Estimating multiple treatment effects in two-phase observational data