out the erroneous observations. Since space and time are treated 

 simultaneously in the regression model rather than separately as 

 in the conventional space/time averaging approach, a single 

 observation is not overly weighted in the averaging process. 



Data taken in the 30 N-latitude strip for fiscal year 1950 will 

 be used to illustrate this editing technique. The original set of 

 raw data contained 199 observations taken in the 18-month period 

 centered on fiscal year 1950. The left-hand section of figure 18 

 shows the statistical results obtained by fitting equation (4) to this 

 complete data set. In addition, a histogram of the differences 

 between the sea-surface temperature obtained from the regression 

 equation and the observed value is presented. It is noted that 

 there are three differences greater than ±3 standard deviations and 

 eight differences greater than ±2 standard deviations. The original 

 data for these 11 observations were examined and in all cases real 

 errors were found. The correspondence suggests that gross 

 errors in data sets may be detected by means of a regression 

 model and eliminated by rejecting data whose differences are 

 greater than some multiple of the standard deviation. The center 

 and right-hand sections present results obtained by rejecting data 

 whose differences were greater than ±3 and ±2 standard devia- 

 tions, respectively. If, for one reason or another, it is not 

 desirable to eliminate the erroneous data, the regression technique 

 affords a method of rapidly identifying the data badly in error. 

 Once identified, data can be examined, corrected, and salvaged 

 for subsequent analysis. 



Several such analyses were made on different data sets and 

 in all cases the data identified by large differences were found to 

 contain real errors. 



71 



