Wednesday, October 20, 2010

Counter-factuals and counter-theories

Thinking about the counter-factual means thinking about something that did not happen. So consider a project involving the provision of savings and credit services, with the expectation of reducing levels of poverty amongst the participating households. The counter-factual is the situation where the savings and credit services were not provided. This can either be imagined, or monitored through the use of a control group, which is a group of similar households in a similar context.

In the course of 20 years work on monitoring and evaluation of development aid projects I have only come across one good opportunity to analyse changes in household poverty levels through the comparison of participating and non-participating households (i.e. the so called double difference method: comparing participants and non-participants, before and after the intervention). This was in Can Loc District, Ha Tinh province, in Vietnam. In 1996 ActionAid Vietnam began a savings and credit program in Can Loc. In 1997 I helped them design and implement a baseline survey of almost 600 households, being a 10% sample of the population in three communes of Can Loc District, covering participants and non-participants in the savings and credit services (which reached about 25% of all households). This was done using the Basic Necessities Survey (BNS) , an instrument that I have described in detail elsewhere.

A few years later the responsibility for the project was handed over to a Vietnamese NGO called the Pro-Poor Centre (PPC), which had been formed by ex-Action Aid staff who used to work in Ha Tinh. They continued to manage the savings and credit program over the following years. In 2006, nine years after the baseline survey, an ex-ActionAid staff member who was now working for a foundation in Hanoi, held discussions with the PPC about doing a follow up survey of the Ha Tinh households. I was brought in to assist the re-use of the same BNS instrument as in 1996. At this stage the main interest was simply to see how much households' situations had improved over the nine year period, a period of rapid economic growth throughout much of Vietnam.

The survey went ahead, and was implemented with particular care and diligence by the PPC staff. A copy of the 2006 survey report can be found here (See pages 23-25 especially). Fortunately the PPC had carefully kept hard copy records of the 1996 baseline survey (including the sample frame) and I had also kept digital copies of the data. This meant it was possible to make a number of comparisons:
  • Of households poverty status in 2006 compared to 1997
  • Of changes in the poverty status of households who were and were not participating in the saving and credit program during these periods i.e
    1. Those who had never participated
    2. Those were in (in 1997) but dropped out (by 2006)
    3. Those who were not in (in 1997) but joined later (before 2007)
    4. Those who were always in (in 1997 and 2006)
    Somewhat to my surprise, I found what seemed an ideal set of results. Poverty levels had dropped the most in the 4th group ("always members"), then almost as much in the 2nd group ("ex-members"), less in the 3rd group ("new members") and least in the 1st group ("never members"). The 3rd group might have been expected to have changed less because over the years the project had expanded its coverage to include the less poor, reaching 43% of all households by 2006.

    However, the project's focus on the poorest was also a problem. The members of the savings and credit program had not been randomly chosen, so the control group was not really a control group. They were not comparable. (and I had not heard of, nor still know, how to use the propensity score matching method)

    The alternative to considering the effects of a counter-factual (i.e. a non-intervention) is, I guess, what could be called “counter-theoretical” That is, an alternative theory of what has happened, with the existing intervention.

    My counter-theoretical centered on the idea of dependency ratios - poor families typically have high dependency ratios (i.e. many young children, relatively few adults). As families age this ratio will change, with dependent children growing up and becoming more able bodied and able to take on workloads and or generate income. Even without the access to a savings and credit program, this demographic fact alone might have explained why the participating families did better over the nine year period. It could also explain why the 2nd group did almost as well, if they were selected on the same basis of being the poorest, but had been participants for the shorter period of time.

    What I could have and should have done, was go back to the PPC and see what data they had on the family structure of the interviewed households. It is quite likely they would have the relevant data: ages of all family members, given their close involvement with the community. Unfortunately at that time there was not much interest in the impact assessment aspect of the survey, by either the foundation, the PPC or ActionAid, and their support was necessary for any further analysis. Perhaps I gave up too quickly…

    Nevertheless, reflection on this experience makes me wonder how often it would be well worthwhile, in the absence of good control group data, giving more attention to identifying and testing “counter-theoreticals” about the existing intervention, as part of a more rigorous process of coming to conclusions about impacts.

    PS1 3rd November 2010: I have since recalled that as part of the 2006 survey I met with the staff of ActionAid in Hanoi to explain the survey process and to solicit from them their views on the likely causes of any improvements. The attached file shows two lists, one relating to ActionAid interventions in the district, and the other relating to interventions in the same district by other organisations, including government. Micro-finance was at the top of the list of the ActionAid interventions seen as likely causes of change, but there were 7 others, as well as 12 non-ActionAid interventions that were possible causes. This raises the spectre of 12 possible alternative hypotheses, let alone various combinations of these. One approach I subsequently toyed with for generating composite predictions in this kind of multiple-location/multiple-intervention situation was the "Prediction Matrix".

    PS2 3rd November 2010: The current edition of Evaluation (16(4), 2010 has an article by Nicoletta Stame, titled " What doesn’t work? Three Failures, Many Answers" which includes a section on "Rival Explanations" which I have taken the liberty of copy and pasting below:
    "The link between complexity and causation has been at the centre of evaluation theory ever since and has nurtured thinking about 'plausible rival hypotheses' (Campbell, 1969). Although it was originally treated as a methodological problem of validity, it has recently been revisited from the substantive perspective of programme theory. Commenting on Campbell's interest in 'reforms', that are by definition 'complex social change', Yin contrasts two strategies of Campbell: that of the experimental design and that of using rival explanations. He concludes that the second - as Campbell himself came to admit in Campbell (199a) - is better suited to complex interventions (that are changing and multifaceted), as it is with the complex case studies that have been Yin's turf for a long times (Yin, 2000: 242). The use of rival explanations is common in other crafts journalism, detective work, forensic science and astronomy), where 'the investigator defines the most compelling explanations, tests them by fairly collecting data that can support or refute them, and - given sufficiently consistent and clear evidence - concludes that one explanation but not the others is the most acceptable' (Yin, 2000: 243). These crafts are empirical: their advantage is that while a 'whole host of societal changes may be amenable to empirical investigation', especially those where stakes are currently the highest, they are 'freed from having to impose an experimental design' ('the broader and in fact more common use of rival explanations covers real-life, not craft, rivals', Yin, 2000:248). Nonetheless, rival explanations are by no means alien to evaluation, as is shown by how Campbell himself has offered Pawson good arguments for criticizing the way systematic reviews are conducted (Pawson, 2006)."
    "The problem that remains is how to identify rival explanations. From a methodological starting point, Yin says that 'evaluation literature offers virtually no guidance on how to identify and define real-life rivals'. He proposes a typology of real-life rivals, that can variously relate to targeted interventions, to implementation, to theory to external conditions; and proposes examples of how to deal with them taken from such fields as decline in crime rates, support for industrial development, technological innovations, etc. However, Yin appears to overlook something that had indeed fascinated theory-based evaluation since its first appearance: the possible existence of different theories to explain the working of a programme, and the need to choose among them in order to test them. And - as Patton (1989: 377) has advised - it should be noted that in this way it would be possible to engage stakeholders in conceptualizing their own programme’s theories of action. Nevertheless, Yin’s contribution in its explicitness and methodological “correctness” is an important step forward."
    "Weiss responded to Win’s provocative stance. In an article entitled “What to do until the random assigners come”, she locates Yin’s contribution as the next step beyond Campbell’s ideas about plausible rival hypotheses: “where Campbell focused primarily on rival explanations stemming from methodological artifacts, Yin proposes to identify substantive rival explanations” (Weiss, 2002: 217). She describes the process whereby the evaluator “looks around and collects whatever information and qualitative data are relevant to the issue at hand” (2002: 219), in order to see “whether any [other factor, such as other programs or policies, environmental conditions, social and cultural conditions] could have brought about the kinds of outcomes that the target program was trying to affect”, thus setting up systematic inquires into the situation. Weiss concludes that alternative means to random assignment in order to solve the causality dilemma can be a “a combination of Theory-Based Evaluation and Ruling-Out” (the rival explanation)."
    I recommend the whole article...

    PS3 6th December. On re-reading this post, especially Nicoletta's quote, I wondered about the potential usefullness of the "Evolving Storylines" method I developed some years ago. It could be used as a means of developing a small range of alternative histories of a project, that could then each be subject to some testing (by focusing on the most vulnerable point in each story)

    Tuesday, October 05, 2010

    Do we need a Minimum Level of Failure (MLF)?

    This week I am attending the 2010 European Evaluation Society conference in Prague. Today I have been involved in a number of interesting discussions, including how to commission better evaluations and the potential and perils of randomised control trials (RCTs). This has prompted me to resurrect an idea I have previously raised partly in jest, but which I now think deserves more serious consideration.

    Background first: RCTs have been promoted as an important means of improving the effectiveness of development aid projects. But there are also concerns that RCTs will become a dominating orthodoxy, driving out the use of other approaches to impact assessment, and in the worst case, discouraging investment in development projects which are not evaluable through the use of RCTs.

    In my PhD thesis many years ago I looked at organisational learning through the lense of an evolutionary epistemology. That school of thought sees evolution (through the re-iteration of variation, selection and retention) as a kind of learning process, and human learning as a sub-set of that process. As I explain below, that view of the process of learning has some relevance to the current debate on how to improve aid effectiveness. It is also worth acknowledging the results of that process - evolution has been very effective in developing some extremely complex and sophisticated lifeforms, against which intentionally designed aid projects pale in comparison.

    The point to be made: A common misconception is evolution is about the “survival of the fittest”. In fact this phrase, coined by Herbert Spencer, is significantly misleading. Biological evolution is NOT about the survival of the fittest, but the non-survival of the least fit. This process leaves room for some diversity amongst those that survive, and it is this diversity that enables further evolution. The lesson here is that the process of evolution is not about picking winners according to some global standard of fitness, but about culling of failures based on their lack of fitness to local circumstances.

    This leads me to my own “modest proposal” for another route to improved aid effectiveness, which is an alternative to the widespread use of RCTs and the replication of the kinds of projects found to be effective via that means. This would be to build a widening consensus about the need for a defined “Minimum Level of Failure” (MLF) within the portfolio of activities funded or implemented by aid agencies. A MLF could be something like a 10% of projects by value. Participating agencies would committ to publicly declaring this proportion of their projects as failed. Each of these agencies would also need to show: (a) how in their particular operating context they have defined these as failures, and (b) what steps they will take to avoid the replication of these failures in the future. There would be no need for a global consensus on evaluation methods, or a hegemony of methods arising through less democratic processes.  PS: Using the current M&E terminology, the consensus would need to be on the desired outcomes, not on the activities needed to achieve them.

    I can of course anticipate, if not already hear, some protests about how unrealistic this proposal is. Let us hear these protests, especially in public. Any agency that did so would probably be implying, if not explicitly arguing, that such a failure rate would be unacceptable, because public monies and poor people’s lives are at stake. However making such a de facto claim of a 90%+ rate of success would be a seriously high risk activity, becaus it would be very vulnerable to disproof, probably through journalistic inquiry alone. For anyone involved with development aid programmes, a brief moment’s reflection would suggest that the reality of aid effectiveness is very different, and that a 10% failure rate is probably way too optimistic and in real life failures are much more common.

    Perhaps the protesting agencies might be better advised to consider the upside of a achieving a minimum level of failure. If taken seriously establishing a norm of a minimal level of failure could help get the public at large, along with journalists and politicians, past the shock-horror of failure itself and into the more interesting territory of why some projects fail. It could also help raise the level of risk tolerance, and enable the exploration of more innovative approaches to the uses of aid. Both of these developments would be in addition to a progressive improvement on the average performance of development projects resulting from a periodic culling of the worst performers.

    It is possible that advocates of specific methods like RCTs (as the route to improved aid effectiveness) might also have some criticisms of the MLF proposal. They could argue that these methods will generate plenty of evidence of what does not work, and perhaps that evidence should be privileged. But the problem with this method-led solution is that there is already a body of evidence from a number of fields of scientific research that negative findings are widely under-reported. People like to publish positive findings. This may not be a big risk while RCTs are funded by one or two major actors, but it will become a systemic risk as the number of actors involved increases.  There needs to be an explicit and public focus on failure.

    Actual data on failure rates

    PS: 15th October 2010: Four days ago I posted below some information on the success and failure rates of  DFID projects. I have re-stated and re-edited that information here with additional comments:

    There is some interesting data on failure within the DFID system, most notably the most recent review of Project Completion Reports (PCRs), undertaken in 2005. See the “An Analysis of Projects and programmes in Prism 2000-2005”report available on the DFID website. The percentage (68%) of projects “defined as ‘completely’ or ‘largely’ achieving their Goals (Rated 1 or 2)” was given at the beginning of the Executive Summary, but information about failures was less prominent. Under section “8. Lessons from Project Failures” on page 61 it is stated “There are only 23 projects [out of 453] within the sample that are rated as failing to meet their objectives (i.e. 4 or 5) and which have significant lessons” (italics added). This is equivalent to about 5% of the sampled projects.

    More importantly are the 20% or so rated 3 = Likely to be partly achieved (see page 64). It could be argued that those with a rating of 3 should also be included as failures, since their objectives are only likely to be partly achieved, versus largely achieved in the case of rating 2. In other words a successful project should be defined as one likely to achieve more than 50% of its Output and Purpose objectives. Others are failures. This interpretation seems to be supported by a comment sent to me (whose author will remain anonymous): " "My understanding is that projects with scores of less than 2 are under real pressure and maybe quickly closed down unless they improve rapidly. I have certainly "felt the pressure" from projects to score them 2 rather than 3. That said I have not buckled to the pressure!"

    I think the fact that DFID at least has a performance scoring system (for all its faults), that it has done this analysis of the project scores, and that it has made the results public, probably puts it well ahead of many other aid agencies. I would like to hear about any other agencies who have done anything like this, along with comments on the strengths and weaknesses of what they have done. I would also like to see DFID repeat the 2005 exercise at the end of this year, this time with more discussion on the projects rated 3 = Likely to be partly achieved, and what subsequently happened to these projects.

    PS 2nd November 2010: See Lawrence Hadad's reference here to the same DFID set of statistics here, recently quoted/misused on the One Show 

    PS 3rd November 2010:  Thanks to Yu-Lan van Alphen, Programmamanager, Stichting DOEN, Amsterdam, for this book reference: Kathryn Schulz – "Being Wrong",  reviewed in the NYT. It sounds like a good read.

    PS 14th February 2011: Computer programs are intolerant of programming errors. So, computer programmers tried to avoid them at all costs, not always successfully. Doing so becomes a much bigger challenge as software grows in size and complexity. Now some programmers are trying a different approach, that involves recognising that there will always be programming errors. For more, see "Let It Crash" Programming" by Craig Stunz at http://blogs.teamb.com/craigstuntz/2008/05/19/37819/ 

    PS 15th February 2011: "Why negative studies are good for health journalism, and where to find them" "
    This is a guest column by Ivan Oransky, MD, who is executive editor of Reuters Health and blogs at Embargo Watch and Retraction Watch.One of the things that makes evaluating medical evidence difficult is knowing whether what's being published actually reflects reality. Are the studies we read a good representation of scientific truth, or are they full of cherry-picked data that help sell drugs or skew policy decisions?..."

    PS: 21 February 2011: See also the Admitting Failure website

    PS: 23 April 2011. See today's Bad Science column in the Guardian by Ben Goldacre, titled "I foresee that nobody will do anything about this problem", on the difficulty of getting negative findings published

    PS: 23 May 2011 .The above analysis of DFID project ratings focuses on the recognition of failures that have already occurred. It is also possible, and important, to take steps to ensure that failures are possible to be recognised in the first place. A project that has no clear theory of change will be difficult to evaluate and thus difficult to classify as a success or failure. The most common means of describing a development project’s theory of change is probably via a LogFrame representation. Within a reasonably well constructed LogFrame representation there is a sequence of “if…and…then…” statements, spelling out what is expected to happen as the project is implemented and takes effect. While there may be positive developments in a project’s Goal level indicators, there also needs to be associated evidence that the expected chain of causation leading to that Goal has also taken place as expected. It is not uncommon, in my experience, to find that while the expected outcomes have occurred, the outputs that were meant to contribute to those outcomes were not successfully delivered. In this situation the project cannot claim to be successful. There is however a more generic point to be made here.  The more detailed a project’s ToC is the more vulnerable it will be to disproof. Any one of the many expected causal links could be found to have not worked as expected. However, if these linkages have not been disproved, then the stronger the project’s claims will be to have contributed to any expected and observed changes. Willingness to allow failure to be identified strengthens the claim of any success that is observed. This seems an important observation in the case of projects where there is no possibility of making comparisons with a control group where there was no intervention. In those circumstances a ToC should be as detailed and articulate as possible.

    PS: 23 May 2011 :Articulating more disprovable theories of change may sound like a good idea, but it could be argued that this requirement risks locking aid agencies into a static view of the world they are working in, and one which is developed quite early in their intervention. In many settings, for example in humanitarian emergencies and highly politicised environments, aid agencies often have to revisit, revise and adapt their views of what is happening and how they should best respond. The best that might be expected in these circumstances is that those agencies are able to construct a detailed (and disprovable) history of what happened.  This could actually produce better (i.e. more disprovable) results. There is some research evidence which shows that people find it easier to imagine events in some detail when they are situated in the past than to imagine the same kind of events taking place in the future[i].



    [i] Bavelas, J.B. (1973) Effects of the temporal context of information, Psychological Reports, 35, 695-698, cited in Dance with Chance, by Makridakis, Hogarth and Gaba, 2009, page189.