Chapter 1
Introduction
Key Points
âą The environment is continuous; only by sampling can one obtain quantitative information about it.
âą The environment is also variable; sampling must therefore be replicated to cover its variability.
âą Sampling according to formal design as distinct from haphazard choice is necessary to provide unbiased information.
âą Sampling should be sufficient for the ensuing statistical analysis of the sample data, to extract the information sought with reasonable confidence.
Why sample?
The human race depends on many complex, interconnected and vulnerable environmental systems â atmospheric, marine, terrestrial and freshwater. There are few, if any, such systems that are not affected by human activity. We cultivate soil, clear forests, emit gases and particulates into the atmosphere, and we use the sea as a source of food and a sink for our wastes. At the same time we know that these systems must be conserved, and that awareness requires that we understand them, that we know their properties, and that we can monitor and predict their behaviour.
We cannot achieve these aims without quantitative information. Whether we are policy makers who want to know how much carbon has been lost from the countryâs soil, regulators who want to know the concentration of arsenic in the groundwater of a particular aquifer, or scientists who want to assess the difference between the rates of emission of nitrous oxide gas from fields under different management, we need reliable information, and we need to obtain it at reasonable cost.
Two difficulties immediately arise. The first is that we cannot measure the properties in the whole of the environment; we cannot measure carbon in all the soil, we cannot analyse all the ground water in a region for arsenic or measure all the nitrous oxide in the air. The best we can do is to make our measurements on small amounts of air, water or soil, that is, on samples, and hope that these measurements are representative. Whatever information we have on the environment is necessarily fragmentary. If we want quantitative information on the environment we have to sample.
Variability, replication and bias
The second difficulty arises because the composition of the air, the amount of arsenic in groundwater and the carbon content of soil vary from place to place. A single measurement on a small volume of any of these materials could give a false view of the generality. An exceptionally small value of arsenic in the water from a particular borehole could indicate that the water everywhere was safe to drink when in fact it was not; an exceptionally large value might suggest that nowhere was it safe to drink the water. An exceptionally small concentration of phosphorus measured in a single soil core might cause a farmer to apply too much fertilizer over the whole of a field or farm. The excess phosphorus might then find its way into water courses and pollute them. The farmer faced with an exceptionally large value might decide not to fertilize at all and as a result get a poor yield of the crop and lose income. An exceptionally large measurement of a pollutant such as dioxin in the soil could prompt a local authority to attempt to clean the whole of a site where only âhot spotsâ were contaminated, and so place an unnecessary burden on its tax payers.
This second difficulty would not be wholly overcome even if the individual sample in each case were not exceptional, that is, if it were âtypicalâ. Such a sample might give a better idea of the generality than the exceptional sample, but it would give no idea of the variation present, and again it could give a false sense of security. Some people in a region could be drinking water containing toxic amounts of arsenic even though most were not. Phosphorus could leak from the few places where amounts were grossly in excess of cropsâ requirements, as beneath former manure-heaps. The air close to a busy road junction could be seriously polluted while the air elsewhere in a city was typically clean. In many instances we also want to know where exceptional conditions exist and to map the variables of interest. Mapping is an important step in understanding our environment and responding to it.
In short, a single measurement of any property is never an adequate sample because we cannot be sure that it is representative, and all but the most trivial problems require some assessment of the variability of the environmental properties of concern. The solution is to use replicated sampling, that is to say, samples comprising more than one small body of material to assess variation in the environment.
Arsenic in groundwater in Bangladesh
We have mentioned arsenic (As) in water with good reason. In Bangladesh 97% of the population draws on the groundwater for drinking. There are estimated to be between 6 million and 11 million boreholes sunk for the purpose, most to depths between 10 m and 150 m. In many places the water contains potentially toxic concentrations of arsenic: the concentration exceeds the Bangladesh standard for the maximum safe concentration 50 ÎŒg As litreâ1 or the more stringent 10 ÎŒg As litreâ1 of the World Health Organization (WHO). The arsenic is of natural origin, sorbed on iron oxides in the young (Holocene) alluvial and deltaic sediments that form so much of Bangladesh. It seems to be released into solution in the anaerobic conditions underground and so contaminates the water.
In the late 1990s the National Hydrochemical Survey of Bangladesh sampled the water from more than 3 000 boreholes less than 150 m deep in an area of approximately 129 000 km2 to assess the magnitude of the situation. Table 1.1, compiled from the British Geological Survey and the Department of Public Health Engineering (BGS and DPHE, 2001) and Gaus et al. (2003), summarizes the findings.
Table 1.1 Summary of arsenic concentration (in ÎŒg litreâ1) in the groundwater of Bangladesh
|
Number of observations | 3208 |
Mean | 60.5 |
Median | 6.08 |
Standard deviation | 123.1 |
|
Percentiles | |
10 | 0.1 |
20 | 0.2 |
30 | 0.6 |
40 | 2.0 |
50 (median) | 6.1 |
60 | 16.6 |
70 | 40.4 |
80 | 83.4 |
90 | 199.7 |
95 | 320.0 |
99 | 570.7 |
|
Table 1.1 reveals immediately several aspects of natural variation and its likely impact on our response to it. The concentration of As in 30% of all boreholes seems trivially small (†0.6 ÎŒg litreâ1) against the WHO guide value, and even the median 50% is less. However, between 40% and 50% (45.2% according to Gaus et al. ) have concentrations exceeding 10 ÎŒg litreâ1, and in between 20% and 30% (27.2% according to Gaus et al. ) the concentration exceeds the Bangladesh limit of 50 ÎŒg litreâ1.
Clearly, any one measurement would be grossly misleading. In more than half of the boreholes the water is safe to drink; in many others it is not. Nor is the mean much of a guide. At 60.5 ÎŒg litreâ1 it exceeds the Bangladesh limit. If one took it to represent the state of the groundwater nationally, then all bore holes should be closed down.
The immediate response to such a finding might be that any, indeed every, borehole should be tested for arsenic. But there is a second aspect of the situation that Table 1.1 cannot reveal, namely where the water is toxic. As it happens, the occurrence of toxic concentrations is patchy, and, using advanced statistical technique, Gaus et al. mapped the distribution of As from the sample data and showed that in large parts of the country there was little likelihood of toxicity. So priority for testing could be given to those regions where the likelihood was greatest.
We need more than one measurement for another reason, namely to give us confidence in our assessments of environmental variables. Environmental scientists commonly use averages to characterize environmental variables â for example, the average concentration of arsenic in groundwater, the average carbon content of the soil. Provided a sample is properly representative, as mentioned below, we can use the arithmetic average of the measured values to estimate the true mean. A mean value calculated from several measurements should give us a more reliable estimate than a single measurement, and provided the sample is chosen according to certain rules we should be able to place the true mean within bounds. Also, we need more than one measurement to make a prediction, whether about the temperature next year, about the strength of the soil ahead of a tractor, or about the arsenic concentration at a place where we might sink a borehole. Again, we want to place our predicted values within bounds with confidence.
Intuitively we might feel that the larger are our samples, the more confidence we can have in our estimates and predictions, and in general we are right to do so. Large samples in general enable us to make more sensitive comparisons between environments than do small ones. For example, we should expect a mean temperature based on data from 30 weather stations in a region to be more reliable than one based on only three. If the difference between two years were only 2°C we should almost certainly want data from more than three stations to say with confidence whether year 2 was warmer than year 1. This seems to be common sense. But then we want to place our notions of reliability and confidence on a quantitative footing. Can we do so? The answer is yes; we can do it with statistics. Statistics enables us to estimate not only mean values and predict actual values, it also enables us to put errors to those estimates and predictions and to place quantitative confidence intervals about them. It enables us to express uncertainty in a quantitative way.
Replication is crucial. It can also be expensive. Getting from place to place in a large region with few roads to visit sampling sites can cost a great deal in time and money. Measuring traces of highly toxic substances such as dioxin requires sensitive laboratory procedures and equipment and meticulous attention to detail, which are again expensive. Research scientists might wonder whether they need large samples to confirm or negate their hunches. Engineers wonder at how many sites they should measure the soil, so as to predict conditions at places between their sampling points. Public bodies and construction companies do not want to spend more than is necessary to obtain the information on which to make decisions. How much sampling should they do to achieve their objectives? Statistics can provide them with answers based on rational criteria.
Not only is replication essential in a variable environment, but so too is sound design. Suppose, for example, that in trying to estimate the quality of groundwater in a region we were to take samples from village boreholes and count bacteria in them. Sampling would be cheap because the boreholes already exist. However, water in those boreholes is likely to be more co...