ABSTRACT The next generation of spectroscopic surveys will target emission-line galaxies (ELGs) to produce constraints on cosmological parameters. We study the large-scale structure traced by ELGs using a combination of a semi-analytical model of galaxy formation, a code that computes the nebular emission from H ii regions using the properties of the interstellar medium, and a large-volume, high-resolution N-body simulation. We consider fixed number density samples where galaxies are selected by their H α, [O iii] λ5007, or [O ii] λλ3727–3729 emission-line luminosities. We investigate the assembly bias signatures of these samples, and compare them to those of stellar mass- and star formation rate-selected samples. Interestingly, we find that the [O iii]- and [O ii]-selected samples display scale-dependent bias on large scales and that their assembly bias signatures are also scale dependent. Both these effects are more pronounced for lower number density samples. The [O iii] and [O ii] emitters that contribute most to the scale dependence tend to have a low gas-phase metallicity and are preferentially found in low-density regions. We also measure the baryon acoustic oscillation (BAO) feature and the β parameter related to the growth rate of overdensities. We find that the scale of the BAO peak is roughly the same for all selections and that β is scale dependent at large scales. Our results suggest that ELG samples include environmental effects that should be modelled in order to remove potential systematic errors that could affect the estimation of cosmological parameters.