We study numerically the metal -insulator transition in the Anderson model on various lattices with dimension 2 < d ≤ 4 (bifractals and Euclidian lattices). The critical exponent ν and the critical conductance distribution are calculated. We confirm that ν depends only on the spectral dimension. The other parameters -critical disorder, critical conductance distribution and conductance cummulants -depend also on lattice topology. Thus only qualitative comparison with theoretical formulae for dimension dependence of the cummulants is possible.PACS numbers: 71.30.+h, 71.23.An, 72.15.Rn It is commonly accepted, though not proved, that metal -insulator transitions (MIT) can be described by oneparameter scaling theory [1]. The critical exponent ν which describes the divergence of correlation length at MIT depends only on the system dimension for a chosen universality class. Microscopic details of models do not affect it. This was confirmed by numerical analysis of quasi-one-dimensional (Q1D) systems [2]. Theoretical dependence of ν on dimension d = 2 + ε was found in [3,4] for small ε. Numerically, ν(ε) was studied on bifractals [5].The conductance g was originally chosen as the order parameter in the scaling theory [1]. Soon it became clear, that the absence of self-averaging of g in the critical region must be taken into account [6,7]. The shape of the critical conductance distribution P (g) in 3D models was numerically analysed in detail [8,9,10,11,12]. Contrary to the critical exponent, P (g) is not universal. Its shape depends not merely on the dimension [10] and physical symmetry [11] but also on boundary conditions [12] and even anisotropy [13]. Nevertheless, for a given physical model the mean conductance and resistance follow one parameter scaling [14].Analytical theory of MIT is restricted to systems with dimension close to the lower critical dimension: 2 + ε with ε ≪ 1 [15]. In spite of predicted non-universality of higher order conductance cummulants δg n ,(1) the distribution P (g) should be universal in the infinite system size limit [16]. For small ε, the bulk of the distribution is approximately Gaussian near the mean value g . The parameters of Gaussian peak,agree with the estimation of the first cummulants (1). For large g, theory predicts long power-law tailNumerical results for 3D systems show completely different P (g), indicating that theory is not applicable to large ε [8].In this Letter we present the critical exponent and the critical conductance distribution for Anderson model obtained numerically on three bifractal lattices with dimension 2 < d < 3. They all possess the same fractal dimension d f = log 3/ log 2 + 1. Two of these lattices have the same spectral dimension d s . Their critical exponents are identical within error bars. This confirms the universality of MIT. The shape of the P (g), and the value of the critical disorder, depend not only on d s but also on the lattice topology. This novel non-universality of P (g) is found also by numerical analysis of two different 3D...