VENRICK: SYSTKMATIC SAMF'LINt; IN KCOSYSTEM 



While it is recognized that the results suffer from 

 small sample sizes, they gain considerable force 

 from the fact that two different approaches give 

 quite similar conclusions. 



It appears that RSS, when applied to the vertical 

 distribution of chlorophyll, may actually give es- 

 timates of total chlorophyll which are, on the av- 

 erage, closer to the true value than are estimates 

 based on SR. This is consistent with results from 

 terrestrial systems. Stratified random designs 

 frequently result in pairs of samples falling closely 

 adjacent to one another. In populations which are 

 varying continuously, the information contained 

 in such an adjacent pair is largely repetitious. 

 Such redundancy is avoided in RSS because the 

 spacing between adjacent samples is controlled. 

 The relative performance of SR is expected to im- 

 prove in populations with more discontinuous dis- 

 tributions such that the strata may be defined to 

 be internally homogeneous, giving maximum pre- 

 cision. On the other hand, and perhaps more im- 

 portant, there appears to be a potential for bias in 

 systematic designs, especially when the sampling 

 occurs within restricted spatial or temporal inter- 

 vals. In the central Pacific, which is relatively 

 homogeneous in time and space, the bias was de- 

 tectable, but the magnitude was small. Unfortu- 

 nately, the results cannot be generalized to other 

 environments. The bias may be expected to di- 

 minish as increasing environmental complexity 

 increases the small-scale variability of the sam- 

 pled population. Whether or not bias is therefore 

 negligible in more complicated neritic environ- 

 ments remains to be investigated. 



On the other hand, in planktonic environments 

 the natural fluctuations produce variability be- 

 tween replicate casts which is generally large rel- 

 ative to the experimental error associated with a 

 single cast. The increased accuracy of systematic 

 designs is not likely to result in a detectable in- 

 crease in the precision of the estimate of the mean 

 of several samples. Depending upon the goals of a 

 study, it may also be true that the magnitude of 

 the bias introduced by systematic sampling is in- 

 significant. For instance, it appears that RSS as 

 routinely used on large-scale oceanic surveys 

 probably introduces no serious error. However, 

 increasing attention is being focused on small- 

 scale, local phenomena, and the routine use of RSS 

 for these studies deserves examination. Possible 

 effects of the interaction between bias and sam- 

 pling scale found in this study include overestima- 

 tion of fluctuations in total population ( if a positive 



bias occurs with higher populations and a nega- 

 tive bias with lower populations), underestima- 

 tion of fluctuations (if a positive bias occurs with 

 lower populations and a negative bias with higher 

 populations), or production of artificial fluctua- 

 tions when in fact the population is stable. 

 Whether or not such artificial effects of RSS might 

 be important enough to overshadow the gain af- 

 forded in terms of ease of sample location and data 

 analysis depends upon the particular study under 

 consideration. 



Study A demonstrated the dependence of the 

 success of a sampling design upon the interaction 

 of that design with the structure of the population 

 being sampled; thus, it would seem that intelli- 

 gent application of knowledge about the sampled 

 population should improve the design. It was, 

 therefore, disconcerting to find that RSS every 20 

 m, RSS-1, consistently performed as successfully 

 as did RSS-3 which was designed by a presumably 

 experienced worker (the author) with total knowl- 

 edge about the population to be sampled. We must 

 conclude either that the location of samples in 

 RSS-3 was not as intelligent as it might have been, 

 or that the natural variability of the population 

 makes intelligent placement of systematic sam- 

 ples impossible. (The latter interpretation has a 

 certain appeal.) In contrast, the dependence of SR 

 on the selection of strata is apparent in both 

 studies. Many narrow strata give more precise and 

 accurate estimates than do fewer, larger ones, un- 

 doubtedly because, in the presence of strong verti- 

 cal gradients, they better satisfy the criterion of 

 internal homogeneity. The most precise and accu- 

 rate results are obtained when the number of 

 strata is equal to the number of samples. The dis- 

 advantage of this strategy is the absence of direct 

 estimates of variability within strata from which 

 to calculate confidence intervals around the final 

 estimate. This is not usually of concern in 

 planktonic work because small-scale patchiness is 

 of such magnitude that more useful estimates of 

 precision are obtained directly from replicate 

 casts, and thus apply to some spatial and temporal 

 interval, rather than to a single cast. 



The logistics of SR in the open ocean presented 

 few major difficulties. The task of allotting sam- 

 ples randomly to strata was time consuming. After 

 the first few casts, the job of sample design was 

 relegated to the computer. Preparation of the cast 

 card demanded more than routine caution (al- 

 though minor errors were readily assimilated into 

 the randomization procedure). The use of random 



625 



