We study the Anderson transition for three-dimensional (3D) N × N × N tightly bound cubic lattices where both real and imaginary parts of onsite energies are independent random variables distributed uniformly between −W/2 and W/2. Such a non-Hermitian analog of the Anderson model is used to describe random-laser medium with local loss and amplification. We employ eigenvalue statistics to search for the Anderson transition. For 25% smallest-modulus complex eigenvalues we find the average ratio r of distances to the first and the second nearest neighbor as a function of W . For a given N the function r(W ) crosses from 0.72 to 2/3 with a growing W demonstrating a transition from delocalized to localized states. When plotted at different N all r(W ) cross at Wc = 6.0±0.1 (in units of nearest neighbor overlap integral) clearly demonstrating the 3D Anderson transition. We find that in the non-Hermitian 2D Anderson model, the transition is replaced by a crossover.Anderson localization is the central concept of solid state physics for more than 60 years 1-3 . It determines electron conductivity of doped crystalline and amorphous semiconductors and many other disordered systems and is observed in experiments 4,5 .In recent years the problem of localization attracted renewed interest as research moved to formerly unexplored area of non-Hermitian systems. Random lasers 6-9 with random dissipation and amplification regions are such prototypical non-Hermitian systems. The other parts of non-Hermitian disorder physics are related to Hatano-Nelson matrices 10 , their biological applications 11,12 or to spin chains [13][14][15] . All these works focus on onedimensional systems.A simple and elegant extension of the 2D Anderson localization problem was proposed in a recent paper by Tzortzakakis, Makris and Economou (TME) 16 . They studied 50 × 50 tight-binding square lattices with real overlap energy I ij = I, and random complex onsite energies E i whose real and imaginary parts are independent random variables distributed uniformly between −W/2 and W/2. The Hamiltonian readswhere i, j in the second term are nearest neighbors, and the hard-wall boundary is employed (no bonds extended out the boundary). Below we call this non-Hermitian Hamiltonian the TME model. By calculating the participation ratio of eigenfunctions of such non-Hermitian Hamiltonian, TME noticed that they become progressively more localized when W (in units of I) grows from 1 to 5. Simultaneously the distribution function P (s) of nearest neighbor distances s between eigenvalues in the complex plane widens, which shows that the repulsion of eigenvalues weakens due to the progressive localization of eigenfunctions. This behavior is similar to what happens in the Anderson model 17 . TME, however, did not raise a question whether there is an Anderson transition or a crossover in the limit of large system.In this paper, we focus on the question of the existence of the Anderson transition in the TME model for 3D cubic and 2D square lattices. We show that in the TME m...