Andy Sutter writes:
It’s been a while (~2 years?) since I was last reading your blog semi-regularly and submitted a comment or two, but I was reading something today that made me recall those days.
At the time, I was curious about why social scientists present data as charts of regression coefficients, since I’d never seen such a presentation in the physical sciences.
Recently, I’ve been reading a collection of essays by Alain Desrosières, « Pour une sociologie historique de la quantification » (Presses de l’École des Mines 2008), which I’m reviewing for a magazine here in Japan. The author is an historian of statistics, and writes especially about the uses of statistics by governments. Aside from one of his older books, though (The Politics of Large Numbers), I’m not aware of much of his work being available in English.
One of the essays in the book (originally published in 1995) claims that French social scientists prefer a visual display of statistical data, which is usually presented as “orthogonal projections” of a “cloud of points, constructed in a space having a large number of dimensions”. In them, he says, “constellations of co-occuring properties … appear according to a statistical and probabalistic logic, rather than a deterministic one.” He points out that visual presentations make it easier to understand the data and “to stabilize and memorize” the forms of the relationships. “In the written commentaries [to these graphics], the subjects of verbs are social groups, or classes of individuals, connected to each other by a probable commonality of behavior, [and regarded] from a holistic perspective of the reconstitution of a person, a group or a locality. With a statistical data-set as their point of departure, these analyses try to approach the semantic richness of a literary description.”
In contrast, he says, Anglo-Saxon social scientists prefer to use regressions, and write their articles in the language of formulas and equations, which include a deterministic part “plus a random residue. The distinction between so-called endogenous and exogenous variables, or those to be explained and those that are explanatory, is in contrast to the symmetry of [data analyses such as are preferred in France], which are directed not to explanation but to description. … The language of econometric regression … is adapted to building models for action, relating effects (the objects of research) to causes (possible actions).”
Are you familiar with Desrosières, and have you observed the cultural differences he alludes to?
My reply: No, I haven’t heard of Desrosières, but this sounds interesting. The only history of statistics I’ve read is by Stigler. Regarding the statistical issue, I have heard that social science statisticians in France like to use correspondence analysis, which I seem to recall is similar to log-linear analysis for contingency tables in that it models multivariate data without picking one variable to be “y” and the others as “x.”
I have to say, though, call me an Anglo-Saxon or an econometrician or whatever, but I find it very useful to pick out a single variable and model it. I’ve never known what to do with the symmetric models that treat all variables the same way.
I read — or tried to read — Desrosières' "The Politics of Large Numbers". It is not so much a history of statistics as a historical deconstruction of the subject, in the manner of Jacques Derrida. There may have been some good ideas in there, but they were obscured by steaming mounds of post-modern jargon.
All things considered, I prefer Alan Sokol.
Historically, Statistics in France saw a split between mathematical statisticians and data analysts in the 60's. The former were attracted by the formalism of the Bourbaki revolution, while the later wanted to let the data speak for themselves and to throw models to the bin in a typical 1968 fashion. The perspective of Desrosières is typical of this data analysis à la francaise, in the spirit of Jean-Paul Benzécri who introduced correspondence analysis. It is however atypical of the current French statistical community where data analysis is no longer separated from Statistics.
I did notice that the correspondence analysis (similar to principal component analysis, a form of dimension reduction) is very well developed and popular in France.
But correspondence analysis is formally a weighted regression where a specially constructed small set of variables (dimensions) predicts deviations from independence (or more generally some assumed log-linear model) and there is an interpretaion for the coefficents…
This is brought out in recent books by Greenacre, especially in regard to the "faulty" generalization to multiple correspondence analysis and the less wrong generalization to joint correspondence analysis.
My PhD advisor was a student of Jean-Paul Benzécri and I've studied in depth the geometrical point of view of the French. The split between statisticians, Xi'an mentioned, was also present in Greece in the '80s and '90s. You would hear about the 'French School of Data Analysis' and the 'Anglo-Saxon' approach. Exploratory, interdependence vs Confirmatory, dependence techniques. Methods like Correspondence Analysis have been reinvented several times and given different names. To date, things have changed a lot. One side has adopted concepts and tools from the other. I think it's a matter of perspective and I personally prefer the middle approach.
PS. I cannot forget my advisor quoting Benzécri in his lectures: "the model should follow the data and not the inverse".
PS2. Excellent blog.
I've not read the book referred to, but this argument, such as it is, seems to reflect one widespread cultural hang-up that resurfaces frequently in French intellectual writing. Crudely put, it's an obsession with which ideas are or are not distinctively French. (There are also plenty of French intellectuals who don't regard this as the deepest of issues.)
Setting aside "Anglo-Saxon" as a silly label, must we now regard R.A. Fisher and John Tukey with their geometrical and graphical orientations as honorary Frenchmen?
Furthermore, correspondence analysis has been discovered and rediscovered under many names times without count, including outside France and long before Benzécri.
my phd adviser is french, she recently wrote a paper on this topic:
"Multivariate Analysis: The French Way." (2007). Holmes, S.
http://www-stat.stanford.edu/~susan/papers/dfc.pd…
A while ago, I did some work on the theory of social representations with a Swiss teacher that studied with French supervisors. Indeed, all the theory of Social Representations (search for Serge Moscovici) is based on those clouds of points mentioned, with identification of groups according to shared perceptions of reality.
Aggrupation methods were prefered, visual scatterplots of data were common, significance levels hardly ever used.
Sorry to be joining this thread late. I appreciate the interesting comments. Apropos of John S.'s comment, I share the Derrida allergy, but I found the writing in Desrosières's recent book (which actually is a bunch of articles written during a roughly 20-year period) generally pretty down-to-earth.
Regarding Nick Cox's observations, I'd ask his indulgence regarding the "Anglo-Saxon" label, and not just because it's handier to use than enumerating all the pertinent countries. I've been reading a lot of European social science stuff in the past year or so, most of it French, and I'd say that in general the Europeans are much better about citing Anglophone authors, even if only to contest them, than vice versa. Other than in empirical studies about a specific country, it is exceedingly rare to find foreign-language references in papers in the fields of economics, law or management by authors based in "Anglo-Saxon" countries. While I wouldn't claim that French scholars are the most open-minded on earth, the intellectual isolationism seems nonetheless pretty lopsided, at least from my first-hand anecdotal POV.
That summary matches my impressions too, in different fields.
"Anglophone", as used here, is surely an improvement on "Anglo-Saxon". Without wanting to look up anyone's ancestry, it seems clear that many of us fairly summarised as Anglophone would not want the label Anglo-Saxon.