|
|
Upon thoughtful consideration of the sample survey approach, several questions
may come to mind. This section answers several commonly asked questions
concerning survey design. These questions are addressed in fairly general
terms. As noted in the introduction, additional technical details are in a
series of methods manuals. (Back to Frequently Asked
Questions
)
Survey Design
What is a survey?
Governmental organizations have monitored the aquatic environment for many
years. This includes monitoring of estuaries, coastal waters, streams, rivers,
lakes, reservoirs, and wetlands. Most, if not all, of these monitoring efforts
have been designed to fulfill a specific purpose; e.g., is a municipal
treatment plant in compliance. More recently, monitoring programs have been
asked to address more regional questions. Examples of the types of question
that these monitoring programs are asked to address are:
-
What is the condition of the Nation's lakes?
-
What is the condition of the streams and rivers in Wyoming?
-
What is the condition of the estuaries in USEPA Regions 9 and 10?
This type of question requires answers that apply to all of the aquatic resource
of interest that occur in a geographic area. Monitoring for results, such as
under the Governmental Performance for Results Act (GPRA), is one motivation
for asking this type of question. Another is the 305(b) reporting requirements
contained in the Clean Water Act. Both of these motivating agents require a
"fair or representative" picture of the resource.
In general there are two approaches to obtaining information on the aquatic
environment. Historically, the most common approach was to collection
information at locations that were chosen based on a variety of judgmental
factors, e.g., thought to be representative, in areas of special interest,
logistical considerations such as access and ease of sampling, etc. The
second approach, historically employed more frequently in other social and
natural resource arenas, relies upon a statistical methodology to provide
quantitative information about the aquatic resource. The interpretation
of the sampling results in the judgmental situation, relies on best
professional judgment, or in some cases modeling approaches, to address the
questions of interest. The statistically based approach utilizes the
scientific methodology developed for surveys to provide quantitative answers
and uncertainty measures for the sampled resource. The section on
sampling survey versus census provides a fuller discussion of this aspect.
The questions are phrased as the general public might initially pose the issue
of concern and are a basis for initiating the design of a monitoring program.
As the monitoring program is developed, the questions will be found to be too
general to complete the design. For example, what is meant by condition?
Answering this question impacts the selection of what indicators should be
measured? What is a lake? A stream? An estuary? Although everyone assumes that
they know what a stream is, the design requires an explicit definition for a
stream. It may be that only streams that are perennial and wadeable are of
interest in the study. In some cases, the geographic region requires further
clarification. For example, what is meant by "the Nation"? Does this
include American Samoa, the Virgin Islands, Puerto Rico, Hawaii, Alaska? The
information that follows focuses on the survey design portion of the general
questions. We will only discuss what will be measured, i.e., the indicators, in
terms of a general discussion on how the field plot design for measuring an
indicator is related to the survey design. Go to Top
What is a target population?
Before monitoring program design process can begin, a clear, concise description
of the aquatic resource is needed. In statistical terminology this description
is called the target population. The target population refers to a different
concept than that associated with the community, population, individual,
genetic concept of biological systems. Both the "target population"
for which information is wanted and the "elements" that make up the
target population must be rigorously defined. The target population is the
collection of elements about which information is wanted (Cochran, 1987;
Särndal et al, 1992). A number of examples will give a better understanding of
the concept of a target population and its elements. Go
to Top
Lakes as a discrete target population
Assume that a study of lakes in Oregon is to be conducted and that questions
focus on determining the number of lakes that have a particular characteristic,
such as its trophic status. The target population might be all lakes in Oregon
that are greater than 1 hectare and are wholly within the boundaries of the
state. The elements of the target population are the individual lakes. Note
that any lake which is partly in an adjoining state would not be included in
the target population. To be rigorous it further requires that a definition be
included for what constitutes a lake. For example, does the definition of a
lake include man-made reservoirs? What if the lake is no deeper than 1 meter
and is over 50% covered by vegetation? Are such lakes of interest for the
study? The definition must be sufficiently rigorous and explicit that it can be
clearly determined if a body of water is part of the target population. Note
that in our example the target population is discrete, i.e., there is a finite
number of lakes (elements) that comprise the target population. It would be
possible to sample every lake, i.e. a census, however, that is generally
logistically and economically unfeasible, an alternative being a sample
survey. Each indicator of trophic status determined for a lake results in
a single value for the lake. Consequently, the summary information about the
target population focuses on the number or proportion of lakes that have a
particular trophic status. Go to Top
Lakes as a continuous target population.
Assume that a study of one of the Great Lakes, e.g., Lake Ontario, requires an
estimate of the percent of the lake area with an unacceptable concentration of
dissolved oxygen. In this case the target population is the entire surface area
of Lake Ontario and the elements of the target population are all points within
the lake. Conceptually, dissolved oxygen can be measured everywhere on the lake
(an infinite number of points). Practically it is impossible to do so;
consequently it is natural to sample only a limited number of locations. That
is, it is impossible to complete a measurement on every element of the target
population. A rigorous definition for the target population would likely
involve questions of exactly what the boundaries are for Lake Ontario and
whether the boundary definition involves a minimum water depth. These
considerations are important for a field crew who must visit a proposed sample
site and determine whether the sample site is included in the target
population. Go to Top
Streams as a continuous target population.
A study is proposed to answer the question: What proportion of streams and
rivers in Wyoming have a fish index of biotic integrity (IBI) greater than 50?
The target population consists of all streams and rivers within the state of
Wyoming. First, some initial questions must be addressed. Does this definition
include the portion of Wyoming in Yellowstone National Park? Is the target
population restricted to only perennial streams? Streams may be thought of as a
linear network, such as is generally used to represent streams on maps and in
geographic information systems (GIS). The elements of the target population are
all the points within the linear stream network. In this case the target
population consists of an infinite number of elements. This is similar to the
Lake Ontario example, except that the elements occur on a linear network rather
than over a two-dimensional area. Field samples would be collected at a sample
of locations (elements) from the stream network and a fish IBI determined at
each location. Go to Top
Estuaries as a continuous target population.
Suppose that a study of the estuaries in California is planned to determine the
concentration of contaminants in sediments. For example, the state wants an
answer to the question: What proportion of the estuarine area in California has
unacceptable concentrations of mercury in sediments? The target population is
all estuaries in California. For the purposes of the study an estuary is
defined as any water body that is tidally influenced, saline, and has less than
50% of its perimeter adjacent to the Ocean. As a result, an estuary is defined
at its lower boundary by its articulation with the Ocean or another estuary and
its upper boundary by the head of tide. This definition results in
approximately 75 different estuarine areas along the California coast. They
range in surface area from 1092 km2 for the main body of San
Francisco Bay to 0.09 km2 for Sweetwater River. The elements of the
target population are all locations of sediments within the bounds of the
estuaries. An infinite number of locations exist within each estuary and the
entire target population consists of the collections of all the locations
across all the sometimes disconnected estuaries along the California coast.
Go to Top
What are Subpopulations and Why are They Important?
The target population defines the main aquatic resource of interest. Usually
subpopulations of the target population are also of interest. For example, in a
study of all lakes in Oregon two potential subpopulations might be natural
lakes and man-made lakes (i.e., reservoirs). The study may also be designed to
compare all lakes between 1 to 50 hectares versus all lakes greater than 50
hectares. Often an estimate of the number of lakes in each trophic status
category is desired for each subpopulation. Note that the subpopulation of
natural lakes overlaps the subpopulation of lakes between 1 to 50 hectares.
Subpopulations do not need to be non-overlapping. They only need to be of
interest in the study and explicitly defined.
Why is the definition of subpopulations important in the planning of a survey?
Subpopulations arise from the questions that a study must answer. For example,
the need for answers on natural and man-made lakes arises from questions posed
at the initiation of the study indicate that the trophic status may differ
between man-made and natural lakes. If such a difference exists, then different
management strategies may be taken for the two types of lakes. During the
initial planning of a study, it is typical for many subpopulations of interest
to be identified. In each case a strong rationale can be given as to why the
information on the subpopulation is important. A study can only meet the
expectations of those requiring the information if clarity is reached on what
subpopulations estimates will be provided.
Subpopulations are also sometimes referred to as domains of study or reporting
units. These phrases imply that the subpopulations are known to be of interest
prior to the conduct of the study. They are sufficiently important that the
study would be viewed as incomplete if estimates for them did not appear in a
report of the study's findings. During the statistical analysis of the study
other subpopulations may be identified and reported on, but they would be
viewed as providing additional information rather than essential to the study.
Examples of subpopulations determined prior to sampling: geographic areas
(States, counties), close to urban areas, resource definitions (wadeable,
non-wadeable), ecoregions, etc.
Others can be determined following sampling, usually the information is only
available at the time of sampling: salinity, substrate types, landuse category,
small scale habitat features. In general these are not available during
the design phase, information that would allow allocation of samples in each
subpopulation to assure sufficient sample sizes.
How does identifying subpopulations impact the design of the survey? The
expectation is that the survey will provide estimates with acceptable precision
for each subpopulation. Achieving acceptable precision requires having a
sufficient number of sample sites occur within a subpopulation. Assume that 50
samples are needed to meet the precision requirements. If only the target
population is of interest, then a total sample size of 50 is all that is
required. However, if three non-overlapping subpopulations are identified, then
150 total samples would be needed: 50 in each of the three subpopulations. Note
that in this case the precision for the target population will be better than
required since it has at least 150 samples. Many times subpopulations overlap
so that multiplying the number of subpopulations by the required sample size
results in many more samples than is actually needed. For example, splitting
lakes into man-made and natural and into 1 to 50 hectare and greater than 50
hectare may only require a total of 100 samples as long as the survey design
results in 50 samples in each of the four subpopulations. The major impact on
the design of the survey is the increased sample size requirements and the need
for the survey design to make sure that the each subpopulation receives the
minimum required number of samples to meet precision requirements. How survey
designs can accomplish the allocation of samples to subpopulations is discussed
under Common Survey Designs. Go to Top
When would I use stratification versus subpopulations?
Some of the usual reasons to stratify include: 1) administrative or operational
convenience, 2) particular portions of the target population require different
survey designs, and 3) increase precision by constructing strata that are
homogeneous. Designs for such strata tend to create more independent
among strata and often contain subpopulations within them. Impacts on the
sample size needed to achieve acceptable precision are the same under both
approaches, i.e. increases in the number of desired estimates also increases
the number of samples needed. Creation of subpopulations is usually
undertaken to support unequal weighting, which provides a method for allocating
sample to the subpopulations. It also can improve the precision of the
resulting estimates. Creating subpopulations requires auxiliary
information for each member of the target population during the design
process. For additional information on subpopulations, see the previous
answer addressing subpopulations.
Go to Top
What is a sample frame?
Särndal et al (1992) define the frame or sampling frame as any material or
device used to obtain observational access to the target population.
Continuing, they state that the frame must make it possible " to (1)
identify and select a sample in a way that respects a given probability
sampling design and (2) establish contact with selected elements." The
definition is abstract, suggesting that many options are possible to construct
a sample frame. That is the case. Generally, the more information available for
use in electronic form, the easier it is to develop a survey design that meets
requirements of a study.
What does the definition imply when studying aquatic resources? The answer
depends upon what information in a usable form currently exists about the
location of the target population and all of its elements. For example, in the
study of all lakes in Oregon, a list of all the lakes and their location may
exist in a computer database. This list could then be used as the sample frame.
Sample sites would be selected from the list and since their location is known
they could be visited to obtain the desired measurements. This lake sample
frame is very simple and easy to use. However, the list only has information on
the lake name and its location. If the survey design required additional
information, such lake size, to complete the sample selection, then the frame
would be inadequate and an alternative sample frame for lakes would be
needed. Currently, for streams and rivers in the US, the RF3 GIS computer
files of stream reaches is available as the frame for designs for these
resources. The primary requirements for the frame is that it cover the
entire target population and be reasonably accurate. Go
to Top
Sample survey versus census
Most people are aware of the term "census" from its use in relation to
the decennial counting of the population of the United States as required by
the constitution. Operationally the objective is to contact every individual in
the United States to elicit basic demographic facts about them: sex, age, race,
etc. The U.S. Bureau of Census devotes appreciable resources every ten years to
carry out the census constitutional mandate. Can a census be completed for
aquatic resources? The practical answer is rarely. A census of lakes in Oregon
would involve visiting every lake in Oregon and obtaining on each lake the
measurements specified for the study. The measurements from all lakes would be
used to determine summary characteristics about all lakes. Although completing
a census of lakes may be possible, available funds, logistics and personnel
will likely make it impractical. A census of streams and rivers in Wyoming is
not only impractical, but essentially impossible. When streams are viewed as a
continuous resource in a linear network, a infinite number of elements exist in
the target population making it impossible to visit each element. The same is
true for estuaries viewed as a continuous population.
A sample survey is a way of collecting information on a subset of the elements
of the target population with the intention of using the information to
determine summary characteristics about the population. The summaries differ
from the same summaries determined from a census in that they contain
uncertainty. The uncertainty arises from the simple fact that not all elements
in the target population were visited. How the sample is selected determines
whether it is possible to know the uncertainty of the estimate.
A sample is any subset of the target population, i.e., any collection of its
elements. Sampling methods may be classified into either probability-based
sample methods or non-probability-based sampling methods. Probability-based
methods are discussed in subsequent sections. Non-probability methods include
chunk samples, expert choice samples, and quota samples.
Go to Top
Chunk Samples. Scientists often draw conclusions using an arbitrary or
fortuitous collection of sites. The sites are gathered haphazardly or
"happen to be handy." Often the scientist implicitly assumes that the
sites are typical for a larger universe of sites about which conclusions are
desired. Such an assumption has only the individual's judgment as a basis and
can not be easily defended. The sites are an unknown "chunks" of the
target population and consequently no basis exists to make a scientific
inference to the target population without invoking assumptions that can not be
verified. Go to Top
Expert Choice Samples. Expert choice sampling is a form of judgment
sampling that is a more developed form of non-random selection. An expert, or
experts, may define a set of criteria to be met for a site to be included in
the sample. Not all sites that meet the criteria are included. Criteria usually
result in the designation of "typical" sites for the study. A fairly
good sample may result given that the expert was skillful in defining the
criteria and locating typical sites that met the criteria. However there is no
way to be sure. A different expert would probably use different criteria or
pick different sites that et the criteria. Without invoking additional
assumptions, no basis exists to make inferences to the target population and
know the uncertainty associated with the inference. Go
to Top
Quota Samples. Quota sampling is commonly used in market research. The
target population is divided according to one or more characteristics, e.g.,
age, sex, and geographic area. For two age groups, two sexes, and three
geographic areas, a total of 12 population cells are defined. The cells are
similar to strata in stratified random sampling. A quota sample then contains a
pre-determined number of individuals in each of the 12 cells. The interviewer
then simply "fills the quota" for each cell. The individuals may be
the first individuals encountered or the interviewer may have the option of
using judgment in selecting the individuals. The sample of individuals in a
cell is either a chunk sample or a judgment sample. Individuals may refuse or
be unavailable, but new individuals are contacted until the quota of
individuals is achieved. However, the problem of selection bias due to
non-response still remains. Hence as before there is no basis for an inference
to the target population. Go to Top
What is a probability sample?
A probability sample is a sample where every element of the target population
has a known, non-zero probability of being selected. That is, it is possible
for every element of the target population to be in the sample. Two important
features of a probability sample are that the probability selection mechanism
(1) guards against site selection bias and (2) is the basis for scientific
inference to characteristics of the entire target population.
Many alternative approaches are available to select a probability sample. Which
approach is used depends on the objectives of the survey, the available
auxiliary information, logistical or operational constraints in conducting the
sampling, the characteristics of the sample frame, and the complexity of the
statistical analysis. A few alternative designs are described below.
Go to Top
Simple random sample. A simple example of a probability sample is one
that gives every sample of a fixed size the same probability of being selected.
This is simple random sampling without replacement. This is the simplest type
of probability sample. Its major advantage is its simplicity not only in design
but in statistical analysis of the survey results. Statistical analyses do not
require any special procedures; consequently, users can analyze the data
essentially ignoring the probability design. In reality the probability design
results in the same assumptions that are used in standard statistical analyses.
Its major disadvantage is that the design does not incorporate any information
about the target population which would improve the efficiency (precision) of
the survey and does not necessarily provide a sample that will address all the
survey objectives. Go to Top
Stratified random sample. A stratified random sample may be the most
common probability survey design used. When auxiliary information is available
on the target population or the survey has multiple objectives, that
information can be used to define strata. For example, a survey conducted over
multiple states may have an objective to provide estimates for each state as
well as for all states combined. In addition, it may be operationally
convenient to have the sample for each state be selected independently from
other states. This can be achieved by defining each state as a stratum and then
selecting a simple random sample within each state. Strata may also be defined
using a known characteristic of each element in the target population, i.e.,
auxiliary data. For example, streams may be categorized according to Strahler
order in a GIS coverage that serves as the sample frame. Survey objectives may
require that approximately an equal number of samples come from 1st, 2nd, 3rd,
and 4th and higher Strahler order categories. This may be achieved by defining
four strata based on the Strahler order of each stream. A stratified sample can
be considered as a set of independent simple random samples, i.e., each stratum
has a simple random sample. Consequently, the statistical analyses are similar
to those of a simple random sample. Each stratum is analyzed as a simple random
sample; then the estimates are combined across the stratum. The latter step
must be completed correctly to avoid biased estimates. Go
to Top
Unequal probability sample. An alternative to a stratified random sample
is an unequal probability sample. An unequal probability sample is achieved by
assigning a probability of selection to each element of the target population,
usually depending on auxiliary information. For example, Strahler order could
be used to assign a probability of selection to each stream segment where 2nd
order streams would be twice as likely to be selected as 1st order streams, 3rd
order streams four times as likely as 1st order, and 4th and higher order eight
times as likely as 1st order. This type of design provides enormous flexibility
in designing to meet objectives as well as a mechanism to increase precision.
Their statistical analysis is more complex and requires that all analyses use
weights derived from the unequal probability of selection.
Go to Top
How many sample sites to use?
The most commonly asked question is: How many sample sites do I need? This is an
important question as it directly determines the precision of any statement
derived from the sample data. An answer requires detailed information on all
the estimates that will be produced from the survey, the precision desired for
each estimate, and knowledge of the variability expected. A reality faced in
most studies is that the number of objectives creates a need for many more
sample sites than budget and operational constraints allow. Consequently, the
total number of sites in many situations is known from these constraints and
the question is which objectives are the most important. It is usual to have
some sub-objectives dropped due to sample size limitations.
Sample size calculations are available in most survey sampling textbook and will
not be discussed here. One specific situation of interest is when the
objectives call for the estimation of a proportion, e.g., proportion of stream
length that meets a designated use. In this case, sample size calculations
depend only on the proportion, precision required, and confidence required.
Approximate precision estimates for proportions can be obtained by assuming the
survey designs are simple random samples. Under this condition the estimated
precision can be estimated using procedures given by Cochran (1987) for
proportions. Go to Top
Precision, as a percent, is determined from precision = Z1-"
* 100 * Sqrt[ p(1-p)/n]
To calculate precision requires knowledge of p, the proportion to be estimated.
However, a conservative estimate of precision can be obtained by assuming p to
be 0.5, which gives the maximum variance. Z1-"
is related to the level of confidence required for the estimate. If desire 90%
confidence, then use 1.645. If desire 95% confidence, then use 2.
Table 1. Precision to achieve 90% confidence in estimates of selected
proportions.
Assumed
Proportion
(percent)
|
Precision with 90% Confidence
for alternative sample sizes
|
Precision with 95% Confidence
for alternative sample sizes
|
|
25
|
50
|
100
|
400
|
1000
|
25
|
50
|
100
|
400
|
1000
|
20%
|
±13
|
±9
|
±7
|
±3
|
±2
|
±16
|
±11
|
±8
|
±4
|
±3
|
50%
|
±17
|
±12
|
±8
|
±4
|
±3
|
±20
|
±13
|
±10
|
±5
|
±3
|
If the survey designs are actually based on the spatially-restricted survey
designs, the actual precision estimates are expected to be lower (better) than
those stated. Go to Top
How do Sites get selected?
Several processes lead to the selection of sites. The first process
identifies the resource characteristics and target population and results in a
sample frame that contains all sites within the target population. The
second process establishes a spatial grid and hierarchical structure that
result in cells containing single, or a small number/area of sites. These
two results are then combined resulting in each site, or small number/area of
sites assigned a hierarchical cell address. Randomization and statistical
weighting processes produce a sequence of all sites from which a systematic
random sample is selected. For additional details see information on
Survey design, Discrete Grid, and
step-by-step site selection example.
Go to Top
Why a sample size of 50?
A general objective of stream or lake surveys is to estimate the proportion of
the population (of sample units) that meet a specific index of condition. For
example, what proportion of the wadeable stream miles in a state don't meet
their aquatic life use designations; or how many lakes in the state are
eutrophic. Given that a statistical probability survey design is used to select
a sample of lake or stream units from the entire population of several thousand
such units, the estimate will have uncertainty associated with it. One measure
of the uncertainty is its precision. The estimate is a proportion, p (i.e., the
proportion of stream length that is impaired; or the proportion lakes that are
eutrophic). Under a general assumption that the survey design is a simple
random sample, we know the variance for the proportion depends only on the true
proportion of units meeting the criterion and the sample size, n. Since the
true proportion is unknown, we can make a conservation estimate of the variance
by assuming the true proportion is 0.5, where the variance is a maximum. Under
this scenario, with a sample size of 50, the precision when 50% of units meet
the criteria (i.e., 50% of stream length is impaired) will be +/-12% with 90%
confidence. If only 20% of the units meet the criterion then the precision will
be +/-9%. If only 25 units are sampled, then the precision changes to +/-17%
and +/-13% respectively. If 250 units are sampled over a 5-year period, then
the precision changes to +/-5% and +/-4%. It is necessary to assume that
conditions during the 5-year period remain constant. Note that for 250 units,
the precision is +/- 6% and +/-5% with 95% confidence.
A critical element in any discussion of precision is the number
"subpopulations" for which estimates are needed. If estimates are
required for each of say, 5 ecoregions, then with 25 units sampled in each of
the ecoregions the precision would be +/-17% and +/-13% (at p=.5 and p=.3,
respectively) for each ecoregion and +/-5% and +/-4% for the entire study area.
The sub-regions do not need to be ecoregions, but could be watersheds or some
other subdivision of the landscape of interest. Another possible grouping of
the units might be by percent of public or private ownership Hence a critical
element on determining total sample size will be determining how many and what
type of groups (subpopulations) will be of management interest. If the 25
samples are obtained over a 5-year period (that is, sample 5 units each year in
each group), then the precision after five years is as stated. For any one
year, the precision is +/- 37% when p=0.5.
Precision will be important when a determination is to be made on whether the
proportion meeting the criteria differs between two different years, say 5
years apart. If the true proportion changes by 10%, e.g., from 20% to 30%, then
what is the chance that the monitoring will detect this change? The better the
precision the more likely the change will be detected. The probability of
detecting the change depends not only on the sample size but also the specific
survey design to be implemented. Decreasing the sample size decreases the
ability to detect the difference. Is 50 a sufficient sample size? That depends
on how confident we must be in detecting a change or estimating the true
proportion. As an example, if the baseline proportion is 50% and after 5 years
the proportion changes to 30%, then with a sample size of 50 the estimated
difference would be 20% +/- 16% with 90% confidence. Since the confidence
interval does not include 0, the conclusion is that a significant (at 90%
confidence) change has occurred. This assumed that the sample units between the
two periods were not paired, i.e., the same in both time periods. If they are
the same, then other procedures can be used and are expected to be able to
detect smaller differences between the time periods. Go
to Top
Response designs - plot level design
considerations
A monitoring design may be divided into two design parts: survey design and
response design. The survey design selects which sites to visit during any
particular sampling period, while the response design determines what and how
to collect information at the selected sites. All the prior discussion
concerned the survey design. Multiple response designs are likely to be used at
each site. Typically, each field measurement requires its own specific field
plot design, measurement protocol, laboratory protocol (if required), and
calculation procedure to obtain a metric or indicator value of interest.
Response designs may designate that only a single site visit each year will be
made during an index time period. This is typical for monitoring designs
conducted over large regional areas such as States. Nothing prevents the
response design from requiring monthly site visits during a year or even
collecting continuous data throughout the year. Choice of a response design is
driven by the objectives of the study. An example of a detailed protocol
for a response design for wadeable streams is available on the web site:
www.epa.gov/wed/pages/EMAPManual.pdf A
diagram of that response design is included in this web site.Go
to Top