Friday 29 January 2010

Part 5: On the futility of models

I write this as I am sitting in the hotel lobby of the Hotel Cardoso in Maputo, trying to recover from a manic week of seemingly never-ending interviews, traffic jams, gate-crashing (of international conferences and cocktail parties) and run-ins with the police. One more interview to go, I lack the energy to work on anything substantive to do with my PhD although I know that I will have to face the inevitable at some point and actually come up with something concrete, even of it is just for the presentation I have been asked to give At Ed Mondlane university on my return to Mozambique or to prepare for adding an extra trip to Zambia to consult with the contacts I have made there, but the real stress and concern sitting on my shoulders at the moment is that of my newly adopted supervisor whose comments and suggestions I need to try and incorporate into a coherent critique of or even a novel type of climatico-economic model.

And so, please find below my thoughts on models as I attempt to answer the questions posed to me...

Why do we need models?

Do we? This question should really be based more on answering the next question...

What do we do with models?

Are they explicative or predictive or merely self-indulgent? Do they make us answer questions with more questions are are they actually useful in leading us down a thought path from which we may alight more enlightened further down the track? What is their function... to understand more about how the world works, to explain things we can't actually manipulate ourselves, to predict the future or simply to see how things could possible work if the world happened to function in this particular way whether it be linearly, non-linearly, stochastically or whatever, only then will we be able to answer the first question. But, do we really know?

Can we argue all of the above? In terms of policy, surely they should be predictive- we can't make decisions based on theory, there needs to be a probability of occurrence, but this probability is calculated based on the representivity of the model to the real world and hence the likelihood of the predicted outcome occurring (taking into account all the stats of how much the model is actually reflected in reality through models of collected data (the real situation... or merely the quantitative aspect of the real situation- here the problem between modelling physical processes and others becomes a tad shifty)

So I digress...

Leaving this for a second, I turn to what types of models we may employ for these carious

What type of models are we then talking about?

Models as thought-experiments can be extremely useful to the individual researcher, even if just to distract him with mathematical relationships that may not necessarily happen in reality, but which are nevertheless interesting building blocks for future research. E.g.s include many ecological models which stick to the general and don't pretend to be examples of reality. Rather they look at relationships between things and processes such as predator-prey models or source-sink population models that include stochastic variables.

Will inductive models that are based on relationships between empirical data and which are then made into general principles (and the deductive approach of taking general principles and matching them to specific situations for that matter) ever get us anywhere because anything that you're bothering to model in the first place is not going to slip nicely into a generic mathematical formula (will it... the more we know the less we know we know, right?) We live shrouded in complexity- the real question is whether down/upscaling processes through dis/aggregating data is actually a true reflection of the world... is the world really one big fractal? Thrift in his non-representational theory bent will argue most definitely not... representations/reflections... models for that matter are all nonsense.

Well, yes, they may be nonsense, but are they useful?

Can models ever be predictive or are they simply projections, handled under varying scenarios... (read here IPCC/Millennium ecosystem assessment scenarios par example)

However, due to the need to give them more credence than their due for making policy based decisions, these models have been accorded higher value than the initial intent of their modellers... take climate change controversy or climategate!!! Yes, the models are not 100% certain- scientists never function in a world of certainty, but the indications are definitely there and the general trends can be modellef pretty successfully, especially if most of the uncertainty points towards worse results than those projected (NOT predicted) by incorporating non-linear relationships and positive feedbacks (ice-caps melting and albedo effect)

BUT when adding in socio-economic factors are we not straying a bit too far into the unknown. We can't predict a future of climate change, but the likelihood that we are heading in that direction is extremely strong based on data collected and events we can already see, but socio-economic models of the current situation are fraught with problems when scaled up because unlike the climate, socio-economic factors are location and context specific and although they have global impacts, they do not work in a similar way to the global climate system. Hence combining the two is extremely difficult, especially when as much credence is given to the model of food security outcomes under climate change (for example) which are modelled on scenarios and take a macro-view of a globally extremely differentiated situation (all conceded by the authors of course because they know they are dealing with the unknown). In order to improve this, we need local level data- it's no good working with international food prices if someone in rural Mozambique is paying far higher prices (or even proportional to their income far higher) for their maize than your average Joe Soap in the USA.

Therefore...

What data is required?

For global studies? LOTS!!!... and a supercomputer to compute it

Rather, instead of predictive/projection models we should be focussing on modelling subsystems that we know a bit more about and for which we have available data at whatever level is required in order to see how certain processes work... but then to stop there and not to draw inferences from one situation into general principles because there will always be a snag. Rather, admit that what is studied is not 100% explicative or predictive, but can lead to assessments of different situations based on the knowledge of a specific example and experimentation.

Not what your average politician wants to hear about certainty, but at least probably more useful

Proviso: please note that my views expressed above are the work of an over-tired mind at ends with trying to come to terms with models and that I will most likely change my mind about their utility tomorrow- this is not aimed to criticise the modelling community, but more of a critique of how we apply the knowledge gained from said activity and how our expectations of what models can offer us (in the practitioner community) often far outweigh what modellers know they can do