Estimation of the number of normal subpopulations in an heterogeneous sample and of their parameters. A recurrent method applied to neurobiology.

A method is proposed to estimate the number of subpopulations which compose a heterogeneous experimental population. The method consists of reconstructing a population sample by adding computed subpopulations, postulated to be independent. The assumption is made that each subpopulation fits a normal distribution; the method may also be generalized to other distributions laws; each subpopulation can thus be entirely defined by its mean, variance and relative weight. Recurrent adjustments of number and parameters of subpopulations are carried out in order to minimize the difference between synthetized and experimental populations. This difference was measured by the quadratic distance chi2 between the two populations. Mean, variance and weight of the computed population must then be close to those of the experimental population in order to consider the results as acceptable. This method is discussed with regard to others proposed in the literature.