SUMMARY AND CONCLUSIONS 



By alternating an autocorrelation analysis and an harmonic analysis of 

 certain time-series of daily sea-surface temperatures, an adequate estimator of 

 these temperatures has been determined. A judicious selection of regression 

 variables would have made the autocorrelation analysis unnecessary, but such an 

 analysis does provide clues to the nature of the time-series. 



For all stations considered, a regression model containing annual and 

 semiannual oscillatory terms (sines and cosines) provides a good statistical fit 

 to the observed daily temperatures. Some of the stations have nonrandom mis- 

 sing data, generally in the winter. Reference 1 examines in detail the effect of 

 this type data on the variances of regression coefficients and autocorrelation co- 

 efficients. These results are extended to the residual variances used in this 

 report. The correction for nonrandom missing data increases the residual vari- 

 ance. Correspondingly, F-ratios are reduced. The effect is conservative. That 

 is, one is less likely to reject null hypotheses after the correction is made than 

 before. 



The question of the existence of trends in ocean temperatures is an impor- 

 tant one. Several statistical tests for trend were performed on the sequences of 

 annual mean sea-surface temperatures, and on the sequences of amplitudes and 

 phases describing the regression functions. No trends were discovered to exist in 

 any of the sequences. It should be pointed out that, if trends did exist, it would 

 be a straightforward statistical problem to isolate their effect on the time-series. 



RECOMMENDATIONS 



1. An analysis of variance of the annual mean sea-surface temperatures 

 indicates that there are real year-to-year differences in the means. In the light 

 of these differences it is recommended that an investigation be made into the 

 length of time-series necessary to produce reliable long-term estimates of sea- 

 surface temperatures. 



2. The yearly means behave like random variables. Shorter sequences 

 of means with a systematic or cyclical appearance may occur by chance. It is 

 recommended that an investigation be made into the frequency of occurrence of 

 such unusual short sequences. 



28 



