We have localised the three OAs to high precision in the STIS images and identify the host as the nearest detected object of the OA position. The GRB 980329 host galaxy redshift is estimated to be $z \approx 3.5$ . For all three candidate hosts we detect faint extended structures within a radius of $\sim$ $1\hbox{$^{\prime\prime}$ }$ ( $\sim$ 7 proper kpc for z >0.5). This scale is similar to that seen between tidally interacting and merging galaxies (e.g. Borne et al. 2000). The hosts show signs of sub-structure (possibly star-forming and/or merging elements); 3-4 blue knots in the GRB 980329 host, a knot in the northern edge of the GRB 980519 host and a point-source within the 3- $\sigma$ localisation error of GRB 990308. The faintness of these hosts suggests that, regardless of the host luminosity, GRBs seem to be associated with star formation (SF). This correlation may allow GRBs to be used as a powerful tracer of star formation, provided the link between GRBs and SF is correct. It also implies that a significant amount of star formation is located in (optically) very faint galaxies. Their faintness may be due to dust extinction or a very steep faint-end slope of the galaxy luminosity function (see also Fynbo et al. 2002). The SF/GRB correlation may therefore be the basis of a unique and new way of finding star-forming galaxies at high redshifts (independent of the otherwise unavoidable surface-brightness bias).

Bloom et al. (2002) present a comprehensive study of 20 GRB hosts, including the three reported here, based on the same HST data and mainly Keck imaging. For GRB 980329 they found an OA position similar too ours, but with somewhat larger errors. However, their selection of the host center coincide with a small object at the center of the host complex, whereas we have chosen the galaxy most nearby the localization. From their thumbnail image (Fig. 2) it seems the northern object is much less significant than in our image and likewise the northern most blob/knot, which is practically coincident with our OA localization. For GRB 990308 Bloom et al. (2002) obtain an OA localisation with significantly larger errors than ours, and identify a large galaxy in the outskirts of their error circle (indicated in Fig. 2). Our host candidate corresponds to the object(s) seen left of the center of the error circle in their image, and which reality was noted by them as questionable. In summary, we identify different host candidates in two out of the three faint hosts investigated. These three hosts are also among the four most extreme outliers (the ones with the largest OA to host-center offset) in the Bloom et al. sample. Adopting our host identifications and OA localisations, the sample does not have obvious outliers. The relevant parameters as measured on our drizzled CL-band images are given in Table 4 including the individual $P_{\rm i}$ values for direct comparison with Bloom et al. (2002). This is quantified by re-computing $P(n_{\rm chance}=0)$ , defined in Bloom et al. (2002), representing the probability that none of the host identifications of the Bloom et al. sample are randomn galaxies (unrelated to the GRB). Bloom et al. (2002) found $P(n_{\rm chance}=0) = 0.483$ , but by using our OA localisations and host identifications we find $P(n_{\rm chance}=0) = 0.69 (0.67)$ . This means that by using our results it is unlikely that any of the 20 hosts in the Bloom et al. sample are false identifications.

Table 4: Measured properties of the host candidates used to estimate the $P_{\rm chance}$ statistic. The Rc magnitude includes the line-of-sight extinction.
Target R₀ $\sigma_{{R}_0}$ $R_{\rm {half}}$ Rc mag $P_{\rm i}$

980329 host 0.16 0.046 0.14 27.0 0.022

980519 host 0.41 0.025 0.11 27.4 0.061

990308 host1 0.048 0.088 0.023 29.4 0.105

990308 host2 0.30 0.088 0.063 29.3 0.125

Target	R₀	$\sigma_{{R}_0}$	$R_{\rm {half}}$	Rc mag	$P_{\rm i}$
980329 host	0.16	0.046	0.14	27.0	0.022
980519 host	0.41	0.025	0.11	27.4	0.061
990308 host1	0.048	0.088	0.023	29.4	0.105
990308 host2	0.30	0.088	0.063	29.3	0.125

7 Summary