Previous: Introduction Next: Tidal effects Up: Ext. Abst.

2. Methodology and data requirements

 Data analyses for trend detection are based on linear regressions. The confidence of the trend coefficient depends upon the variance of the difference between the data and the fit model. This residual variability includes the short term fluctuations of the atmosphere and the measurement noise. The confidence is related to this short-term residual variance and to the length of the data set. In practice in the atmosphere, two successive measurements are correlated by the non-random changes such as waves. The determination of the trend confidence interval can take into account this non-independence data properties with autocorrelation coefficient. As more measurements will not improve trend estimates, one can reduce the confidence interval in reducing the residual atmospheric variance in averaging successive measurements. The optimum interval depends upon several factors and is somewhere between a week and a month.
However, the fundamental factor limiting our hability to detect and quantify trends is the instrumental bias inducing changes of the mean measurements. Most of the instrumental changes may impact on the bias and may generated spurious changes which can be interpreted as trends. So it is deeply recommended to well documented any instrumental changes for efficient trend data analyses.