Published on November 10th, 20129
Multilevel Analysis – The New Way to Find Gaps in the Research
By Tor G. Jakobsen
Regression models are considered the work horse of the social sciences. Hundreds of thousands of social scientists have investigated thoroughly how individual level characteristics influence other individual level characteristics, and how country level characteristics influence other country level characteristics. How can today’s aspiring social scientist find new correlations to write research papers about? The answer: through multilevel modeling.
It is Professor in Statistics Harvey Goldstein who is considered the founder of multilevel modeling. In 1991 he developed software that allowed for running so-called two- and three-level models. Goldstein’s assumption was that a unit at the lowest level (level-1) was nested into a higher-level unit, such as a region, country or school (level-2).
Multilevel modeling soon became popular within educational research, where students were nested into school classes, which again were nested into schools. Ordinary regression models assume the independence of units, which in such data is breached because of the nesting. This is taken into account when running hierarchical (another name for multilevel) models. The object of a multilevel analysis is to account for variance in a dependent variable measured at the lowest level, by investigating information from all levels of analysis.
There are both theoretical and statistical reasons for using this approach. From a theoretical point of view the researcher might be concerned with the relationship between the individual its surroundings, for example arguing that the individual is influenced by the features of his or her country, region, or school. Observations that are close in space are likely to be more similar than observations far apart in space. Thus, respondents from the same country are more similar than respondents from different countries, due to shared history, experiences, environment etc. Many of these relationships are unexplored, and pave the way for the up-and-coming social scientist trying to generate new research.
This also implies a statistical reason for using multilevel modeling. Such a shared context is a cause of dependency among observation. If the individual level dependent variable is influenced by for example country level variables, the observation at the lowest level are not independent, that is, they are clustered.
An additional feature is that multilevel modeling is a good statistical reply to the critique forwarded by several proponents of the qualitative method, namely that one needs to take into account the context of the individuals when studying these. This is actually one of the advantages of multilevel analysis: by including level-2 factors in the regression equation one allows for the context surrounding the individuals to be accounted for.
Hox, Joop (2002). Multilevel Analysis: Techniques and Applications. Mahwah, NJ: Lawrence Erlbaum Associates.
Steenbergen, Marci R. & Bradford S. Jones (2002) «Modeling Multilevel Data Structures» American Journal of Political Science, 46(1): 218–237.