This paper analyzes the impact of high order light reflections on indoor optical wireless communication (IOWC) channel models. Based on observing the results of computer simulations, a calibration method is proposed to reduce model errors. Channel models are generated by tracing and adding up diffuse light reflections and sequential sub-reflections along its traveling path. As computation complexity increases significantly with the number of reflection orders considered, researchers traditionally, though incorrectly, take the contribution of first a few reflection orders, most commonly three, to represent the complete channel. Discarded high-order reflections bring no significant performance difference to low-speed transmission systems; however, major contemporary IOWC research institutions focus on high-speed Gigabits per second (Gbps) communications and the model errors resulting from discarded high-order reflections are no longer negligible. This is where the importance of our proposed method lies. root-mean-square (RMS) delay-spread, for instance, is severely underestimated by neglecting higher-order reflections. We simulate an IOWC system in an ordinary 6 m × 6 m × 3 m room and calculate the contributions of each order of reflections at 841 locations. It shows the RMS delay-spread estimation using the first three orders underestimates the true value by 15.3% on the average and by at most 26.6% as maximum. To limit error within half a symbol period, 1 Gbps and 10 Gbps systems tolerate underestimations up to 13.7% and 1.4%, respectively. These must be achieved by applying first five and nine orders. To maintain the computation efficiency of low-order reflection models and improve their accuracies, we propose a statistical calibration method. It reduces average model error of first three reflection orders from 15.7% to 4.3%. The numbers of orders required by 1 and 10 Gps systems are individually reduced to 3 and 7.