VENRICK: SYSTEMATIC SAMPI.IN'd IN ECOSYSTEM 



TEMPERATURE (°C) 



15 20 25 



I — I — I — I — I — I — I — \ — \ — \ — \ — \ — I 



CHLOROPHYLL (mg/m3) 

  05 .10 .15 .20 .25 



I I I I I I I I I M I I I I I M I I I I I I I 



value and may be measured as a percent of the 

 true value ( 0): 



50 



CL 100 



UJ 

 Q 



,- kAh 



- A 



- A 



- A 



- A 



hAH 



hAH 



- A 



150 



hAH 



Figure 3. — Chlorophyll values observed in two 18-bottle casts 

 preliminary to study B, together with the temperature trace 

 from the associated STD lowering. Triangles indicate location of 

 samples in restricted systematic design 1; bars represent the 

 boundaries of strata used in stratified random design 3. 



paired, taken within a few hours and within 3 

 n.mi. of each other. These have been considered 

 replicate casts. 



When the combined systematic-stratified ran- 

 dom design called for bottles to be spaced more 

 closely than 3 m, it was necessary to use a mes- 

 senger heavier than the standard Nansen mes- 

 senger (such as a Niskin bottle messenger) in 

 order that it develop enough momentum to trip the 

 second bottle. When both sampling designs called 

 for the same depth, the extra bottle was arbitrarily 

 positioned, usually filling in the largest gap in the 

 region of the chlorophyll maximum layer. This 

 "free" sample was used only in the calculations of 

 total chlorophyll in the water column. 



Statistical Procedures 



Bias is evaluated by the consistency with which 

 n observations (.v, , / = 1, n) from a given sampling 

 design fall above or below the true population 



2 iXi-9) 



n 



X 100% 



le. 



Precision is measured by the variance of a series of 

 n observations about their mean (x): 



^j ^X j' X) 



In an analogous way, accuracy is measured by the 

 mean square deviation of a series of observations 

 from the true population total: 



2 {x,-Qy 



n 



Both accuracy and precision are inversely related 

 to their statistical measures, increasing as the 

 numerical value of the measure decreases. Since 

 most scientists are used to thinking in terms of 

 variances and sums of squares, it did not seem 

 desirable to invert these measures to achieve di- 

 rect correspondence. 



In the analysis of the results, limited use was 

 made of the parametric analysis of variance. Most 

 statistical tests were nonparametric tests which 

 make few assumptions about the characteristics of 

 the data (e.g., Dixon and Massey 1957; Tate and 

 Clelland 1957; Conover 1971; Hollander and 

 Wolfe 1973). Unless stated otherwise, the prob- 

 abilities associated with conclusions in the text 

 are derived from the binomial distribution withp 



= y-i. 



In several analyses in these studies, the problem 

 of multiple testing arose, as when all four sys- 

 tematic designs were tested for bias. Unfortu- 

 nately, the tabulation on most nonparametric 

 procedures is not sufficiently complete to allow 

 correction for multiple testing to be made without 

 making the tests extremely conservative. Since 

 this was deemed undesirable, the probabilities 

 given for the statistical tests are uncorrected. It is 

 unlikely that this makes any real difference in the 

 outcome of these studies which gain most of their 

 force from the similarity of results in the two ap- 

 proaches. 



621 



