This paper proposes a semi-parametric method for poverty decomposition, which
combines the data-generating procedure of Shorrocks and Wan (2004) with the Shapley
value framework of Shorrocks (1999). Compared with the popular method of Datt and
Ravallion (1992), our method is more robust to misspecification errors, does not require
the predetermination of functional forms, provides better fit to the underlying Lorenz
curve and incorporates the residual term in a rigorous way. The method is applied to
decomposing variations of urban poverty across the Chinese provinces into three
components – contributions by the differences in average nominal income, inequality
and poverty line. The results foreground average income as the key determinant of
poverty incidence, but also attach importance to the influence of distribution. The
regional pattern of the decomposition suggests provincial groupings based not entirely
on geographical locations.