AbstractA number of recent studies have examined the statistics of tropical storms simulated by general circulation models (GCMs) forced by observed sea surface temperatures. Many GCMs have demonstrated an ability to simulate some aspects of the observed interannual variability of tropical storms, in particular, variability in storm frequency. This has led to nascent attempts to use GCMs as part of programs to produce operational seasonal forecasts of tropical‐storm numbers.In this study, the sensitivity of the statistics of GCM‐simulated tropical storms to changes in the model's physical parametrizations is examined. After preliminary results indicated that these statistics were most sensitive to details of the convective parametrization, GCM simulations with identical dynamical cores but different convective parametrizations were created. The parametrizations examined included moist convective adjustment, two variants of the Arakawa‐Schubert scheme, and several variants of the relaxed Arakawa‐Schubert (RAS) scheme; the impact of including a shallow‐convection parametrization was also examined.The simulated tropical‐storm frequency, intensity, structure, and interannual variability were all found to exhibit significant sensitivities to changes in convective parametrization. A particularly large sensitivity was found when the RAS and Arakawa‐Schubert parametrizations were modified to place restrictions on the production of deep convection.Climatologies of the GCM tropical atmosphere and composites of tropical storms were examined to address the question of whether the tropical‐storm statistics were directly impacted on by changes in convection associated with tropical storms, or if they were indirectly affected by parametrization‐induced changes in the tropical mean atmosphere. A number of results point to the latter being the primary cause. A regional hurricane model, initialized with mean states from the GCM simulation climatologies, is used to further investigate this point. Particularly compelling is the fact that versions of the RAS scheme that produce significantly less realistic simulations of tropical storms nevertheless produce a much more realistic interannual variability of storms, apparently due to an improved tropical mean climate.A careful analysis of the background convective available potential energy (CAPE) is used to suggest that this quantity is particularly relevant to the occurrence of tropical storms in the low‐resolution GCMs, although this may not be the case with observations. If the tropical CAPE is too low, tropical storms in the low‐resolution GCMs cannot form wit