We revisit systematics in determining the local dark matter density ρdm from the vertical motion of stars in the solar neighbourhood. Using a simulation of a Milky Way like galaxy, we determine the data quality required to detect ρdm at its expected local value. We introduce a new method for recovering ρdm that uses moments of the Jeans equations, combined with a Markov chain Monte Carlo technique, to marginalize over the unknown parameters. Given sufficiently good data, we show that our method can recover the correct local dark matter density even in the face of disc inhomogeneities, non-isothermal tracers and a non-separable distribution function. We illustrate the power of our technique by applying it to Hipparcos data. We first make the assumption that the A- and F-star tracer populations are isothermal. This recovers ρdm= 0.003+0.009- 0.007 Msun pc-3 (ρdm= 0.11+0.34- 0.27 GeV cm-3, with 90 per cent confidence), consistent with previous determinations. However, the vertical dispersion profile of these tracers is poorly known. If we assume instead a non-isothermal profile similar to that of the blue disc stars from SDSS DR-7 recently measured, we obtain a fit with a very similar χ2 value, but with ρdm= 0.033+0.008- 0.009 Msun pc-3 (ρdm= 1.25+0.30- 0.34 GeV cm-3 with 90 per cent confidence). This highlights that it is vital to measure the vertical dispersion profile of the tracers to recover an unbiased estimate of ρdm.