A stochastic simulation of a pig breeding unit with 100 sows and 15 boars was used to examine the effects of culling before testing on the estimates of (co)variance components and parameters. The traits were age at test (AGE) and sidefat thickness (FAT) measured on farm. Data were simulated over 10 years and missing data were generated by deletion of records based on the phenotypic value of AGE. A situation with no culling (C-NO) was compared with culling at random (C-RAND), within litters (C-W/IN) and over litters (C-OVER). A multivariate restricted maximum likelihood procedure was applied to estimate (co)variance components, using the same model as that for the simulation of records. The heritabilities of AGE were over estimated in C-W/IN and C-OVER, and genetic, residual and phenotypic correlations became weaker. The litter variance ratio of AGE was seriously overestimated in C-W/IN and underestimated in C-OVER. A reduced model, without the litter effect included, resulted in even larger differences in parameter estimates between culling strategies.