, serious attention has to be paid to a range of factors:
the context of the intervention and the power of the situation could affect the outcomes and behaviours of participants (the Hawthorne effect or the Lucifer effect (Zimbardo, 2007a));
the same causes do not always produce the same effects;
inappropriate timing of the pre-test and post-test measurements of effects could undermine the reliability of the statement of the effects of the cause;
there is a problem of accuracy, as groups and individuals cannot both be in a group that is and is not receiving an intervention (Holland’s (1986: 947) ‘fundamental problem of causal inference’, which may not be sufficiently attenuated by randomization) (see Chapter 16));
process variables and factors, and not only input variables, as these feature in understanding causation;
the characteristics, personae and specific individual features of participants and their agency, as these influence interventions and their effects.
Experimental techniques, particularly randomized controlled trials (RCT), have considerable potency in establishing causation, and it is here that the identification, isolation and control of independent variables is undertaken, manipulating one independent variable to see if it makes a difference to the outcome. The other variables are held constant and, if a change of outcome is found by manipulating the one independent variable, then the change can be attributed to that independent variable (it becomes the cause), as the other variables have been held constant, i.e. their influence has been ruled out.
In experimental approaches, randomization is an important element in determining causation in order to overcome the myriad range of variables present in, and operating in, participants (the ceteris paribus condition discussed earlier). RCTs and experiments (see Chapter 16) are an example of strongly interventionist approaches that seek to establish the effects of causes by introducing one or more interventions into a situation and observing the outcomes of these under controlled conditions.
However, RCTs are often not possible in education and, indeed, are not immune to criticism. For example, the assumptions on which they are founded may be suspect (e.g. oversimplifying the variables at work in a situation, and overriding the influence of mediating or process variables). They frequently do not establish the causal processes or causal chains that obtain in the situation. They neglect participants’ motives and motivations. They neglect the context in which the action is located, and they might neglect the moral agency of participants and the ethics of researchers. Indeed context can exert a more powerful causal force than the initial causal intervention, as evidenced in the examples of the Stanford Prison Experiment and the Milgram experiments on obedience (see Chapter 26).
Caution must be exercised in supposing that RCTs, the epitome of causal manipulation and the ‘gold standard’ in the determination of the effects of causes, will yield sufficient evidence of causation, as these overlook the significance of context and conditions, of processes, of human intentionality, motives and agency, in short of the contiguous causal connections between the intervention and its putative effects. Indeed, even the issue of when and whether an effect is an effect (short-term to long-term, immediate or delayed) is problematic, and attention has to be given to effects that have been caused by the intervention other than those in which the researcher might be initially interested. For example, a researcher might find that pressuring students to learn improves their mathematics scores but leads to an enduring dislike of mathematics. In the contest of moves towards judging ‘what works’, deciding ‘what works’ is also as much a matter of values and judgement as it is of empirical outcomes of causation. Success is a value judgement, not simply a measure or a matter of performance. Judging ‘what works’ in terms of cause and effect is an incomplete analysis of the situation under investigation. A more fitting question should be ‘what works for whom, under what conditions, according to what criteria, with what ethical justifiability, and with what consequences for participants?’.
A range of issues in judging the reliability and validity of experimental approaches in establishing causation include the acceptability of laboratory experiments that are divorced from the ‘real world’ of multiple human behaviours and actions. Here field experiments and natural experiments (see Chapter 16) may attenuate the difficulties posed by laboratory experiments, though these, too, may also create their own problems of reliability and validity.
As an alternative to action research and experimental methods in determining effects from causes, observational approaches can be used, employing both participant and non-participant approaches (see Chapter 23). Whilst these can catch human intentionality, agency and perceptions of causality and events more fully than experimental methods, nevertheless they encounter the same difficulty as action research and experiments, as they, too, have to provide accounts of causal processes and causal chains. Further, in addressing intentionality and agency in causal processes and chains, it is also possible that whilst perceptions might be correct, they might also be fallacious, partial, incomplete, selective, blind and misinformed. I might think that there is a mouse in the room (a cause), and act on the basis of this (an effect), but, in fact, there may be no mouse in the room.
Interventionist approaches, and the determination of the effects of causes, risk mixing perception with fact, and, regardless of evidence, human inclinations may be to judge data and situations on the basis of personal perceptions and opinions that may fly in the face of evidence (the ‘base rate fallacy’ (Morrison, 2009: 170–1)). This is only one source of unreliability, and it is important to consider carefully what actually are the effects of causes, rather than jumping to statements of causation based on premature evidence of connections.
In determining the causes of effects, the enterprise is even more provisional, tentative and inferential than determining the effects of causes, as data are incomplete and backtracking along causal chains and/or searching within causal nets is difficult, as it requires the searching for clues and testing rival hypotheses about causation. It is possible to generate a huge number of potential causes of observed effects, and the problem is in deciding which one(s) is/are correct. Morrison (2009) suggests that approaches that can be adopted in tracing causes from effects include: (a) variants of ex post facto research (see Chapter 15); and (b) a seven-stage sequence of steps (see below).
In different forms of ex post facto research – quasi-experiments – difficulties arise in their sometimes in ability to control and manipulate independent variables or to establish randomization in the sample.
A seven-stage process of tracing causes from effect can be set out thus:
Stage 1: Establish exactly what has to be explained.
Stage 2: Set out possible theoretical foundations for the investigation.
Stage 3: Examine, evaluate and eliminate rival theoretical foundations, selecting the most fitting.
Stage 4: Hypothesize a causal explanation on the basis of the best theoretical foundation.
Stage 5: Set out the assumptions underlying the causal explanation.
Stage 6: Test the causal hypotheses empirically.
Stage 7: Draw conclusions based on the test.
As this poses several concerns for researchers, a worked example is provided here, from Goldthorpe (2007). Goldthorpe seeks to explain the causes of ‘persistent differentials in educational attainment’ despite increased educational expansion, provision and uptake across the class structure (p. 21), i.e. in the context of increased educational opportunity and its putative weakening influence on class-based determination of life chances.
He proceeds in the seven stages indicated above. Only after that test does he provide a causal explanation for his observed effects.
First, Goldthorpe observes some ‘regularities’ (effects) (p. 45):
a In all economically advanced societies over the previous 50 years there has been an expansion of education provision and in the numbers of children staying on in full-time education beyond the minimum schooling age (e.g. going into higher education).
b At the same time, class differentials in educational attainment have remained stubbornly stable and resistant to change, i.e. though children from all classes have participated in expanded education, class origins and their relationship to the likelihood of children staying on in education or entering higher education has only reduced slightly, if at all, and this applies to most societies.
He is establishing social regularities that any causal and theoretical account should seek to explain: the creation, persistence and continued existence of class stratification in modern societies, and the continuing classrelatedness of educational inequality and life chances (p. 24).
Goldthorpe’s work is premised on the view that theories are necessary to provide explanatory foundations for how established regularities came to be as they are (p. 21). He initially suggests four theoretical foundations: Marxist theory, liberal theory, cultural theory and rational choice theory.
For several reasons which he gives (pp. 22–34), he rejects the first three of these and argues that rational choice theory provides a fitting theoretical foundation for his investigation of the causes of the effects observed (pp. 34–41). True to rational action theory, Goldthorpe (2007: 31) places emphasis on aspirations, in particular noting their relative rather than their absolute status, that is to say, aspirations are relative to class position, as working-class aspirations may not be the same as those of other classes. Different social classes have different levels and kinds of aspiration, influenced – as rational action theory suggests – by the constraints under which they operate, and the perceived costs and benefits that obtain when making decisions (p. 32). Taking relative rather than absolute views of aspiration enable accounts to be given that include the fact of increased provision of, and participation in, education by students from all classes, i.e. class differentials have not widened as education provision and participation have widened.
Goldthorpe (2007: 32) suggests that cultural theory may account for what Boudon (1973) terms ‘primary effects’, that is initial levels of achievement and ability in the early stages of schooling. However Goldthorpe is more concerned with Boudon’s ‘secondary effects’, that is those effects that come into play when children reach branching points (transition points, e.g. from primary to secondary schooling, from secondary education to university) (p. 32) and which have increasingly powerful effects as one progresses through schooling. ‘Secondary effects’ take account of the aspirations and values that children and their parents hold for education, success and life options, i.e. the intentionality and agency of rational action theory in a way that ‘primary effects’ do not. Goldthorpe notes that at each successive ‘branching point’, children from more advantaged backgrounds remain in the educational system and those from less advantaged backgrounds either leave school or choose courses that lead to lower qualifications (hence reducing their opportunities for yet further education).
Goldthorpe (2007: 33) argues that more ambitious options may be regarded less favourably by those from less advantaged class backgrounds as they involve: (a) greater risk of failure; (b) greater cost; and (c) relatively less benefit. In other words, the level of aspiration may vary according to class and the associated levels of assessed cost and risk by members of different classes, and children from less advantaged backgrounds have to be more ambitious than those from more advantaged backgrounds if they are to meet the aspirations and success levels of those from more advantaged backgrounds. Class origins influence risk assessment, cost assessment and benefit assessment – all aspects that are embraced in rational action theory. These determine the choices made by children and their parents.
Goldthorpe (2007: 34) argues that class differentials in educational attainment have persisted because even though there has been expansion and reform of education, and even though the overall costs and benefits that are associated with having more ambitious options have encouraged their take-up, in practice, there has been little concurrent change in the ‘relativities between class-specific balances’: different classes view the costs, risks and benefits differently (p. 34). This is his working hypothesis in trying to establish cause from effect.
Goldthorpe tests his theory initially by drawing attention to the ongoing income differentials between classes; indeed he argues that they have widened (p. 35), with manual labourers more prone to unemployment than professional or managerial workers, i.e. the costs of education are still a factor for less advantaged families, particularly at the end of the period of compulsory schooling. At the time when their children come to the end of compulsory schooling, the income of manual workers will already have peaked (e.g. when they are in their forties), whereas for professional and managerial workers it will still be rising, that is costs are more of a problem for manual workers than for professional and managerial workers, i.e. the costs of higher education relative to income, and the consequent effects on family lifestyle if families are having to finance higher education, are much higher for manual workers. This increased proportion of family income to be spent on education for less advantaged families is coupled with the fact that if children from these families are to succeed, then they need even more ambition than their professional and managerial class counterparts, that is they are at a potential double disadvantage, i.e. relative advantage and disadvantage are not disturbed, a feature on which liberal theory is silent (p. 36).
Goldthorpe makes the point that class position conditions educational decisions made by members of different classes. These different class positions influence different evaluations of the costs and benefits of education, and these are socially reproductive, i.e. the social class position is undisturbed.
Another element of his argument concerns risk aversion. His view is that a major concern of members of different classes is to minimize their risk of downward class mobility, and to maximize their chances for upward class mobility or, at least, maintenance of their existing class location (p. 37). This exerts greater pressure on the already-advantaged classes (e.g. the salariat) to have their children complete higher education (in order to preserve intergenerational class stability) than it does on the children from less advantaged classes (e.g. the waged). It costs more for the children of the advantaged classes to preserve their class position than it does for children of the less advantaged classes to preserve theirs.
With regard to families in the less advantaged classes, Goldthorpe (2007: 38) suggests that they regard higher education in a much more guarded light. Not only does it cost less for them to maintain their class position, but it costs relatively more to achieve upward class mobility; their best options might be for vocational education, as it is cheaper and gives a strong guarantee of not moving downwards in class situation (e.g. to be unemployed or unskilled).
Further, for children in this class, the costs (and likelihood) of failure in higher education could be proportionately greater than those for children from more advantaged families. For example in terms of: the relative costs of the higher education; lost earning time; lost opportunity to follow a vocational route in which they have greater likelihood of being successful (p. 38); loss of social solidarity if working-class children pursue higher education, the consequences of which may be to remove them from their class origin and community (pp. 38–9). These factors combine to suggest that children and families from less advantaged backgrounds will require a greater assurance, or expectation, of success in higher education before committing themselves to it, than is the case for children and families from more advantaged backgrounds (p. 68).
Goldthorpe then offers his causal explanation of the effects observed: the persistence of class differentials in educational attainment despite expansion of educational provision and participation (p. 39):
1 Class differentials in the uptake of more ambitious educational options remain because the conditions also remain in which the perceived costs and benefits of these options operate, and these lead to children from less advantaged families generally requiring a greater assurance of success than children from more advantaged families before they (the former) pursue more ambitious educational options.
2 There is a rational explanation for the persistence of these different considerations of ambitious options by class over time, which is rooted in class-based conditions.
These are the two main hypotheses that he seeks to test.
Goldthorpe then proceeds to test his two hypotheses (pp. 39–44, 53–6, and his chapters 3 and 4), adducing evidence concerning several factors, e.g.
the greater sensitivity of working-class families to the chances of success and failure in comparison to middle-class families (p. 40);
different levels of ambition in working-class and middle-class families (p. 40);
relative (class-based) risk aversion in decision making, e.g. the risk of failure and/or of closing options (pp. 55–6);
the loss of foregone earnings (pp. 53–5);
expectations of success (pp. 55–6);
evaluation of the potential benefits, value and utility of higher education (pp. 38–9);
influences on choices and decision making in different classes (his chapter 3);
actual choices made by members of different classes;
fear of downward social mobility (pp. 53–4);
the need to preserve, or improve on, intergenerational mobility (pp. 53–4);
financial costs (p. 56).
Goldthorpe indicates that students from lower socioeconomic groups either cannot afford, or cannot afford to take risks in, higher education, and he identifies three clusters of possible explanations of persistence of class differentials in educational attainment, including (but not limited to):
Cluster 1: Differences in aspirations and decisions are caused by perceptions of costs: (a) loss of earnings during study time (a bigger drawback for families and students from low income households than for those from privileged backgrounds); (b) students from low-income households have to work harder than privileged students in order to compete with them; (c) students from low-income households must have greater ambition than privileged students in order to be successful in a higher social class; (d) the financial costs of higher education (HE), proportional to income, are higher for less advantaged students than for more advantaged students and families.
Cluster 2: Differences in aspiration and decisions are caused by relative risk aversion: (a) the risk of failure in HE is greater for students from disadvantaged classes; (b) the risk of loss of further educational opportunities if failure ensues or incorrect options are
followed is greater for students from disadvantaged classes than for students from more privileged classes; (c) the risk of loss of social solidarity is greater for students from working-class groups than for students from more privileged classes; (d) less advantaged students must have greater ambition than privileged students in order to be successful in higher social classes.
Cluster 3: Differences in aspiration and decisions are caused by perceptions of relative benefit: (a) the opportunity for upward social mobility through HE is an attraction for students from lower class backgrounds; (b) HE is differentially necessary for preferred or likely employment for those from privileged and less privileged groups.
Goldthorpe indicates that class differentials have, indeed, continued to affect the take-up of educational options. He finds that class differentials in terms of the take-up of more ambitious educational options have been maintained because so too have the conditions in which the perceived costs and benefits of these options lead to children from less advantaged families requiring, on average, a greater assurance of success than their more advantaged counterparts before they decide to pursue such options. There are class differences in terms of relative ambition, risk aversion, perceived costs and benefits, amounts of effort required, assurances of success (and the significance of this), fear of downward social mobility, income, occupational choices and the need for qualifications.
He concludes that the results of empirical tests support his explanation of the factors of relative risk aversion and fear of downward social mobility exerting causal power on educational decision making which, in turn, lead to class differentials in educational attainment being maintained (p. 99).
Goldthorpe argues that this hypothesis is better supported than alternative hypotheses (e.g. educational choices being predetermined by culture, class identity and the class structure).
This example here offers a robust account of how to track backwards from an effect to a cause and how to evaluate the likelihood that the putative cause of the effect actually is the cause of that effect. In summary, for researchers seeking to establish the causes of effects, the task has several aspects:
Indicate what needs to be done to test the theory and to falsify it.
Identify the kinds of data required for the theory to be tested.
Identify the actual data required to test the theory.
Identify the test conditions and criteria.
Construct the empirical test.
Consider the use of primary and secondary data.
Consider using existing published evidence as part of the empirical test.
Ensure that action narratives and intentionality are included in causal accounts.
The fundamental problem in determining causes from effects is the uncertainty that surrounds the status of the putative cause; it can only ever be the best to date, and the researcher does not know if it is the best in absolute terms. One effect stems from many causes, and to try to unravel and support hypotheses about these may present immense difficulties for the researcher. Morrison (2009: 204) suggests that there are several ways in which causes may be inferred from effects:
recognizing that a high level of detail may be required in order to establish causation: high granularity;
identifying several causal chains, mechanisms and processes in a situation;
combining micro-and macro-levels of analysis;
addressing both agency and structure;
underpinning the data analysis and causal explanation with theory;
using different kinds of ex post facto analysis;
using correlational and causal-comparative, criterion group analysis;
ensuring matching of groups in samples and that similar causes apply to both groups;
adopting the seven-stage process set out above, of generating, testing and elimination of hypotheses and rival hypotheses;
ensuring clarity on the direction of causation;
using empirical data to test the causal explanation;
identification of which is cause and which is effect, and/or which effect then, subsequently, becomes a cause;
avoiding the problem of over-selective data;
ensuring that the data fairly represent the phenomenon under investigation;
recognizing that cause and effect may be blurred;
accepting that effects may become causes in a cyclical sequence of causation;
seeking out and recognizing over-determination at work in causal accounts;
keeping separate the explanans (the explanation) from the explanandum (that which is to be explained);
ensuring that alternative theories and causal explanations are explored and tested;
drawing conclusions based on the evidence, and the evidence alone.
In seeking to establish the causes of effects, there is a need to review and test rival causal theories and to retain those with the greatest explanatory potential and which fit the evidence most comprehensively and securely. Testing of rival hypotheses must be done with data that are different from those that gave rise to the hypotheses, in order to avoid circularity.
The determination of causes from effects does not have the luxury afforded to causal manipulation available in determining effects from causes. Whilst this renders the determination of causes from effects more intractable, nevertheless this is not to say that it cannot be attempted or achieved, only that it is difficult.
Morrison (2009) argues that, in seeking to identify the causes of effects, there is a need for a theoretical foundation to inform causal explanation. Possible causal explanations should be evaluated against rival theories and rival explanations, being operationalized in considerable detail (high granularity), and tested against data that are different from those that gave rise to the causal explanation. Causal explanations should link micro-and macro-factors, include agency and intentionality as well as structural constraints, and contain a level of detail that is sufficiently high in granularity to explain the phenomenon to be explained without concealing or swamping the main points with detail overload, that is the researcher must be able to distinguish the wood from the trees.
In approaching causal research, then, the researcher is faced with a range of challenges, including for example (Morrison, 2009: 213–14):
focusing more on causal processes than input/output/results models of causation;
establishing causation other than through reduction and recombination of atomistic, individual items and elements;
regarding causation as the understanding of the emergent history of a phenomenon or a whole;
investigating multiple and simultaneous causes and their multiple and simultaneous effects in a multiply-connected and networked world;
separating causation from predictability, and drawing the boundaries of predictability for an understanding of the frequent uniqueness of a causal sequence, that may not be repeatable, i.e. living with uncertainty and with unpredictability;
learning to work with causation in a situation in which randomness often ‘trumps’ causation (Gorard, 2001a: 21);
indicating the utility of an understanding of causation if it has little subsequent predictive strength;
understanding how to investigate causation in holistic webs of connections, i.e. how it is possible to discover or demonstrate causation when looking at events holistically;
understanding causation and causal processes in a multi-causal, multi-effect, non-linear and multiply-connected world;
identifying the causal processes at work in determining social and macro-structures from the actions and interaction of individuals (the micro-worlds) and, conversely, in determining the actions and interactions of individuals from the structures of society and its institutions (the macro-worlds), their ontologies and epistemologies.
The researcher has to decide whether the research is investigating the cause of an effect, the effect of a cause, or both, and when causation is demonstrated, given that absolute certainty is illusory. If one is investigating the effects of causes then the methodologies and approaches to be used might include experiments, action research, survey analysis, observational approaches, or a combination of these (and others). If one is investigating the causes of effects then, in the context of the likelihood of greater uncertainty than in establishing the effects of causes, one can employ numerical and qualitative data in backtracking from effects to causes and in testing hypothesized causes of effects. In all these approaches this chapter has suggested that probabilistic rather than deterministic causation is a more fitting description of the nature of the conclusions reached. It has suggested that, even if it sounds simplistic at first, nevertheless it is both important yet difficult to establish what actually constitutes a cause and an effect. The chapter has suggested that causal processes, with high granularity, are often closer to identifying the operations of causes and effects and the links between them, and that here qualitative data might hold pre-eminence in educational research. However, the chapter has also suggested that there is an important role for numerical approaches, for examining the ‘regularities’ that might be evidenced in survey approaches, and in the isolation and control of variables in experimental approaches. In short, the chapter is arguing for the power of mixed methodologies and mixed methods in investigating and establishing causation.
Companion WebsiteThe companion website to the book includes PowerPoint slides for this chapter, which list the structure of the chapter and then provide a summary of the key points in each of its sections. This resource can be found online at www.routledge.com/textbooks/cohen7e.
