The FCC in the US published research by the Berkman Centre at Harvard to assess the impact of certain policy parameters on the speed of broadband adoption.
This research has been much criticised, with one particular criticism by George Ford focussing on the econometrics involved.
The issue has a particular Australian resonance because the original econometric model used is one developed by our own John de Ridder who first presented this work at the 2006 Communications Policy and Research Forum.
At issue is the appropriate econometric techniques to be used in estimating broadband demand, with the study utilising a simple OLS approach estimating demand only, rather than a simultaneous version that estimates demand and supply. I do not propose to enter that debate at all, except to note that this debate reflects an ongoing problem with almost all econometric studies. That is, the degrees of fredom available to researchers.
It also highlights another factor which is the relationship between published results and data sets. The Berkman report acknowledges John for making the data set available. I am surprised that in the age of the Internet it is not manadatory (or at least the expected norm) for datasets (unless confidential and the confidential detail cannot be "masked") to be published online.
The question for me is whether Panel data can ever be meaningfully used in either form to assess factors that determine an adoption rate. My reasoning is simple, that all adoption cycles follow some kind of S-curve and it is the shape of the curve that needs to be "explained" and that the points on the curve (the observed data) need to be first fitted to the curve and then the parameters of the curve estimated through a second exercise. Put another way, what needs to be explained is not the adoption level, but the rate of adoption.
Using Panel data (i.e. cross section data) in either of the methods used above carries an implicit assumption about what effects the shape of the curve.
First let's consider why adoption follows an S-shaped curve. I can identify three simple factors that would individually result in S-shaped adoption curves. The first is a simple demand effect, and the relationship of demand to income. Assuming price remains unchanged there will be an adoption process as progressively more people for whom the product is of interest can afford to buy it - that is as relative incme increases compared to supply costs.
The second is a supply effect, that the well known concept of the "experience curve" in production means that the cost of supply declines with increases in cumulative production. That is mathematically a similar effect to the above, but is driven by decline in cost rather than increae in income.
The third, and perhaps most important, is a variety of network effect. In the study of economic dynamics a simple exercise is to develop an adoption model that assumes that an individual's propensity to acquire a service is proportional to how many other people have already acquired it. This generates the simplest of all adoption curves, the logistic curve. This model also provides an interesting interpretation of the often claimed great Australian propensity to adopt. In practice Australian adoption curves start out very slowly, and then increase rapidly. An interpretation of this is that Australians are more influenced by what others do than other countries. In popular culture it looks like fast adoption because it progresses rapidly once it starts, but this ignores the slow start.
Fitting adoption to S-shaped curves suffers from a lack of clarity about what is the appropriate form of the curve, the fact that as a non-linear estimate it requires the study to choose some initial values and the fact that the "fit" may not be unique. The flip side is that the cumulative OECD data provides some very good data for the analysis - that shows all countries with some kind of S-curve with vastly varying slopes.
Finally, one has to wonder whether trying to use econometric analysis to determine the effectiveness of different policy settings on broadband adoption isn't just an overall inappropriate application of a tool. The data set cannot be adjusted forv the varying quality of the product, nor the complicated pricing structures.
Full credit to John for trying the initial study, I thought it was dodgy then and I still do. But equally I think the Ford conclusions are equally dodgy.
Note: Many thanks to Grahame Lynch of Communications Day for bringin the controversy to my attention through the pages of his newsletter.