This thesis describes a series of experiments which investigate preschool children's spatial abilities. To overcome the problems of extrapolating from traditional laboratory task to abilities in the real world, the children were tested in 'natural' environments, such as buildings and streets, and which were large-scale, that is, they could not be viewed in their entirety from one position but instead had to be constructed from successive views. The measure of spatial knowledge chosen was direction estimation, a task which has been successfully used by other authors with older subjects, and which avoids the problems of interpretation and comprehension inherent in more traditional methods of investigating spatial representation such as map drawing and model building. The findings are discussed in relation to Piaget's distinction between topological and Euclidean spatial knowledge (Piaget et al. 1960; Piaget and Inhelder 1967; Piaget 1977), but traditionally used interpretations of his theory are avoided (e.g. Siegel and White 1975) as being based upon methodologically problematical experimentation. The results are interpreted in terms of Byrne's (1979, 1982) network-map/vector-map theory of spatial knowledge, which has only previously been applied to adults. It was found that preschool children can show both network-map knowledge (topological), and vector-map knowledge (Euclidean). Piaget's stage theory of development is inappropriate as within the age and ability range tested here, the type of spatial knowledge shown was more dependent upon qualities of the environment than of the child. Preschool children are most likely to show vector-map knowledge in small, over-learned, and actively explored environments than in larger passively explored but familiar environments, and lastly in novel large environments. Preschool children's network-map knowledge, built up by walking in natural environments, is coded in one direction only; and two separately learned but overlapping routes are encoded as an integrated network.