Complex Theories of Change: Recipes for failure or for learning?

The diagram below is a summary of a Theory of Change for interventions in the education sector in Country X. It did not stand on its own, it was supplemented by an extensive text description.

Its complex in the sense that there are many different parts to it and many interconnections between them, including some feedback loops. It seems realistic in the sense of capturing some of the complexity of social change. But it may be unrealistic if it is a prescription for achieving change. Whether it is the later depends on how we interpret the diagram, which I discuss below.

One way of viewing the Theory of Change is in terms of conditions (the elements in the diagram) that may or may not be necessary and/or sufficient for the final outcome to occur. The ideas of necessary and/or sufficient causal conditions are central to the notion of “configurational” models of causation, described by Mahoney and Goertz (2012) and others. A configuration is a set of conditions that may be either sufficient or necessary for an outcome e.g. Condition X + Condition T + Condition D + Condition P -> Outcome. This is in contrast to simpler notions of an outcome having a single cause e.g. Condition T -> Outcome.

The philosopher John Mackie (1974) argued that most of the “causes” that we talk about in everyday life are what are called INUS causes. That is, they are about a condition that is an Insufficient but Necessary part of a configuration of conditions but one which is Unnecessary but Sufficient for an outcome to occur. For example, smoking is a contributory cause of lung cancer, but it is neither necessary nor sufficient to get cancer. There are other ways of getting cancer and all smokers do not get cancer.

The interesting question for me is whether the above Theory of Change represents one or more than one causal configuration. I look at both possibilities and their implications.

If the Theory of Change represents a single configuration then each element, such as “More efficient management of teacher recruitment and deployment”, would be insufficient by itself, but a necessary part of the whole configuration. In other words, every element in the Theory of Change has to work or else the outcome won’t occur. This is quite a demanding expectation. The more complex this “single configuration” model becomes (i.e. by having more conditions), the more vulnerable it will becomes to implementation failure, because even if only part does not work, the whole process will fail. One saving grace is that it would be relatively easy to test this kind of theory. In any locations where the outcome did occur it would be expected that all elements would be present. If some were not, then the missing elements would not qualify as insufficient but necessary conditions.

 The alternative perspective is to see the above Theory of Change as representing multiple causal configurations i.e. multiple possible combinations of conditions, each of which can lead to the desired outcome. So any condition, again such as “More efficient management of teacher recruitment and deployment” may not be necessary under all circumstances. Instead it may be insufficient but necessary part of one of the configurations, but not the others. Viewed from this perspective, the Theory of Change seems less doomed to implementation failure, because there is more than one route to success.

However if there are multiple routes the challenge is then how to identify the different configurations that may be associated with successful outcomes. As it stands the current Theory of Change gives little guidance. Like many Theory of Change at this macro-level / sector perspective it tends towards showing “everything connected to everything”. In fact this limitation seems unavoidable, because with increasing scale there is often a corresponding increase in the diversity of actors, interventions and contexts. In such circumstances there are likely to be many more causal pathways at work. This view suggests that at such a macro level it might be more appropriate for a Theory of Change to initially have relatively modest ambitions and to limit itself to identifying the conditions that are likely to be involved in the various causal configurations.

The focus then would move to on what can be done through subsequent monitoring and evaluation efforts. This could involve three tasks: (a) Identifying where the outcomes have and have not occurred, (b) identifying how they differed in terms of the configuration of conditions that were associated with the outcomes (and absent where the outcomes did not occur). This would involve across-case comparisons. (c) Establishing plausible causal linkages between the observed conditions within each configuration. This would involve within-case analyses. Ideally, the overall findings about the configurations involved would help ensure the sustainability and replicability of the expected outcomes.

The Theory of Change will still be useful in as much as it successfully anticipates the various conditions making up the configurations associated with outcomes, and their absence. It will be less useful if it has omitted many elements, or included many that are irrelevant. Its usefulness could actually be measured! Going back to the recipe metaphor in the title, a good Theory of Change will have at least an appropriate list of ingredients but it will be really up to subsequent monitoring and evaluation efforts to identify what combinations of these produce the best results and how they do so (e.g. by looking at the causal mechanisms connecting these elements).

Some useful references to follow up:
• Causality for Beginners, Ray Pawson, 2008
• Qualitative Comparative Analysis, at Better Evaluation
• Process Tracing, at Better Evaluation
• Generalisation, at Better Evaluation

Postscript:

I have just read Owen Barder's review of Ben Ramalingam's new book "Aid on the Edge of Chaos" In that review he makes two comments that are relevant to the argument presented above:

"As Tim Harford showed in his book Adapt, all successful complex systems are the result of adaptation and evolution.  Many in the world of development policy accepted intellectually the story in Adapt but were left wondering how they could, practically and ethically, manage aid projects adaptively when they were dealing with human lives"

"Managing development programmes in a complex world does not mean abandoning the drive to improve value for money. Iteration and adaptation will often require the collection of more data and more rigorous analysis - indeed, it often calls for a focus on results and 'learning by measuring' which many people in development may find uncomfortable."

The point made in the last paragraph about requiring the collection of more data needs to be clearly recognised, as early as possible. Where there are likely to be many possible causal relationships at work, and few if any of these can be confidently hypothesised in advance, the coverage of data collection will need to be wider. Data collection (and then analysis) in this situation is like casting a net onto the waters, albeit still with some idea of where the fish may be. The net needs to be big enough to cover the possibilities.

Measuring the impact of ideas: Some testable propositions

Evaluating the impact of research on policy and practice can be quite a challenge, for at least three reasons: (a) Our ideas of the likely impact pathways may be poorly developed, (b) Actors within those pathways may not provide very reliable information about exposure to and use of the research we are interested in. Some may be over-obliging, others may be very reluctant to acknowledge its influence. Others may not even be concious of the influence that did occur, (c) It is quite likely that that there are many more pathways through which the research results travel that we cant yet imagine, let alone measure. Even more so when we are looking at impact over a longer span of time. When I look back to the first paper I wrote about MSC, which I put on the web in 1996, I could never have imagined the diversity of users and usages of MSC that have happened since then.

I am wondering if there is a proxy measure of impact that might be useful, and whose predictive value might even be testable, before it is put to work as a proxy. A proxy is conventionally defined as "a person authorized to act on behalf of another". In this case it is a measure that can be justifably used in place of another, because that other measure is not readily available.

What would that proxy measure look like? Lets start with an assumption that the more widely dispersed an idea is, the more likely someone will encounter it, if only by chance, and then make some use of it. Lets make a second assumption, that impact is greater when not only is the idea widely dispersed, say amongst 1000 people rather than 100, but when it is dispersed amongst a wide variety of people, not just one kind of people. Combined together, the proxy measure could be descirbed as Availability.

While one can imagine some circumstances where  impact will be bigger when the idea is widely dispersed but within a single type of people I would argue the success of these more "theory led" predictions will often be outnumbered by serindipitous encounters and impact, especially where there has been large scale dissemination, as will often be the case when research is disseminated via the web. This is a view that could be tested, see below.

How would the proxy measure be measured? As suggested by the assumptions above, Availability could be tracked using two measures. One is the number of references to the research that can be found (e.g. on the web), which we could call Abundance. The other is the Diversity of sources that make these references. The first measure seems relatively simple. The second, the measurement of diversity, is an interesting subject in its own right , and one which has been widely explored by ecologists and other disciplines for some decades now (For a summary of ideas, see Scott Page - Diversity and Complexity, 2001, chapter 2). One simple measure is Simpson's Reciprocal Index (1/D), which combines Richness ( the number of species [/ number of types of reference sources]) and Evenness, the relative abundance of species [/number of references] across those types). High diversity is a combination of high Richness and high Evenness (i.e. all species are similarly abundant). A calculation of the index is shown below:

How could the proxy measure be tested, before it can be widely used? We would need  a number of test cases where not only can we measure the abundance and diversity of references to a given piece of research, but we can also access some known evidence of impact(s) of that research. With the latter we may be able to generate a rank ordering of impact, through a pair comparison process - a process that can acknowledge the differences in the kinds of impact. We could then use data from these cases to identify which of the following distributions existed:

We could also compare cases with different combinations of abundance and diversity. It is possible that abundance is all that matters and diversity is irelevant.

Now, does anyone have a set of cases we could look at, to test the propositions outlined above?

Postscript: There are echoes of evolutionary theory in this proposal. Species that are large in number and widely dispersed, across many different habitats, tend to have better long term survival prospects in the face of changing climates and the co-evolution of competitors

A reverse QCA?

I have been talking to a project manager who needed some help clarifying their Theory of Change (and maybe the project design itself). The project aims to improve the working relationships between a particular organisation (A) and a number of organisations they work with (B). There is already a provisonal scale that could be used to measure the baseline state of relationships, and changes in those relationships thereafter. Project activities designed to help improve the relationships have already been identified and should be reasonably easy to monitor. But the expected impacts of the improved relationships on what B's do elsewhere via their other relationships have not been clarified or agreed to, and in all likelihood they could be many and varied. It will probably be easier to identify and categorise after the activities have been carried out, rather than during at any planning stage.

I have been considering the possible usefullness of QCA as a means of analysing the effectiveness of the project. The cases will be the various relationships between A and Bs that are assisted in different ways. The conditions will be different forms of assistance provided as well as differences in the context of these relationships (e.g. the people, organisations and communities involved). The outcome of interest will be the types of changes in the relationships between A and Bs. Not especially problematic, I hope.

Then I thought..., perhaps one could do a reverse QCA analysis to identify associations between specific types of relationship changes and the many different kinds of impacts that were subsequently observed on other relationships. The conditions in this analysis would be various categories of observed change (with data on their presence and absence). The configurations of conditions identified by the QCA analysis would in effect be a succinct typology of impact configurations associated with each kind of relationship change. As distinct from causal configurations sought via a conventional QCA.

This reversal of the usual QCA analysis should be possible and legitimate because relations between conditons and outcomes are set theoretic relations, not temporal relationships. My next step, will be to find out if someone has already tried to do this elsewhere (that I could learn from). These days this is highly likely.

Postscript 1: The same sort of reverse analyses could be done with Decision Tree algorithms, whose potential for use in evaluations has been discussed in earlier postings on this blog and elsewhere.

Postscript 2: I am slowly working my way through this comprehensive account of QCA, published last year:

Schneider, Carsten Q., and Claudius Wagemann. 2012. Set-Theoretic Methods for the Social Sciences: A Guide to Qualitative Comparative Analysis. Cambridge University Press.

Another perspective on the uses of control groups

I have been reading Eric Siegel's book on Predictive Analytics. Though it is a "pop science" account, with the usual "this will change the world" subtitle, it is definitely a worthwhile read.

In chapter 7 he talks the reader through what are called "uplift models", which are Decision Tree models that can not only differentiate groups who respond differently to an intervention, but how much differently when compared to a control group where there is no intervention. All this is in the context of companies marketing their products to the population at large, not the world of development aid organisations.

(Temporarily putting aside the idea of uplift models...) In this chapter he happens to use the matrix below, to illustrate the different possible sets of consumers that exist, given two types of scenarios that can be found where both a control and intervention group are being used.

But what happens if we re-label the matrix, using more development project type language? Here is my revised version below:
 

Looking at this new matrix it struck me that evaluators of development projects may have a relatively improverished view of the potential uses of control groups. Normally the focus is on the net difference in the improvement, between households in the control and intervention groups: How big is it and is it statistically significant? In other words, how many of those in the intervention group were really "self-helpers" who would have improved anyway, versus being "Need help'ers" who would not have improved without the intervention.

But this leaves aside two other sets of households who also surely deserve at least equal attention.One are the "hard cases", that did not improve in either setting. Possibly the poorest of the poor. How often are their numbers identified with the same attention to statistical detail? The other are the "Confused", who have improved in the control group, but not in the intervention group. Perhaps these are the ones we should really worry about, or at least be able to enumerate. Evaluators are often asked, in their ToRs, to also give attention to negative project impacts, but how often do we systematically look for such evidence?

Okay, but how will we recognise these groups? One way is to look at the distribution of cases that are possible. Each group can be characterised by how cases are distributed in the control and intervention group, as shown below. The first group (in green) are probably "self-help'ers" because the same proportion also improved in the control group. The second group are more likely to be "need-help'ers" because fewer people improved in the control group. The third group are likely to be the "confused" because more of them did not improve in the intervention group than in the control group. The fourth group are likely to be the "hard cases" if the same high proportion did not improve in the control group either.

At an aggregate level only one of the four outcome combinations shown above can be observed at any one time. This is the kind of distribution I found in the data set collected during a 2012 impact assessment of a rural livelihoods project in India. Here the overall distribution suggests that the “need-helpers” have benefited. 

How do we find if and where the other groups are? One way of doing this is to split the total population into sub-groups, using one household attribute at a time, to see what difference it makes to the distribution of results. For example, I thought that household’s wealth ranking might be associated with differences in outcomes. So I examined the distribution of outcomes for the poorest and least poor of the four wealth ranked groups. In the poorest group, those who benefited were the “need-help’ers” , but in the “Well-Off” group those who benefited were the “self-help’ers”, perhaps as expected

There are still the two other kinds of outcomes that might exist in some sub-groups – the “hard cases” and the “confused” How can we find where they are? At this point my theory-directed search fails me. I have no idea where to look for them. There are too many household attributes in the data set to consider manually examining how different their particular distribution of outcomes is from the aggregate distribution.

This is the territory where an automated algorithm would be useful. Using one attribute at a time, it would split the main population into two sub-groups, and search for the attribute that made the biggest difference. The difference to look for would be extremity of range, as measurable by the Standard Deviation.  The reason for this approach is that the most extreme range would be where one cell in the control group was 100 and the other was 0, and similarly in the intervention group. These would be pure examples of the four types of outcome distributions shown above. [Note that in the two wealth ranked sub-groups above, the Standard Deviation of the distributions was 14% and 15% versus 7% in the whole group]

This is the same sort of work that a Decision Tree algorithm does, except Decision Trees usually search for binary outcomes and use different “splitting” criteria. I am not sure if they can use the Standard Deviation, or if they can use a another measure which would deliver the same results (i.e. identify four possible types of outcomes).

Predicting evaluability: An example application of Decision Tree models

The project: In 2000 ITAD did an Evaluablity Assessment of Sida funded democracy and human rights projects in Latin America and South Africa. The results are available here:Vol.1 and Vol.2. Its a thorough and detailed report.

The data: Of interest to me were two tables of data, showing how each of the 28 projects were rated on 13 different evaluablity assessment criteria. The use of each of these criteria are explained in detail in the project specific assessments in the second volume of the report.

Here are the two tables. The rows list the evaluability criteria and the columns list the projects that were assessed. The cell values show the scores on each criteria: 1 = best possible, 4 = worst possible. The bottom row summarises the scores for each project, and assumes an equal weighting for each criteria, except for the top three, which were not included in the summary score.

CLICK ON THE TABLE TO VIEW AT FULL SIZE

The question of interest: Is it possible to find a small sub-set of these 13 criteria which could act as good predictors of likely evaluability? If so, this could provide a quicker means of assessing where evaluablity issues need attention.

The problem: With 13 different criteria there are conceivably 2 to the power of 13 possible combinations of criteria that might be good predictors i.e 8,192 possiblities

The response:  I amalgamated both tables into one, in an Excel file, and re-calculated the total scores, by including scores for the first three criteria (recoded as Y=1, N=2). I then recoded the aggregate score into a binary outcome measure, where 1 = above average evaluablity scores and 2 below average scores.

I then imported this data into Rapid Miner, an open source data mining package. I then used the Decision Tree module within that package to generate the following Decision Tree model, which I will explain below.

 CLICK ON THE DIAGRAM TO VIEW AT FULL SIZE

The results: Decision Tree models are read from the root (at the top) to the leaf, following each branch in turn.

This model tells us, in respect to the 28 projects examined, that IF a project scores less than 2.5 (which is good) on "Identifiable outputs"  AND if it scores less than 3.5 on "project benefits can be attributed to the project intervention alone"  THEN there is a 93% probability that the project is reasonably evaluable (i.e has above average aggregate score for evaluability in the original data set). It also tells us that 50% of all the cases (projects) meet these two criteria.

Looking down the right side of the tree we see that IF the project scores more than 2.5 (which is not good) on"Identifiable outputs" AND even though it scores less than 2.5 on "broad ownership of project purpose amongst stakeholders THEN there is a 100% probability that the project will have low evaluability. It also tells us that 32% of all cases meet these two criteria.

Improvements: This model could be improved in two ways. Firstly, the outcome measure, which is an above/below average aggregate score for each project could be made more demanding, so that only the top 25th percentile were rated as having good evaluability. We may want to set a higher standard.

Secondly, the assumption that all criteria are of equal importance, and thus their scores can simply be added up, could be questioned. Different weights could be given to each criterion, according to their perceived causal importance (i.e. the effects they will have). This will not necessarily bias the Decision Tree model towards using those criteria in a predictive model. If all projects were rated highly on a highly weighted criteria that criteria would have no particular value as a means of discriminating between them, so it would be unlikely to feature in the Decision Tree at all.

Weighting and perhaps subsequent re-weighting criteria may also help reconcile any conflict between what are accurate prediction rules and what seems to make sense as a combination of criteria that will cause high or low evaluability. For example in the above model, it seems odd that a criteria of merit (broad ownership of project purpose) should help us identify projects that have poor evaluablity.

Your comments are welcome

PS: For a pop science account of predictive modelling see Eric Siegel's book on Predictive Analytics

My two particular problems with RCTs

Up till now I have tried not to take sides in the debate, when crudely cast as between those "for" and those "against" RCTs (Randomised Control Trials)  I have always thought that there are "horses for courses" and that there is a time and place for RCTs, along with other methods, including non-experimental methods, for evaluating the impact of an intervention. I should also disclose that my first degree included a major and sub-major in psychology, much of which was experimental psychology. Psychologists have spent a lot of time thinking about rigorous experimental methods. Some of you may be familiar with one of the more well known contributors to the wider debates about methodology in the social sciences - Donald T Campbell - a psychologist whose influence has spread far beyond psychology. Twenty years after my first degree, his writings on epistemology subsequently influenced the direction of my PhD, which was not about experimental methods. In fact it was almost the opposite in orientation - the Most Significant Change (MSC) technique was one of its products.

This post has been prompted by my recent reading of two examples of RCT applications, one which has been completed and one which has been considered but not yet implemented. They are probably not exemplars of good practice, but in that respect they may still be useful, because they point to where RCTs should not be used. The completed RCT was of a rural development project in India. The contemplated RCT was on a primary education project in a Pacific nation. Significantly, both were large scale projects covering many districts in India and many schools in the Pacific nation.

Average effects

The first problem I had is with the use of the concept of Average Treatment Effect (ATE) in these two contexts. The India RCT found a statistically significant difference in the reduction in poverty of households involved in a rural development project, when compared to those who had not been involved. I have not queried this conclusion. The sample looked decent in size and the randomisation looked fine. The problem I have is with what was chosen as the "treatment" The treatment was the whole package of interventions provided by the project. This included various modalities of aid (credit, grants, training) in various sectors (agriculture, health, education, local governance and more) It was a classic "integrated rural development project, where a little bit of everything seemed to be on offer, delivered partly according to the designs of the project managers, and partly according to beneficiary plans and preferences. So, in this context, how sensible is it to seek the average effects on households of such a mixed up salad of activities? At best it tells us that if you replicate this particular mix (and God knows how you will do that...) you will be able to deliver the same significant impact on poverty. Assuming that can be done, this must still be about the most inefficient replication strategy available. Much more preferable, would be to find which particular project activities (or combinations thereof) were more effective in reducing poverty, and then to replicate those.

Even the accountability value of the RCT finding was questionable. Where direct assistance is being provided to households a plausible argument could be made that process tracing (by a decent auditor) would provide good enough assurance that assistance was reaching those intended. In other words, pay more attention to the causal "mechanism"

The proposed RCT of the primary education project had similar problems, in terms of its conception of a testable treatment. It proposed comparing the impact of two project "components", by themselves and in combination. However, as in India, each of these project components contained a range of different activities which would be variably made available and variably taken up locally across the project location.

Such projects are commonplace in development aid. Projects focusing on a single intervention, such as immunization or cash transfers are the exception, not the rule. The complex design of most development projects, tacitly if not explicitly, reflects a widespread view that promoting development involves multiple activities, whose specific composition often needs to be localised.

To summarise: It is possible to calculate average treatment effects, but its is questionable how useful that is in the project settings I have described - where there is a substantial diversity of project activities and combinations thereof


Context

Its commonplace amongst social scientists, especially the more qualitatively oriented, to emphasis the importance of context. Context is also important in the use of experimental methods, because it is a potential source of confounding factors, confusing the impact of a independent variable under investigation.

There are two ways of dealing with context. One is by ruling it out e.g. by randomising access to treatment so that historical and contextual influences are the same for intervention and control groups. This was done in both the India and Pacific RCT examples. In India there were significant caste and class variations that could have influenced project outcomes. In the Pacific there were significant ethnic and religious differences. Such diversity often seems to be inherent in large scale development projects.

The result of using this ruling-out strategy is hopefully a rigorous conclusion about the effectiveness of an intervention, that stands on its own, independent of the context. But how useful will that be? Replication of the same or similar project will have to take place in a real location where context will have its effects. How sensible is it to remain intentionally ignorant of those likely effects?

The alternative strategy is to include potentially relevant contextual factors into an analysis. Doing so takes us down the road of a configurational view of causation, embodied in the theory-led approaches of Realist Evaluation and QCA, and also in the use of data mining procedures that are less familiar to evaluators (Davies, 2012).

Evaluation as the default response

In the Pacific project it was even questionable if an evaluation spanning a period of years was the right approach (RCT based or otherwise). Outcomes data, in terms of student participation and performance data will be available on a yearly basis through various institutional monitoring mechanisms. Education is an area where data abounds, relative to many other development sectors, notwithstanding the inevitable quality issues. It could be cheaper, quicker and more useful to  develop and test (annually) predictive models of the outcomes of concern. One can even imagine using crowdsourcing services like Kaggle to do so. As I have argued elsewhere we could benefit by paying more attention to monitoring, relative to evaluation.

In summary, be wary of using RCTs where development interventions are complex and variable, where there are big differences in the context in which they take place, and where an evaluation may not even be the most sensible default option.