Trends in satellite-based temperature datasets created by different groups differ by up to 0.2 K/decade in the troposphere, well beyond the uncertainty estimates for the individual datasets. This discrepancy creates considerable uncertainty in our understanding of changes in vertical temperature patterns. Radiosonde temperature data could provide an alternative source of information, but are subject to similar uncertainties. Trends in global mean temperatures from the Lanzante-Klein-Seidel (LKS) and Hadley Center (HadRT) radiosonde datasets differ by up to 0.1 K/decade.

To narrow these uncertainties, it is important to understand the sources of differences between datasets. Reasons for differing trends include differences in spatial and temporal coverage, differences in the input data at individual locations, and differences in adjustments for inhomogeneities in the data. We have explored the role of each of these factors in the LKS and HadRT radiosonde datasets and the role of space and time sampling errors in the differences between radiosonde and satellite datasets.

To examine sampling effects on radiosonde dataset differences we compared trends in global and hemispheric mean time series from 71 locations common to the LKS and HadRT datasets, dropping months of data that were not present in both time series. We also compared unadjusted and adjusted versions of the datasets to assess the effects of the adjustments. Differences in large-scale mean trends between LKS and HadRT are due roughly equally to sampling differences, homogeneity adjustments and differences in input data. It follows that to narrow uncertainties in radiosonde temperatures we must deal with all three sources of disagreement.

To get another estimate of sampling effects, we subsampled (globally complete) NCEP reanalysis and MSU satellite data according to 6 radiosonde networks corresponding to existing datasets and compared the resulting large-scale means to those from the globally complete datasets. In this work, the larger sonde networks do not give consistently smaller spatial sampling error than the smaller networks. However, these estimates of coverage effects from reanalysis and satellite data differ noticeably from those seen in actual radiosonde data as well as from each other. Thus the usefulness of reanalysis and MSU datasets for estimating sampling error in radiosonde datasets is unclear.

Subsampling MSU satellite data according to the radiosonde coverage does not generally bring the trends in the MSU datasets significantly closer to those in the radiosonde datasets, and in some cases increases the differences. The differences between MSU and radiosonde global mean trends appear to be due primarily to factors other than large-scale differences in space and time coverage.