This thesis studies the properties of the family of chi-square copulas. This is a generalization of the multidimensional normal copulas obtained by squaring the components of normal random vector. These copulas are indexed by a correlation matrix and by a shape parameter. This thesis shows how this family can be used to perform spatial interpolation and to model multidimensional data. First, the usefulness of this class of dependence structures is demonstrated with an application in spatial statistics. An important problem in that context is to predict the value of a stationary random field at a position where it has not been observed. This thesis shows how to construct such predictions using spatial models based on copulas. One focusses on the use of the family of chi-square copulas in that context. One must first assumes that the correlation matrix has a standard parametric form, such as that of Matérn, indexed by an unknown parameter associated with the force of the spatial association. This parameter is first estimated using a composite pseudo-likelihood constructed from the bivariate distributions of the observed data. Then, a spatial interpolation method using the ranks of the observations is suggested to approximate the best prediction of the random field at an unobserved position under a chi-square copula. In a second work, the fundamental properties of the chi-square copulas are studied in detail. This family allows a lot of flexibility to model multidimensional data. In the bivariate case, this family is adapted to symmetric and asymmetric dependence structures. In larger dimensions, the shape parameter controls the degree of radial asymmetry of the two-dimensional marginal distributions. Parameter estimation procedures of the correlation matrix and of the shape parameter are compared under independent and identically distributed repetitions. Finally, the formulas of the conditional expectation for the best prediction in a spatial context are established. Goodness-of-fit tests for the family of chi-square copulas are then developed. These new tests can be applied to data in any dimension. These procedures are based on two association measures based on the ranks of the observations, which avoids having to specify the marginal distributions. It is shown that the joint behavior of these two measures is asymptotically normal. The efficiency of the new goodness-of-fit procedures is demonstrated through a simulation study and is compared to a classical goodness-of-fit test based on the empirical copula.