Financial leverage is weakly but consistently negatively associated with profitability in a cross-section of Polish firms (2008-2010)

Summary

Using audited financial statements of Polish firms (2008-2010), we pre-registered and tested a single association: whether financial leverage (total debt / assets) is rank-correlated with profitability (return on assets, ROA = net profit / assets). In the primary 2010 cross-section the association is negative, weak, and statistically significant (Spearman rho = -0.058, 95% CI [-0.092, -0.024], n = 3581, p = 5.03e-04). The direction replicates in every other cross-section and survives outlier handling and firm-level aggregation, but the effect size is small in the most recent year. This is an associational finding only; no causal interpretation is made.

Provenance and methods

Source. A single pinned spreadsheet of Polish firm financial statements (DOI 10.6084/m9.figshare.92633.v1, licence CC BY 4.0), sheet MPB2008-2010PolishFirmsData. The analysis script downloads the file from https://ndownloader.figshare.com/files/94217 and verifies its MD5 before use; observed MD5 2701c5774c800badaad5b29c31a661d0 matches the pinned value 2701c5774c800badaad5b29c31a661d0 (size 3396741 bytes). If the canonical download is unreachable, an MD5-gated local copy is used (the same checksum is enforced, so the bytes analysed are identical either way).

Variables. ROA = Net profit (loss) / Assets; Leverage = (Long-term debt (Dl) + Short-term debt (Ds)) / Assets.

Inclusion. A row enters a cross-section if its year label matches and Assets > 0 (needed for finite ratios). The file contains 14421 rows across year labels 2008, 2009, 2009b and 2010; 1 row has a missing year (and is therefore in no cross-section). The 2009b label is a largely separate collection (few registry-number matches with 2009) and is treated as a separate batch: reported but never pooled with, or multiple-comparison-corrected alongside, the primary family [2008, 2009, 2010].

Tests (pre-registered). Primary: two-sided Spearman rank correlation of Leverage vs ROA in 2010 (H1: rho < 0). Robustness: (A) temporal replication in 2008 and 2009 with Holm-Bonferroni correction across the family [2008, 2009, 2010]; (B) winsorization/trimming at 1/99% to rule out tiny-denominator artifacts; (C) firm-level aggregation (one mean observation per registry-identified firm) as a pseudoreplication guard. Spearman is used throughout because both ratios have heavy, non-normal tails; 95% confidence intervals for rho are bootstrap percentile intervals (10000 resamples). Global seed = 20240517.

Data records

Derived tables are in tables/ with a Frictionless datapackage.json describing each field:

• tbl-1-sample-profile.csv - per-year rows, exclusions, ROA/Leverage medians and IQRs.
• tbl-2-primary-result.csv - primary Spearman result (2010).
• tbl-3-year-replication.csv - Spearman by year, Holm-corrected p (2009b as separate batch).
• tbl-4-winsorization.csv - primary Spearman under raw/winsorized/trimmed handling.
• tbl-5-firm-level.csv - Spearman on firm means.

Every statistic is also recorded in results.json. Figures are in figures/ (one per test, >=150 DPI, colourblind-safe).

Technical validation

• Integrity: MD5 of the analysed bytes is verified equal to the pinned value

on every run; the script aborts on mismatch.

• Determinism: the seed is fixed (20240517); re-running the script

regenerates byte-identical results.json (verified across two runs).

• Robustness of the primary result: the 2010 rho is essentially unchanged

under winsorizing (rho = -0.058) and trimming (rho = -0.060), so it is not an outlier artifact.

• Replication: the negative association holds in 2008

(rho = -0.212, Holm p = 5.44e-38), 2009 (rho = -0.068, Holm p = 1.01e-04), the separate 2009b batch (rho = -0.200), and at the firm level (rho = -0.186, n = 5710 firms, p = 1.94e-45).

Usage notes

The effect is weak in the most recent cross-section (rho ~ -0.06; the shared variance is well under 1%), so leverage is at best a minor rank-predictor of profitability in 2010 even though the sign is stable and highly significant in the larger 2008 and firm-level samples. Because ROA and Leverage share the same denominator (Assets), a mechanical component to the association cannot be excluded; the finding should be read as a descriptive rank association, not evidence that leverage lowers profitability or vice versa. No firm-level panel model, industry control, or causal identification is attempted.

Code availability

analysis.py is self-contained: it downloads and MD5-verifies the source, sets all seeds, runs only the pre-registered tests, and writes every figure, table, results.json, and this narrative. environment.txt records the interpreter version and installed packages.

Claims

See claims.json. Each claim traces to a numeric field in results.json.