Extended Data Fig. 9: Sensitivity analysis of annual structural equation models fitted to the balanced dataset.

Panel (a) illustrates the original structural equation model (SEM) structure (Fig. 5, fitted to the full dataset). The balanced dataset yielded an acceptable SEM fit to the original model in (a) Fisher’s C = 25.05, P = 0.199, df = 20, n = 120, AIC = 4,939, R²_m = 0.64, R²_c = 0.71 (NEE) but interactions between GPP ~ SW + DtNUP, R_e ~ T + VPD, and NEE ~ GS Length were no longer significant (p > 0.05, red lines in a). Retained pathways (blue lines in a) were fitted to the balanced dataset alongside R_e ~ GS Length, which was identified as a missing pathway in the balanced SEM. Standardized path coefficients are shown for the balanced SEM in (b), adjacent to arrows with line thickness indicating the relative strength of relationships. Significance levels for path coefficients (solid black arrows in b), p < 0.0001: ****, p < 0.001 ***, p < 0.01 **, p < 0.05 *) are based on two-sided tests, with no multiple comparison adjustments. The refitted SEM to the balanced dataset (b) had a better global fit (SEM diagnostics shown in plot) than (a) but yielded a higher AIC due to the original SEM in (a) having better explanatory power for R_e (a: R²_m = 0.51; b: R²_m = 0.38).