I present a study of the sizes (semimajor axes) of bars in disc galaxies, combining a detailed R-band study of 65 S0-Sb galaxies with the B-band measurements of 70 Sb-Sd galaxies from Martin (1995). As has been noted before with smaller samples, bars in early-type (S0-Sb) galaxies are clearly larger than bars in late-type (Sc-Sd) galaxies; this is true both for relative sizes (bar length as fraction of isophotal radius R25 or exponential disc scalelength h) and absolute sizes (kpc). S0-Sab bars extend to ~1-10 kpc (mean ~ 3.3 kpc), ~0.2-0.8R25 (mean ~ 0.38R25) and ~0.5-2.5h (mean ~ 1.4h). Late-type bars extend to only ~0.5-3.5 kpc, ~0.05-0.35R25 and 0.2-1.5h their mean sizes are ~1.5 kpc, ~ 0.14R25 and ~0.6h. Sb galaxies resemble earlier-type galaxies in terms of bar size relative to h; their smaller R25-relative sizes may be a side effect of higher star formation, which increases R25 but not h. Sbc galaxies form a transition between the early- and late-type regimes. For S0-Sbc galaxies, bar size correlates well with disc size (both R25 and h); these correlations are stronger than the known correlation with MB. All correlations appear to be weaker or absent for late-type galaxies; in particular, there seems to be no correlation between bar size and either h or MB for Sc-Sd galaxies. Because bar size scales with disc size and galaxy magnitude for most Hubble types, studies of bar evolution with redshift should select samples with similar distributions of disc size or magnitude (extrapolated to present-day values); otherwise, bar frequencies and sizes could be mis-estimated. Because early-type galaxies tend to have larger bars, resolution-limited studies will preferentially find bars in early-type galaxies (assuming no significant differential evolution in bar sizes). I show that the bars detected in Hubble Space Telescope (HST) near-infrared(IR) images at z~ 1 by Sheth et al. have absolute sizes consistent with those in bright, nearby S0-Sb galaxies. I also compare the sizes of real bars with those produced in simulations and discuss some possible implications for scenarios of secular evolution along the Hubble sequence. Simulations often produce bars as large as (or larger than) those seen in S0-Sb galaxies, but rarely any as small as those in Sc-Sd galaxies.