The key to our investigation is an improved (and in a sense sharp) understanding of the survival time of the contact process on star graphs. Using these results, we show that for the contact process on Galton-Watson trees, when the offspring distribution (i) is subexponential the critical value for local survival $λ_2=0$ and (ii) when it is geometric($p$) we have $λ_2 \le C_p$, where the $C_p$ are much smaller than previous estimates. We also study the critical value $λ_c(n)$ for "prolonged persistence" on graphs with $n$ vertices generated by the configuration model. In the case of power law and stretched exponential distributions where it is known $λ_c(n) \to 0$ we give estimates on the rate of convergence. Physicists tell us that $λ_c(n) \sim 1/Λ(n)$ where $Λ(n)$ is the maximum eigenvalue of the adjacency matrix. Our results show that this is not correct.