Arnaud Bringé and Valérie Golaz
Arnaud Bringé, the head of INED’s Statistical Methods service, and Valérie Golaz, a research officer at INED, answered our questions on their recently published Manuel pratique d’analyse multiniveau [Practical manual of multilevel analysis].
(Interview conducted in October 2017)
What is multilevel analysis?
Social facts fit into complex realities; different levels of observation are needed to define and explain them. Statistical modelling never explains reality entirely, but it does bring in several levels of analysis. For example, in his renowned study of suicide, Durkheim distinguished characteristics particular to individuals from other characteristics related to their degree of integration into society. If we are studying older persons, their situation is the result of their occupational and family histories; social relations, often constructed over a long period of time; the residential frame in which they live (housing, access to services, public policies targeting or affecting this group, etc.). Residential mobility is a function of individuals and families (occupation, marital situation) but also of more general characteristics pertaining to the places they leave and those they move to (quality of public services, environment, etc.). Similarly, to take the example we develop throughout the manual using census data recently made available to the academic community by IPUMS-International, the fact that an African child is not in school is due simultaneously to the child’s characteristics, those of his or her family, and those of the milieu in which he or she lives—the social milieu as well as the administrative context (degree of public service development, school distances, how crowded schools are, etc.).
The power of multilevel analysis is that it can take several levels of analysis into account in a single statistical model. The model takes account of data structuration and correlations by level. This gives us better estimations, particularly for the role played by contextual variables in individual behaviours.
How does multilevel analysis improve estimations?
In classic regression, contextual data are treated as individual variables like the others. In multilevel regression, they are identified with the administrative area or social group they refer to and their effects can therefore be estimated with more accuracy. Above all, though, multilevel analysis is a great tool for refining analyses, because by measuring in terms of level you can better identify whatever the model does not explain. So users can distinguish what follows from differences between individuals and what follows from differences between milieus that are not explained by the model variables. This will guide them in identifying complementary invariables whose inclusion will then improve the model and the findings.
What need was there for a manual on multilevel analysis?
Multilevel analysis can now be done with all the standard programmes but it is rife with pitfalls: many models don’t work, and you need to know how to interpret findings and non-findings in order to move forward. Though there are a considerable number of theoretical books on the subject, few offer practical illustrations. Our manual was designed as a teaching tool for beginners in multilevel analysis, to get them over the first stumbling blocks by explaining programming, finding comparison, and errors that may result from three programmes currently used in statistical analysis: SAS®, Stata® et R.
Who was the manual written for?
All users of statistics who analyse structured data by different observation levels, and anyone who would like to do so. In this time of Bigdata, users need to have a well-adapted methodology to take on the massive amounts of data now available and to ensure that their approaches and analyses are relevant. It is this challenge we seek to address in the manual.