Is the education-health gradient inflated because both education and health are associated with unobserved socio-emotional skills? Revisiting the literature, we find that the gradient is reduced by 30-45% by fine-grained personality facets and Locus of Control. Traditional aggregated Big-Five scales, in contrast, have a much smaller and mostly insignificant contribution to the gradient. We decompose the gradient into its components with an order-invariant method, and use sibling-fixed effects to address that much of the observed education-health gradient reflects associations rather than causal relationships. There are education-health gradients even within sibling pairs; personality facets reduce these gradients by 30% or more. Our analyses use an extraordinarily large survey (N=28,261) linked to high-quality administrative registers with information on SES background and objective health outcomes.