The modifiable areal unit problem in multivariate statistical analysis

Abstract

The examination of the modifiable areal unit problem is extended into multivariate statistical analysis. In an investigation of the parameter estimates from a multiple linear regression model and a multiple logit regression model, conclusions are drawn about the sensitivity of such estimates to variation in scale and zoning systems. The modifiable areal unit problem is shown to be essentially unpredictable in its intensity and effects in multivariate statistical anlaysis and is therefore a much greater problem than in univariate or bivariate analysis. The results of this analysis are rather depressing in that they provide strong evidence of the unreliability of any multivariate analysis undertaken with data from areal units. Given that such analyses can only be expected to increase with the imminent availability of new census data both in the UK and in the US, and the current proliferation of GIS technology which permits even more access to aggregated data, this paper serves as a topical warning.