Heed the data gap: Guidelines for using incomplete datasets in annual stream temperature analyses

•Incomplete temperature datasets may affect interpretation of watershed process and resilience.•Missing up to 7 to 9 consecutive weeks of data still produces accurate thermal signal estimates.•Imputation techniques could extend the applicable missing data periods by up to 13 weeks.•Timing of missing data affected thermal signal estimates, but inconsistently in space and time.•The methods described here can be efficiently applied to readily available data. Stream temperature data are useful for deciphering watershed processes important for aquatic ecosystems. Accurately extracting signal trends from stream temperature is essential for predicting responses of environmental and ecological indicators to change. Missing data periods are common for various reasons, and pose a challenge for scientists using temperature signal analysis to support stream research and ecological management objectives. However, the sensitivity of estimated temperature signal patterns to missing data has not been thoroughly evaluated, despite the potentially large impact on interpretation. In this study, we explored the effects of simulated missing daily data on the characterization of annual water temperature signals measured at headwater sites in the Pacific Northwest and Mid-Atlantic regions of the USA. For each site, we used linear regressions of sine-waves fitted to complete (365-d) and partial (7–357 consecutive missing data points) annual datasets of daily mean water temperature and computed three thermal parameters (mean, phase, and amplitude), which together can indicate thermally and ecologically influential watershed processes (e.g., depth and magnitude of groundwater discharge). Expected values (derived from complete datasets) ranged from 7.0 to 12.6 °C, 205 to 254 d, and 1.9 to 9.5 °C for annual mean, phase, and amplitude, respectively. While annual phase and amplitude could be accurately estimated (i.e., within 95–99% confidence intervals of expected values) with up to approximately two months of consecutively missing data, annual mean temperature required more complete datasets. We found that datasets with less than seven weeks of consecutively missing data enabled estimation of all annual signal parameters with reasonable accuracy (>75% probability of being within the 95–99% confidence intervals of expected values). Imputation of missing data expanded this range to approximately 20 weeks, with the greatest improvements in parameter estimation between 9 and 27 weeks of