The author analyses the estimation quality of unemployment spell distribution on the basis of economic activity data (BAEL) - collected in Poland in 2001-2002 for Eurostat and discusses the adequacy of formal model assumptions needed for standard econometric inference on the distribution. The data show high and possibly selective non-response censoring, which collides with standard formal assumptions needed for a reliable estimation. The Kaplan-Meier estimation was applied to the original data, where non-response censoring during the study was treated as independent of time variable, and to the modified data, where non-response was taken equivalent to unemployment termination. A large discrepancy between the two estimations, giving the region of actual unemployment time distribution, was observed. Similar effects were noticed in a supplementary simulation experiment. The estimations applied to four Polish provinces show no clear relationship between the local unemployment rate and the flexibility of labor market described by the unemployment spell distribution. Difficulties in a reasonably accurate estimation of the unemployment time distribution for the BAEL data are in the absence of information explaining individual's non-response during the study, and also in frequent illogical partial information necessary for the final determination of individual's active job time search.