Intensity–duration–frequency (IDF) analyses of rainfall extremes provide critical information to mitigate, manage, and adapt to urban flooding. The accuracy and uncertainty of IDF analyses depend on the availability of historical rainfall records, which are more accessible at daily resolution and, quite often, are very sparse in developing countries. In this work, we quantify performances of different IDF models as a function of the number of available high-resolution (Nτ) and daily (N24h) rain gauges. For this aim, we apply a cross-validation framework that is based on Monte Carlo bootstrapping experiments on records of 223 high-resolution gauges in central Arizona. We test five IDF models based on (two) local, (one) regional, and (two) scaling frequency analyses of annual rainfall maxima from 30-min to 24-h durations with the generalized extreme value (GEV) distribution. All models exhibit similar performances in simulating observed quantiles associated with return periods up to 30 years. When Nτ > 10, local and regional models have the best accuracy; bias correcting the GEV shape parameter for record length is recommended to estimate quantiles for large return periods. The uncertainty of all models, evaluated via Monte Carlo experiments, is very large when Nτ ≤ 5; however, if N24h ≥ 10 additional daily gauges are available, the uncertainty is greatly reduced and accuracy is increased by applying simple scaling models, which infer estimates on subdaily rainfall statistics from information at daily scale. For all models, performances depend on the ability to capture the elevation control on their parameters. Although our work is site specific, its results provide insights to conduct future IDF analyses, especially in regions with sparse data.