Abstract Satellite measurements are used to evaluate the glaciation, particle shape, and effective radius in cloud-resolving model simulations of tropical deep convection. Multidirectional polarized reflectances constrain the ice crystal geometry and the thermodynamic phase of the cloud tops, which in turn are used to calculate near-infrared reflectances so as to constrain the simulated ice effective radius, thereby avoiding inconsistencies between retrieval algorithms and model simulations. Liquid index values derived from Polarization and Directionality of the Earth’s Reflectances (POLDER) measurements indicate only ice-topped clouds at brightness temperatures (BTs) lower than −40°C, only liquid clouds at BT > −20°C, and both phases occurring at temperatures in between. Liquid index values calculated from model simulations generally reveal too many ice-topped clouds at BT > −20°C. The model assumption of platelike ice crystals with an aspect ratio of 0.7 is found consistent with POLDER measurements for BT < −40°C when very rough ice crystals are assumed, leading to an asymmetry parameter of 0.74, whereas measurements indicate more extreme aspect ratios of ~0.15 at higher temperatures, yielding an asymmetry parameter of 0.84. MODIS-retrieved ice effective radii are found to be 18–28 μm at BT < −40°C, but biased low by about 5 μm owing primarily to the assumption of pristine crystals in the retrieval. Simulated 2.13-μm reflectances at BT < −40°C are found to be about 0.05–0.1 too large compared to measurements, suggesting that model-simulated effective radii are 7–15 μm too small. Two simulations with contrasting ice nucleation schemes showed little difference in simulated effective radii at BT < −40°C, indicating that homogeneous nucleation is dominating in the simulations. Changes around −40°C in satellite observations suggest a change in cloud-top ice shape and/or size in natural deep convection possibly related to a change in the freezing mechanism.