Context. It is becoming clear that determination of the abundance of Si using lines of Si ii and Si iii can lead to quite discordant results in mid to late B-type stars. The difference between the Si abundances derived from the two ion states can exceed one dex in some cases. Aims. We have carried out a study intended to clarify which kinds of B stars exhibit this discrepancy, to try to identify regularities in the phenomenon, and to explore possible explanations such as abundance stratification by comparing models to observed spectra. Methods. We used spectra from the ESPaDOnS spectropolarimeter and FEROS spectrograph, supplemented with spectra from the ESO and ELODIE archives, of magnetic Bp, HgMn, and normal B-type stars ranging in effective temperature from about 10 500 to 15 000 K. Using these spectra, we derived abundances using the spectrum synthesis program zeeman, which can take the influence of magnetic fields into account. For each star, accurate abundances of Si ii, Si iii, Ti, Cr, and Fe were derived from two separate ∼100 Å windows. Si ii abundances were deduced from multiple lines, and Si iii abundances were found using λλ 4552, 4567, and 4574.Results. All magnetic Bp stars in our sample show a discordance between the derived abundances of the first and second ions of silicon, with the latter being between 0.6−1.7 dex higher. The same behaviour is observed in the non-magnetic stars but to a much smaller extent: Si iii is enhanced by between 0.3−0.8 dex compared to Si ii. We do not detect the discrepancy in three stars, HD 22 136 (normal), HD 57 608 (HgMn) and HD 27 295 (HgMn); these are the only stars in our sample for which the microturbulence parameter is significantly different from zero, and which therefore probably have convection occurring in their atmospheres. Conclusions. We find that vertical stratification of silicon in the atmospheres of B-type stars may provide an explanation of this phenomenon, but our detailed stratification models do not completely explain the discrepancies, which may, in part, be due to non-LTE effects.