Uncertainties are quantified in atmospheric temperature changes derived from satellites, radiosondes, and the reanalyses of the National Center for Environmental Prediction and European Center for Medium-Range Weather Forecasts (NCEP and ERA). To facilitate intercomparison, we compute from the reanalyses and radiosonde data deep layer temperatures equivalent to those estimated from the satellite-based Microwave Sounding Unit (MSU). Equivalent MSU temperatures generated using global mean weighting functions and a radiative transfer code give similar results. NCEP's pre-1979 global mean lower stratospheric temperature anomalies diverge markedly from radiosonde data. A smaller offset occurs in the midtroposphere. These differences are attributed to a likely warm bias in the tropical lower stratosphere in the temperature retrievals used by NCEP from November 1978 onward, and changes in the error characteristics of the assimilation model's simulation of the lower stratosphere. In the lower troposphere, ERA and NCEP show different global mean trends due to differences in assimilation strategy, input observational data, quality control procedures, and model physics. Over 1979--1993, ERA warms by 0.106 ¿C/decade, while NCEP cools by 0.028 ¿C/decade. Applying the HadRT1.1 (radiosonde) data availability mask to NCEP improves the agreement between these data sets. Neglecting coverage differences can yield misleading results in MSU-radiosonde trend comparisons. Substantial trend uncertainties also arise from coverage differences between various radiosonde data sets. Version c of the MSU lower tropospheric temperature retrieval fails to adjust explicitly for orbital decay. If this were applied without any additional adjustments, it would resolve an important discrepancy: in MSUc the lower troposphere has cooled in relation to the midtroposphere, while the reverse is the case for both reanalyses and for the radiosonde data examined here. ¿ 1999 American Geophysical Union