Steel and steel-concrete composite bridges made of plate girders represent an economical structural form and are largely used for short and medium spans. Two parallel plate girders connected by cross-beams define a twin-girder bridge. This type of bridge transfers the loads from the deck to the supports essentially through bending and shear of I-girders. Associated with the two resistance modes are the following instability modes: lateral torsional buckling (LTB) and local buckling. Design experience of steel bridges made of plate girders shows that LTB determines the dimension of the upper flange of an I-girder at mid-span during construction phase, and the dimension of the bottom flange next to the intermediate supports during construction by launching or in service. In practice, LTB verification is carried out using reduction curves (i.e. buckling curves) which decrease the resistance of a beam with respect to the sensitivity to LTB; thus, the more sensitive the beam, more severe the reduction. A comparison of different standards shows reduction curves available for steel bridge girders have non-negligible differences which can reach 30 % of the resistance. A more refined study of the buckling curves shows that it is semi-empirical in nature, where the empirical part is mainly based on numerical and experimental research for building sections. These findings justify the goal of this thesis which is to propose a verification method for LTB that better takes into account the special features of bridge girders. To reach this goal, study of parameters that influence LTB, like re-sidual stresses, geometric imperfections, cross-beams, loading and beam geometry, is required. In this research, the experimental studies carried out on the residual stresses propose a new model which takes into account flame-cutting and welding of thick plates used in the fabrication of steel bridge girders. Experimental measurements of geometric imperfections performed on real bridge girders show the difference between fabrication tolerances and imperfections measured. These results are then used as input data for numerical simulations of LTB by two approaches: beam and bridge. The beam approach, which considers a simple span I-girder with fork support and a constant bending moment, reveals that the influence on the resistance to LTB can reach 10 % for the residual stresses and 12 % for the geometric imperfections. The bridge approach, which considers a twin-girder bridge with cross-beams and a constant load on the upper flange, shows that the influence reaches 17 % for cross beam spacing and 27 % for the type of cross-beams (frame, truss, diaphragm). Based on the experimental and numerical results, a design method for LTB is proposed. It recommends the use of reduction curve c for SIA263:2013 or reduction curve c of Eurocode EN1993-2:2006.