Abstract

Commercial firm-level databases are not probability samples, and their selective coverage can distort estimates of unconditional population objects such as aggregate productivity and firm-size distributions. At the same time, researchers may want to work with a smaller analytical sample because the full database is costly to process or uneven in data quality. This paper proposes a benchmark-guided stratified subset-selection algorithm for this setting. Given a target sample size and official cell-level aggregate statistics, the method selects firms so that the retained sample matches official sector-by-size cell counts and aligns within-cell benchmark moments. The selected sample is then adjusted using mild post-stratification weights. The algorithm can be viewed as a sparse, design stage analogue of calibration. We evaluate the method using Italian manufacturing firms from Orbis in 2022, with OECD Structural Business Statistics as external benchmarks. In the raw Orbis pool, cell-weighted turnover per employee is overstated by about 14%. The proposed algorithm reduces this distortion to about 5% and produces near-uniform final weights, while post-stratification alone leaves the distortion essentially unchanged. Full-pool calibration reproduces the moments it constrains, but in this application it performs less well on non-targeted functionals, and bounded cell-level calibration becomes infeasible in many cells under finer stratifications. The results show that benchmark-guided subset selection can complement weighting-based corrections when researchers work with large non-probability firm databases and want a smaller sample aligned with official aggregate statistics.