Geothermal resources assessments use data that are unevenly distributed in space, with more data collected in areas with known thermal features. To meet the assumptions for geostatistical modeling (e.g., variography and kriging) such as having a random sample representative of the population, declustering may be needed to correct for spatial sample bias. Several declustering methods exist and to understand how best to use these methods, we apply these to real data and samples of that data. The work described herein summarizes the application of cell-based declustering to shallow temperature data (~20 cm) collected in a survey across a thermal feature in the Lower Geyser Basin, Yellowstone National Park, Wyoming. The sample dataset is a regular grid (3-m spacing) of temperatures across a 72-m square area, providing a shallow, subsurface temperature dataset collected with minimal spatial bias (a few grid locations near a hot spring could not be sampled). To test the influence of sample clustering on geothermal estimates, this dense dataset is sub-sampled irregularly to evaluate bias on temperature estimation. Three sampling strategies were tested: a simple random sample, a stratified random sample, and a stratified biased random sample. The naive mean (before declustering) values for each dataset were compared to the post-declustering mean to evaluate the effectiveness of declustering on correcting the mean for spatial bias. For the limited number of sample datasets evaluated, we found that although cell-based declustering did partially correct the mean, some bias remained (i.e., the estimate was improved, but not fully corrected). It is possible that the procedure documented herein (applied here to only a few random samples) could be applied to many random samples, so that robust conclusions might be drawn (e.g., Is there always some remaining bias in declustered estimates? Does it depend on the number of sample points?). In particular, bias could be evaluated for persistency, and uncertainty could be evaluated.