Series editor: Kenneth N.Ross
Module
Kenneth N. Ross
Sample design for educational survey research
3
UNESCO International Institute for Educational Planning
Module 3
Sample design for educational survey research
Content
1. Basic concepts of sample design for educational survey research
Populations : desired, defined, and excluded Sampling frames Representativeness Probability samples and nonprobability samples Types of nonprobability samples
1. Judgement sampling 2. Convenience sampling 3. Quota sampling
1
2 3 4 5 6
7 7 8
Types of probability samples
1. Simple random sampling 2. Stratified sampling 3. Cluster sampling 1. Mean square error
8
9 10 11 13 15 16
The accuracy of estimates obtained from probability samples 13
2. The accuracy of individual sample estimates 3. Comparison of the accuracy of probability samples
Sample design for twostage cluster samples
1. The design effect for twostage cluster samples 2. The effective sample size for twostage cluster samples
17
18 19
Sample design tables for twostage cluster samples
20
© UNESCO
1
Module 3
Sample design for educational survey research
2. How to draw a national sample of students: a hypothetical example for ‘country x’ 3. How to draw a national sample of schools and students: a real example for Zimbabwe 4. The estimation of sampling errors
The Jackknife Procedure
27 48 64
66
5. References Appendix 1
Random number tables for selecting a simple random sample of twenty students from groups of students of size 21 to 100
70
73 78 81
Appendix 2 Appendix 3
Sample design tables (for roh values of 0.1 to 0.9)
Estimation of the coefficient of intraclass correlation
II
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
Sampling in educational research is generally conducted in order to permit the detailed study of part, rather than the whole, of a population. The information derived from the resulting sample is customarily employed to develop useful generalizations about the population. These generalizations may be in the form of estimates of one or more characteristics associated with the population, or they may be concerned with estimates of the strength of relationships between characteristics within the population. Provided that scientific sampling procedures are used, the selection of a sample often provides many advantages compared with a complete coverage of the population. For example, reduced costs associated with gathering and analyzing the data, reduced requirements for trained personnel to conduct the fieldwork, improved speed in most aspects of data summarization and reporting, and greater accuracy due to the possibility of more intense supervision of fieldwork and data preparation operations. The social science research situations in which sampling is used may be divided into the following three broad categories: experiments – in which the introduction of treatment variables occurs according to a prearranged experimental design and all extraneous variables are either controlled or randomized; surveys – in which all members of a defined target population have a known
1
© UNESCO
1
Module 3
Sample design for educational survey research
nonzero probability of selection into the sample; and investigations – in which data are collected without either the randomization of experiments or the probability sampling of surveys. Experiments are strong with respect to internal validity because they are concerned with the question of whether a true measure of the effect of a treatment variable has been obtained for the subjects in the experiment. Surveys, on the other hand, are strong with respect to external validity because they are concerned with the question of whether the findings obtained for the subjects in the survey may be generalized to a wider population. Investigations are weak with respect to both internal and external validity and their use is due mainly to convenience or low cost.
Populations : desired, defined, and excluded
In any educational research study it is important to have a precise description of the population of elements (persons, organizations, objects, etc.) that is to form the focus of the study. In most studies this population will be a finite one that consists of elements which conform to some designated set of specifications. These specifications provide clear guidance as to which elements are to be included in the population and which are to be excluded. In order to prepare a suitable description of a population it is essential to distinguish between the population for which the results are ideally required, the desired target population, and the population which is actually studied, the defined target population. An ideal situation, in which the researcher had complete control over the research environment, would lead to both of these populations containing the same elements. However, in most studies, some differences arise due, for example, to (a) noncoverage:
2
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
the population description may accidently omit some elements because the researcher has no knowledge of their existence, (b) lack of resources: the researcher may intentionally exclude some elements from the population description because the costs of their inclusion in data gathering operations would be prohibitive, or (c) an ageing population description: the population description may have been prepared at an earlier date and therefore it includes some elements which have ceased to exist. The defined target population provides an operational definition which may be used to guide the construction of a list of population elements, or sampling frame, from which the sample may be drawn. The elements that are excluded from the desired target population in order to form the defined target population are referred to as the excluded population.
Sampling frames
The selection of a sample from a defined target population requires the construction of a sampling frame. The sampling frame is commonly prepared in the form of a physical list of population elements – although it may also consist of rather unusual listings, such as directories or maps, which display less obvious linkages between individual list entries and population elements. A wellconstructed sampling frame allows the researcher to ‘take hold’ of the defined target population without the need to worry about contamination of the listing with incorrect entries or entries which represent elements associated with the excluded population. Generally the sampling frame incorporates a great deal more structure than one would expect to find in a simple list of elements. For example, in a series of largescale studies of Reading Literacy carried out in 30 countries during 1991 (Ross, 1991), sampling frames were constructed which listed schools according to a number
© UNESCO
3
Module 3
Sample design for educational survey research
of stratification variables: size (number of students), program (for example, comprehensive or selective), region (for example, urban or rural), and sex composition (single sex or coeducational). The use of these stratification variables in the construction of sampling frames was due, in part, to the need to present research results for sample data that had been drawn from particular strata within the sampling frame.
Representativeness
The notion of ‘representativeness’ is a frequently used, and often misunderstood, notion in social science research. A sample is often described as being representative if certain percentage frequency distributions of element characteristics within the sample data are similar to corresponding distributions within the whole population. The population characteristics selected for these comparisons are referred to as ‘marker variables’. These variables are usually selected from among those demographic variables that are readily available for both population and sample. Unfortunately, there are no objective rules for deciding which variables should be nominated as marker variables. Further, there are no agreed benchmarks for assessing the degree of similarity required between percentage frequency distributions for a sample to be judged as ‘representative of the population’. It is important to note that a high degree of representativeness in a set of sample data refers specifically to the marker variables selected for analysis. It does not refer to other variables assessed by the sample data and therefore does not necessarily guarantee that the sample data will provide accurate estimates for all element characteristics. The assessment of the accuracy of sample data can only be discussed meaningfully with reference to the value of the mean square error, calculated separately, for particular sample estimates (Ross, 1978).
4
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
The most popular marker variables in the field of education have commonly been demographic factors associated with students (sex, age, socioeconomic status, etc.) and schools (type of school, school location, school size, etc.). For example, in a series of educational research studies carried out in the United States during the early 1970’s, Wolf (1977) selected the following marker variables: sex of student, father’s occupation, father’s education, and mother’s education. These variables were selected because their percentage frequency distributions could be obtained for the population from tabulations prepared by the Bureau of the Census.
Probability samples and nonprobability samples
The use of samples in educational research is usually followed by the calculation of sample estimates with the aim of either (a) estimating the values of population parameters from sample statistics, or (b) testing statistical hypotheses about population parameters. These two aims require that the researcher has some knowledge of the accuracy of the values of sample statistics as estimates of the relevant population parameters. The accuracy of these estimates may generally be derived from statistical theory – provided that probability sampling has been employed. Probability sampling requires that each member of the defined target population has a known, and nonzero, chance of being selected into the sample. In contrast, the stability of sample estimates based on nonprobability sampling cannot be discovered from the internal evidence of a single sample. That is, it is not possible to determine whether a nonprobability sample is likely to provide very accurate or very inaccurate estimates of population parameters. Consequently, these types of samples are not appropriate for
© UNESCO
5
Module 3
Sample design for educational survey research
dealing objectively with issues concerning either the estimation of population parameters or the testing of hypotheses. The use of nonprobability samples is sometimes carried out with the (usually implied) justification that estimates derived from the sample may be linked to some hypothetical universe of elements rather than to a real population. This justification may lead to research results which are not meaningful if the gap between the hypothetical universe and any relevant real population is too large. In some circumstances, a wellplanned probability sample design can be turned accidentally into a nonprobability sample design if subjective judgement is exercised at any stage during the execution of the sample design. Some researchers fall into this trap through a lack of control of field operations at the final stage of a multistage sample design. The most common example of this in educational settings occurs when the researcher goes to great lengths in drawing a probability sample of schools, and then leaves it to the initiative of teaching staff in the sampled schools to select a ‘random sample’ of students or classes.
Types of nonprobability samples
There are three main types of nonprobability samples: judgement, convenience, and quota samples. These approaches to sampling result in the elements in the target population having an unknown chance of being selected into the sample. It is always wise to treat research results arising from these types of sample design as suggesting statistical characteristics about the population – rather than as providing population estimates with specifiable confidence limits.
6
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
1. Judgement sampling
The process of judgement, or purposive, sampling is based on the assumption that the researcher is able to select elements which represent a ‘typical sample’ from the appropriate target population. The quality of samples selected by using this approach depends on the accuracy of subjective interpretations of what constitutes a typical sample. It is extremely difficult to obtain meaningful results from a judgement sample because no two experts will agree upon the exact composition of a typical sample. Therefore, in the absence of an external criterion, there is no way in which in the research results obtained from one judgement sample can be judged as being more accurate than the research results obtained from another.
2. Convenience sampling
A sample of convenience is the terminology used to describe a sample in which elements have been selected from the target population on the basis of their accessibility or convenience to the researcher. Convenience samples are sometimes referred to as ‘accidental samples’ for the reason that elements may be drawn into the sample simply because they just happen to be situated, spatially or administratively, near to where the researcher is conducting the data collection. The main assumption associated with convenience sampling is that the members of the target population are homogeneous. That is, that there would be no difference in the research results obtained from a random sample, a nearby sample, a cooperative sample, or a sample gathered in some inaccessible part of the population.
© UNESCO
7
Module 3
Sample design for educational survey research
As for judgement sampling, there is no way in which the researcher may check the precision of one sample of convenience against another. Indeed the critics of this approach argue that, for many research situations, readily accessible elements within the target population will differ significantly from less accessible elements. They therefore conclude that the use of convenience sampling is likely to introduce a substantial degree of bias into sample estimates of population parameters.
3. Quota sampling
Quota sampling is a frequently used type of nonprobability sampling. It is sometimes misleadingly referred to as ‘representative sampling’ because numbers of elements are drawn from various target population strata in proportion to the size of these strata. While quota sampling places fairly tight restrictions on the number of sample elements per stratum, there is often little or no control exercised over the procedures used to select elements within these strata. For example, either judgement or convenience sampling may be used in any or all of the strata. Therefore, the superficial appearance of accuracy associated with proportionate representation of strata should be considered in the light that there is no way of checking either the accuracy of estimates obtained for any one stratum, or the accuracy of estimates obtained by combining individual stratum estimates.
Types of probability samples
There are many ways in which a probability sample may be drawn from a population. The method that is most commonly described in textbooks is simple random sampling. This method is rarely used in practical social research situations because
8
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
(a) the selection and measurement of individual population elements is often too expensive, and (b) certain complexities may be introduced intentionally into the sample design in order to address more appropriately the objectives and administrative constraints associated with the research. The complexities most often employed in educational research include the use of stratification techniques, cluster sampling, and multiple stages of selection.
1. Simple random sampling
The selection of a simple random sample is usually carried out according to a set of mechanical instructions which guarantees the random nature of the selection procedure. For example, Kish (1965) provides the following operational definition in order to describe procedures for the selection of a simple random sample of elements without replacement from a finite population of elements: From a table of random digits select with equal probability n different selection numbers, corresponding to n of the N listing numbers of the population elements. The n listings selected from the list, on which each of the N population elements is represented separately by exactly one listing, must identify uniquely n different elements. (p. 36) Simple random sampling, as described in this definition, results in an equal probability of selection for all elements in the population. This characteristic, called ‘epsem sampling’ (equal probability of selection method), is not restricted solely to this type of sample design. Equal probability of selection can result from either the use of equal probabilities of selection throughout the sampling process, or from the use varying probabilities that compensate for each other through several stages of multistage sampling. Epsem sampling is widely applied in educational research because it usually leads to selfweighting samples in which the simple arithmetic mean obtained from the sample data is an unbiased estimate of the population mean.
© UNESCO
9
Module 3
Sample design for educational survey research
2. Stratified sampling
The technique of stratification is often employed in the preparation of sample designs because it generally provides increased accuracy in sample estimates without leading to substantial increases in costs. Stratification does not imply any departure from probability sampling – it simply requires that the population be divided into subpopulations called strata and that probability sampling be conducted independently within each stratum. The sample estimates of population parameters are then obtained by combining information from each stratum. In some studies, stratification is used for reasons other than obtaining gains in sampling accuracy. For example, strata may be formed in order to employ different sample designs within strata, or because the subpopulations defined by the strata are designated as separate ‘domains of study’ (Kish, 1987, p. 34). Variables used to stratify populations in education generally describe demographic aspects concerning schools (for example, location, size, and program) and students (for example, age, sex, grade level, and socioeconomic status). Stratified sampling may result in either proportionate or disproportionate sample designs. In a proportionate stratified sample design the number of observations in the total sample is allocated among the strata of the population in proportion to the relative number of elements in each stratum of the population. That is, a stratum containing a given percentage of the elements in the population would be represented by the same percentage of the total number of sample elements. In situations where the elements are selected with equal probability within strata, this type of sample design results in epsem sampling and therefore ‘selfweighted’ estimates of population parameters.
10
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
In contrast, a disproportionate stratified sample design is associated with the use of different probabilities of selection, or sampling fractions, within the various population strata. This can sometimes occur when the sample is designed to achieve greater overall accuracy than proportionate stratification by using ‘optimum allocation’ (Kish,1965:92). More commonly disproportionate sampling is used in order to ensure that the accuracy of sample estimates obtained for stratum parameters is sufficiently high to be able to make meaningful comparisons between strata. The sample estimates derived from a disproportionate sample design are generally prepared with the assistance of ‘weighting factors’. These factors, represented either by the inverse of the selection probabilities or by a set of numbers proportional to them, are employed in order to prevent inequalities in selection probabilities from causing the introduction of bias into sample estimates of population parameters. The reciprocals of the selection probalities, sometimes called ‘raising factors’, refer to the number of elements in the population represented by a sample element (Ross, 1978). In the field of educational research, the weighting factors are often calculated so as to ensure that the sum of the weighting factors over all elements in the sample is equal to the sample size. This ensures that the readers of research reports are not confused by differences between actual and weighted sample sizes.
3. Cluster sampling
A population of elements can usually be thought of as a hierarchy of different sized groups or ‘clusters’ of sampling elements. These groups may vary in size and nature. For example, a population of school students may be grouped into a number of classrooms, or it may be grouped into a number of schools. A sample of students may then be selected from this population by selecting clusters of students
© UNESCO
11
Module 3
Sample design for educational survey research
as classroom groups or school groups rather than individually as would occur when using a simple random sample design. The use of cluster sampling in educational research is sometimes undertaken as an alternative to simple random sampling in order to reduce research costs for a given sample size. For example, a cluster sample consisting of the selection of 10 classes – each containing around 20 students – would generally lead to smaller data collection costs compared with a simple random sample of 200 students. The reduced costs occur because the simple random sample may require the researcher to collect data in as many as 200 schools. Cluster sampling does not prevent the application of probability sampling techniques. This may be demonstrated by examining several ways in which cluster samples may be drawn from a population of students. Consider the hypothetical population, described in Figure 1, of twentyfour students distributed among six classrooms (with four students per class) and three schools (with two classes per school). A simple random sample of four students drawn without replacement from this population would result in an epsem sample with each element having a probability of selection, p, equal to 1/6 (Kish. 1965:40). A range of cluster samples, listed below with their associated p values, may also be drawn in a manner which results in epsem samples. • • • Randomly select one class, then include all students in this class in the sample. (p = 1/6 x 4/4 = 1/6). Randomly select two classes, then select a random sample of two students from within these classes. (p = 2/6 x 2/4 = 1/6). Randomly select two schools, then select a random sample of one class from within each of these schools, then select a random sample of two students from within these classes. (p = 2/3 x 1/2 x 2/4 = 1/6).
12
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
Figure 1. Schools
Hypothetical Population of 24 students
School 1 / \ School 2 / \ School 3 / \
Classes
Class 1 Class 2 Class 3 Class 4 Class 5 Class 6      
Students
abcd
efgh
ijkl
mnop
qrst
uvwx
The accuracy of estimates obtained from probability samples
The degree of accuracy associated with a sample estimate derived from any one probability sample may be judged by the difference between the estimate and the value of the population parameter which is being estimated. In most situations the value of the population parameter is not known and therefore the actual accuracy of an individual sample estimate cannot be calculated in absolute terms. Instead, through a knowledge of the behaviour of estimates derived from all possible samples which can be drawn from the population by using the same sample design, it is possible to estimate the probable accuracy of the obtained sample estimate.
1. Mean square error
Consider a probability sample of n elements which is used to calculate the sample mean as an estimate of the population mean. If an infinite set of samples of size n were drawn independently from
© UNESCO
13
Module 3
Sample design for educational survey research
this population and the sample mean calculated for each sample, the average of the resulting sampling distribution of sample means would be referred to as the expected value. The accuracy of the sample mean as an estimator of the population parameter may be summarized in terms of the mean square error (MSE). The MSE is defined as the average of the squares of the deviations of all possible sample estimates from the value being estimated (Hansen et al, 1953).
A sample design is unbiased if the expected value of the sample mean is equal to the population mean. It is important to remember that ‘bias’ is not a property of a single sample, but of the entire sampling distribution, and that it belongs neither to the selection nor the estimation procedure alone, but to both jointly. For most welldesigned samples in educational research the sampling bias is usually very small – tending towards zero with increasing sample size. The accuracy of sample estimates is therefore generally assessed in terms of the magnitude of the variance term in the above equation.
14
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
2. The accuracy of individual sample estimates
In educational settings the researcher is usually dealing with a single sample of data and not with all possible samples from a population. The variance of sample estimates as a measure of sampling accuracy cannot therefore be calculated exactly. Fortunately, for many probability sample designs, statistical theory may be used to derive formulae which provide estimates of the variance based on the internal evidence of a single sample of data. For a simple random sample of n elements drawn without replacement from a population of N elements, the variance of the sample mean may be estimated from a single sample of data by using the following formula: where s2 is the usual sample estimate of the variance of the element values in the population (Kish, 1965:41).
���� � �� ���� � � � � � �
For sufficiently large values of N, the value of the finite population correction, (N  n)/N, tends toward unity. The variance of the sample mean in this situation may be estimated to be equal to s2/n. The sampling distribution of the sample mean is approximately normally distributed for many educational sampling situations. The approximation improves with increased sample size even though the distribution of elements in the parent population may be far from normal. This characteristic of sampling distributions is known as the Central Limit Theorem and it occurs not only for the sample mean but also for most estimators commonly used to describe survey research results (Kish, 1965).
© UNESCO
15
Module 3
Sample design for educational survey research
From a knowledge of the properties of the normal distribution we know that we can be “68 per cent confident” that the population mean lies within the range specified by:
(the sample mean ± one standard error of the sample mean) where the standard error of the sample mean is equal to the square root of the variance of the sample mean. Similarly the we can be “95 percent confident” that the population mean lies within the range specified by:
(sample mean ± two standard errors of the sample mean)
While the above discussion has concentrated mostly on sample means derived from simple random samples, the same approach may be used to set up confidence limits for many other population values derived from various types of sample designs. For example, confidence limits may be calculated for complex statistics such as correlation coefficients, regression coefficients, multiple correlation coefficients, etc. (Ross, 1978).
3. Comparison of the accuracy of probability samples
The accuracy of probability samples is usually compared by considering the variances associated with a particular sample estimate for a given sample size. This comparison has, in recent years, been based on the recommendation put forward by Kish (1965) that the simple random sample design should be used as a standard for quantifying the accuracy of a variety of probability sample designs which incorporate such complexities as stratification and clustering. Kish (1965:162) introduced the term ‘deff’ (design effect) to describe the ratio of the variance of the sample mean for a complex sample design, denoted c, to the variance a simple random sample, denoted srs, of the same size.
16
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
����
�
���� � � � ���� � ��� �
The potential for arriving at false conclusions in educational research by using incorrect sampling error calculations has been demonstrated in a study carried out by Ross (1976). This study showed that it was highly misleading to assume that sample size was, in itself, an adequate indicator of the sampling accuracy associated with complex sample designs. For example, Ross (1976: 40) demonstrated that a twostage cluster sample of 150 students (that was selected for a study conducted in Australia by randomly selecting 6 classes followed by the random selection of 25 students within these classes) had the same sampling accuracy for sample means as would a simple random sample of 20 students.
Sample design for twostage cluster samples
The twostage cluster sample is probably the most commonly used sample design in educational research. This design is generally employed by selecting either schools or classes at the first stage of sampling, followed by the selection of either clusters of students within schools or clusters of students within classes at the second stage. In many studies the twostage cluster design is preferred because this design offers an opportunity for the researcher to conduct analyses at more than one level of data aggregation. For example, the selection of students within classes at the second stage of sampling would, provided there were sufficient numbers of classes and numbers of students selected within classes, permit analyses to be carried out at (a) the betweenstudent level (by
© UNESCO
17
Module 3
Sample design for educational survey research
using data describing individual students), (b) the betweenclass level (by using data based on class mean scores), or (c) both levels simultaneously using ‘multilevel analysis’ procedures. In the following discussion some properties of twostage cluster samples have been described by using the simplifying assumption that the clusters in the population are of equal size. This assumption permits the use of ‘sample design tables’ (see Appendix 2) which provide an excellent aid for designing twostage cluster samples with prespecified levels of sampling accuracy.
1. The design effect for twostage cluster samples
The value of the design effect (Kish, 1965:257) for a twostage cluster sample design depends, for a given cluster size, on the value of the coefficient of intraclass correlation.
���� �
���� � � � � � � ������� ��� ���� � ��� �
where b is the size of the selected clusters, and roh is the coefficient of intraclass correlation
The coefficient of intraclass correlation, often referred to as roh, provides a measure of the degree of homogeneity within clusters. In educational settings the notion of homogeneity within clusters may be observed in the tendency of student characteristics to more homogeneous within schools, or classes, than would be the case if students were assigned to schools, or classes, at random. This homogeneity may be due to common selective factors (for example,
18
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
residential zoning of schools), or to joint exposure to the same external influences (for example, teachers and school programs), or to mutual interaction (for example, peer group pressure), or to some combination of these.
2. The effective sample size for twostage cluster samples
The “effective sample size” (Kish, 1965:259) for a given twostage cluster sample is equal to the size of the simple random sample which has a level of sampling accuracy, as measured by the variance of the sample mean, which is equal to the sampling accuracy of the given twostage cluster sample. A little algebra may be used to demonstrate that the actual sample size, nc, and the effective sample size, n*, for a twostage cluster sample are related to the design effect associated with that sample in the following manner (Ross, 1978:137138).
�� �
� � � ����
From previous discussion, we may replace deff in this formula by an expression which is a function of the cluster size and the coefficient of intraclass correlation.
��
�
� � � �� � �� ����� ����
For example, consider a twostage cluster sample based on a sample of 10 schools followed by the selection of 20 students per school. In addition, consider a student characteristic (for example, a test score or attitude scale score) for which the value of the coefficient of intraclass correlation is equal to 0.1. This value of roh would
© UNESCO
19
Module 3
Sample design for educational survey research
be typical for clusters of students selected randomly from within secondary schools in Australia (Ross, 1983). In this situation, the above formula simplifies to the following expression.
���
�
� � � �� � ��� �� ����
Solving this equation for n* gives a value of 69 for the value of the effective sample size. That is, given the value of 0.1 for roh, a twostage cluster sample of size 200 that is selected by sampling 10 schools followed by sampling 20 students per school would have sampling accuracy equivalent to a simple random sample of 69 students. For a given population of students, the value of roh tends to be higher for clusters based on classes rather than clusters based on schools. Ross (1978) has obtained values of roh as high as 0.5 for mathematics test scores based on classes within Australian secondary schools.
Sample design tables for twostage cluster samples
Sample design tables are often prepared for welldesigned research studies in which it is intended to employ twostage cluster sampling. These tables present a range of sample design options – each designed to have a prespecified level of sampling accuracy. A hypothetical example has been presented in the following discussion in order to illustrate the sequence of calculations and decisions involved in the preparation of these tables. Consider an educational research study in which test items are administered to a twostage cluster sample of students with the aim
20
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
of estimating the percentage of students in the population that are able to obtain correct answers. In addition, assume that a sampling accuracy constraint has been placed on the design of the study so as to ensure that the sample estimate of the percentage of students providing the correct answer, p, will provide p + 5 per cent as 95 per cent confidence limits for the value of the percentage in the population. For reasonably large samples it is possible to assume normality of the distribution of sample estimates (Kish, 1965:1314) and therefore confidence limits of p + 5 per cent are approximately equivalent to an error range of plus or minus two standard errors of p. Consequently, the error constraint placed on the study means that one standard error of p needs to be less than or equal to 2.5 per cent. Consider a simple random sample of n* students selected from this population in order to calculate values of p. Statistical theory may be employed to show that, for large populations, the variance of the sample estimate of p as an estimate of the population value may be calculated by using the following formula (Kish, 1965:46).
��� ��� �
� ���� ������� �� � �
The maximum value of p(100  p) occurs for p = 50. Therefore, in order to ensure that we could satisfy the error constraints described above, the following inequality would need to be valid.
���� �� �
�� ������� ��� �� � �
© UNESCO
21
Module 3
Sample design for educational survey research
That is, the size of the simple random sample, n*, would have to be greater than, or equal to, about 400 students in order to obtain 95 per cent confidence limits of p + 5 per cent. Now consider the size of a twostage cluster sample design which would provide equivalent sampling accuracy to a simple random sample of 400 students. The design of this cluster sample would require knowledge of the numbers of primary sampling units (for example, schools or classes) and the numbers of secondary sampling units (students) which would be required. From previous discussion, the relationship between the size of the cluster sample, nc, which has the same accuracy as a simple random sample of size n* = 400 may be written in the following fashion. This expression is often described as a ‘planning equation’ because it may be used to explore sample design options for twostage cluster samples.
� � � ��� � �� � ������� ����
The value of nc is equal to the product of the number of primary sampling units, a, and the number of secondary sampling units selected from each primary sampling unit, b. Substituting for nc in this formula, and then transposing provides the following expression for a in terms of b and roh.
� �
��� � �� � ������� ���� �
22
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
As an example, consider roh = 0.1, and b = 20. Then we have the following value for a.
� � �
��� � �� � ��� ����� ���� �� ��
That is, for roh = 0.1, a twostage cluster sample of 1160 students (consisting of the selection of 58 primary sampling units followed by the selection of clusters of 20 students) would have sampling accuracy equivalent to a simple random sample of 400 students. In Table 1.1 the planning equation has been employed to list sets of values for a, b, deff, and nc which describe a group of twostage cluster sample designs that have sampling accuracy equivalent to a simple random sample of 400 students. Three sets of sample designs have been listed in the table – corresponding to roh values of 0.1, 0.2, and 0.4. In a study of school systems in ten developed countries, Ross (1983: 54) has shown that values of roh in this range are typical for achievement test scores obtained from clusters of students within schools. The most striking feature of Table 1.1 is the rapidly diminishing effect that increasing b, the cluster size, has on a, the number of clusters that are to be selected. This is particularly noticeable when the cluster size reaches 10 to 15 students.
© UNESCO
23
Module 3
Sample design for educational survey research
Table 1
Sample design table for twostage cluster samples with sampling accuracy equal to a simple random sample of 400 roh = 0.1 deff 1.0 1.1 1.4 1.9 2.4 2.9 nc 400 440 560 760 960 1160 a 400 220 112 76 64 58 52 49 48 deff 1.0 1.2 1.8 2.8 3.8 roh = 0.2 nc 400 480 720 1120 1530 a 400 240 144 112 102 96 91 88 87 deff 1.0 1.4 roh = 0.4 nc 400 560 a 400 280 208 184 176 172 168 166 165
Cluster Size b 1 2 5 10 15 20 30 40 50
2.6 1040 4.6 1840 6.6 2640 8.6 3440 12.6 5040 16.6 6640 20.6 8250
4.8 1920 6.8 2730 8.8 3520 10.8 4350
3.9 1560 4.9 1960 5.9 2400
Consider, for example, two sample designs applied in a situation where a value of roh = 0.4 may be assumed: (a) a total sample of 2640 students obtained by selecting 15 students per cluster from 176 clusters, and (b) a total sample of 8250 students obtained by selecting 50 students per cluster from 165 clusters. From Table 1.1, it may be observed that both of these sample designs have sampling accuracy equivalent to a simple random sample of 400 students.These two sample designs have equivalent sampling accuracy, however there is a striking difference between the each design in terms of total sample size. Further, the magnitude of this difference is not reflected proportionately in the difference between the number of clusters selected.
24
http://www.sacmeq.org and http://www.unesco.org/iiep
Basic concepts of sample design for educational survey research
This result illustrates an important point for the planning of educational research studies that seek to make stable estimates of student population characteristics. This is that the sampling accuracy levels of twostage cluster sample designs, for cluster sizes of 10 or more, tend to be greatly influenced by small changes in the number of clusters that are selected at the first stage of sampling, and relatively less influenced by small changes in the size of the selected clusters. The main use of sample design tables like the one presented in Table 1.1 is to permit the researcher to choose, for a given value of roh, one sample design from among a list of equally accurate sample design options. The final choice between equally accurate options is usually guided by cost factors, or data analysis strategies, or a combination of both of these. For example, the cost of collecting data by using ‘group administration’ of tests, questionnaires, etc. often depends more on the number of selected schools than on the number of students surveyed within each selected school. This occurs because the use of this methodology usually leads to higher marginal costs associated with travel to many schools, compared with the marginal costs of increasing the number of students surveyed within each selected school. In contrast, the cost of collecting data by using ‘individual administration’ of onetoone tests, interviews, etc., normally depends more on the total sample size than on the number of selected schools. The choice of a sample design option may also depend upon the data analysis strategies that are being employed in the research. For example, analyses may be planned at both the betweenstudent and betweenschool levels of analysis. In order to conduct analyses at the betweenschool level, data obtained from individual students may need to be aggregated to obtain files consisting of school records based on student mean scores. This type of analysis
© UNESCO
25
Module 3
Sample design for educational survey research
requires that sufficient students be selected per school so as to ensure that stable estimates are able to be made for individual schools. At the same time, it requires that sufficient schools are available so as to ensure that meaningful results may be obtained for estimates of population characteristics.
26
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
In this section a hypothetical example has been presented in order to illustrate the 12 main steps involved in selecting a national sample of schools and students for ‘Country X’. The numbers of schools and students in the example have been made very small so that all calculations and tabulations can be carried out by hand. In the next section of this module a ‘real’ example has been presented in which the required calculations and tabulations must be carried out by computer. The desired target population for the example consists of 1100 students attending 20 schools located in two administrative regions of Country X. The sample design requires that five schools be selected across four strata with probability proportional to the number of students in the defined target population. Within the five selected schools a simple random sample of 20 students is to be selected. The 12 steps required to complete the sample design cover the following five areas.
• Specification of Sample (Steps 1 to 4)
2
The development and listing of ‘specifications’ for the sample design – including target population, sample size, sampling stages, stratification, sample allocation, cluster size, and pseudoschool
© UNESCO
27
Module 3
Sample design for educational survey research
rules. In most research situations the specifications are prepared following a detailed investigation of the availability of information that is suitable for the construction of sampling frames, and an evaluation of the nature and scope of the data that are to be collected. • Stratification and Sample Allocation (Steps 5 to 7)
A description of the stratification variables and an analysis of the breakdown of the defined target population across strata. The breakdown is used with the sample design specifications to guide the allocation of the sample across strata. • Marker Variables (Step 8)
This is an ‘optional’ area. It provides information that can be used to make comparisons between the population and sample. However, it has no real impact on the scientific quality of the final sample. • Construction of Sampling Frame (Steps 9 and 10)
The construction of a separate list of schools for each stratum. Each school is listed in association with the number of students in the defined target population. Schools that are smaller than a specified size are linked to a similar (and nearby) school to form a ‘pseudoschool’. • Sample Selection (Steps 11 and 12)
The selection of the sample of schools within strata (by using a ‘lottery tickets’ approach), and the selection of a sample of students within schools (by using a table of random numbers).
28
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
E XERCISE A
The reader should work through the 12 steps of the sample design. Where tabulations are presented, the figures in each cell should be verified by hand calculation using the listing of schools that has been provided in step 3. After working through all steps, the following questions should be addressed:
1.
What is the probability of a student in the first school of the first stratum being selected into the sample? Is this probability the same for students in other schools in the first stratum? Is it the same for students in other strata?
2. How would the sampling of students be undertaken for a pseudoschool that was selected at the first stage of sampling? 3. What other variables might be used to stratify the defined target population? Are any of these ‘better’ than the ones used in the example? 4. What would Table 3 and Table 4 look like if the stratification variables were ‘Region’ and ‘School Type’?
© UNESCO
29
Module 3
Sample design for educational survey research
Step 1 List the basic characteristics of the sample design
1.
Desired target population: Grade Six students in Country X.
2. Stratification variables: ‘Region’ and ‘School size’. 3. Sampling stages: First Stage: Schools selected within strata with probability proportional to size. Second Stage: Students selected within schools by using simple random sampling. 4. Minimum cluster size: Fixed cluster of twenty students per school. 5. Number of schools to be selected: A total of five schools. 6. Allocation of the sample across strata: Proportionate allocation across strata. That is, the size of the sample within a stratum should be proportional to the total size of the stratum. 7.
Pseudoschools: Each school in the defined target population with less than 20 students in the defined target population is to be combined with another similar (and nearby) school to form a ‘pseudoschool’.
8. Selection equation
Probability
= =
ah × ( ah × nhi Nh
Nhi nhi )×( ) Nh nh
where ah Nh Nhi nhi
= = = =
the number of schools selected in stratum h, the total number of students in stratum h, the total number of students in school i in stratum h, and the number of students selected from school i.
30
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Step 2 Prepare brief and accurate written descriptions of the desired target population, the defined target population, and the excluded population
1.
Desired Target Population: “Grade Six students in Country X”.
2. Defined Target Population: “All students at the Grade Six level who are attending registered Government or Nongovernment primary schools in all administrative regions of Country X”. 3. Excluded Population: “All students at the Grade Six level attending special schools for the handicapped in Country X”.
© UNESCO
31
Module 3
Sample design for educational survey research
Step 3 Locate a listing of all primary schools in Country_X that includes the following information for each school that has students in the desired target population
1.
Description of listing a. School name. The official (unique) name of the school or a suitable school identification number. b. Region. The major state, province, or administrative region in which the school is located. c. District. The district within an administrative region that is normally supervised by a District Education Officer. d. School type. The name of the authority that administers the school. For this example, a simple classification into ‘Government’ and ‘Nongovernment’ schools. (Many studies use more detailed subgroups.) e. School location. The location of the school in terms of urbanisation. For this example, ‘Urban’ or ‘Rural’ schools. f. School size. The size of the total school enrolment. For this example, a simple classification into ‘Large’ and ‘Small’ schools. (Many studies use more detailed subgroups and some use the ‘exact’ enrolment.) g. School program. A description of school program that is suitable for identifying schools in the ‘excluded population’. For this example, ‘Regular’ and ‘Special’ schools. h. Target population. The number of students within the school who are members of the desired target population. For this example, the total enrolment of students at the Grade Six level.
32
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
2. Contents of Listing School_A School_B School_C School_D School_E School_F School_G School_H School_I School_J School_K School_L School_M School_N School_O School_P School_Q School_R School_S School_T Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Dist_1 Dist_1 Dist_1 Dist_2 Dist_2 Dist_3 Dist_3 Dist_4 Dist_4 Dist_4 Dist_5 Dist_5 Dist_6 Dist_6 Dist_6 Dist_7 Dist_7 Dist_8 Dist_8 Dist_9 Govt Govt NonGovt Govt NonGovt Govt Govt NonGovt NonGovt Govt Govt NonGovt Govt Govt NonGovt Govt Govt Govt NonGovt NonGovt Urban Urban Urban Urban Urban Rural Rural Urban Urban Urban Urban Urban Urban Urban Urban Rural Rural Urban Urban Urban Large Large Small Large Large Small Large Small Small Large Large Large Small Small Large Small Large Large Small Small Regular Regular Regular Special Regular Regular Regular Regular Regular Regular Regular Regular Regular Regular Regular Regular Regular Special Regular Regular 100 60 60 60 70 40 80 50 50 90 60 40 10 80 50 30 50 40 50 30
© UNESCO
33
Module 3
Sample design for educational survey research
Step 4 Use the listing of schools in the desired target population to prepare a tabular description of the desired target population, the defined target population, and the excluded population
There are 1100 students attending 20 schools in the desired target population. Two of these schools are ‘special schools’ and are therefore their 100 students are assigned to the excluded population – leaving 1000 students in 18 schools as the defined target population.
Table 2
The desired, defined and excluded populations
Defined Schools 18 Students 1000 Excluded Schools 2 Students 100
Desired Schools 20 Students 1100
34
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Step 5 Select the stratification variables
The stratification variables are ‘Region’ (which has two categories: ‘Region_1’ and ‘Region_2’) and ‘School Size’ (which has two categories: ‘Large’ and ‘Small’ schools). These two variables combine to form the following four strata. Stratum_1: Region_1 Large Schools Stratum_2: Region_1 Small Schools Stratum_3: Region_2 Large Schools Stratum_4: Region_2 Small Schools
© UNESCO
35
Module 3
Sample design for educational survey research
Step 6 Apply the stratification variables to the desired, defined, and excluded population
Table 3 Schools and students in the desired, defined and excluded populations listed by the four strata
Desired
Schools Students
Stratum Stratum_1 Stratum_2 Stratum_3 Stratum_4 Country X 6 4 5 5
Defined
Schools
Excluded
Students
Students Schools
460 200 240 200 1100
5 4 4 5 18
400 200 200 200 1000
1 0 1 0 2
60 0 40 0 100
20
36
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Step 7 Etablish the allocation of the sample across strata
The sample specifications in Step 1 require five schools to be selected in a manner that provides a proportionate allocation of the sample across strata. Since a fixedsize cluster of 20 students is to be drawn from each selected school, the number of schools to be selected from each stratum must be proportional to the total stratum size. Proportionate Number of Schools. The proportionate number of schools required from the first stratum is 5 x (400/1000) = 2 schools. The other three strata have 200 students each and so the proportionate number of schools for each is 5 x (200/1000) = 1 school.
Table 4
The sample allocation for the defined target population
Population of students
Number Percent
Sample allocation Schools
Exact Rounded
Stratum
Students
Number Percent
Stratum_1 Stratum_2 Stratum_3 Stratum_4
400 200 200 200
40.0 20.0 20.0 20.0
2.0 1.0 1.0 1.0
2 1 1 1
40 20 20 20
40.0 20.0 20.0 20.0
© UNESCO
37
Module 3
Sample design for educational survey research
Step 8 (an optional step) Prepare tabular displays of students and schools in the defined target population – broken down by certain variables that may be of interest for the construction of ‘marker variables’ after the data have been collected
This is an optional step in the sample design because it has no impact upon the scientific quality of the final sample. The main justification for this step is that it may provide useful information (sometimes referred to as ‘marker variables’) that can be used to compare several characteristics of the final sample and the defined target population.
Table 5
Schools with students in the defined target population listed by categories of school size, school location, and school type
School size School location Rural 1 1 1 1 4 Urban 4 3 3 4 14 School type Govt 4 1 2 3 10 NonGovt 1 3 2 2 8 Stratum
Stratum
Large 5 0 4 0 9
Small 0 4 0 5 9
Stratum_1 Stratum_2 Stratum_3 Stratum_4 Country X
5 4 4 5 18
38
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Table 6
Students in the defined target population listed by categories of school size, school location, and school type
School size School location Rural 80 40 50 30 200 Urban 320 160 150 170 800 School type Govt 330 40 110 120 600 NonGovt 70 160 90 80 400 Stratum
Stratum
Large 400 0 200 0 600
Small 0 200 0 200 400
Stratum_1 Stratum_2 Stratum_3 Stratum_4 Country X
400 200 200 200 1000
© UNESCO
39
Module 3
Sample design for educational survey research
Step 9 For schools with students in the defined target population, prepare a separate list of schools for each stratum with ‘pseudoschools’ identified by a bracket ( [ )
The sample design specifications in Step 1 require that 20 students be drawn for each selected school. However, one school in the defined target population, School_M, has only 10 students. This school is therefore combined with a similar (and nearby) school, School_N, to form a ‘pseudoschool’. The pseudoschool is treated as a single school of 90 students during the process of selecting the sample. If a pseudoschool happens to be selected into the sample, it is treated as a single school for the purposes of selecting a subsample of 20 students. Note that School_D and School_R do not appear on the list because they are members of the excluded population.
40
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Stratum 1. Region_1 Large schools
School_A School_B School_E School_G School_J Reg_1 Reg_1 Reg_1 Reg_1 Reg_1 Dist_1 Dist_1 Dist_2 Dist_3 Dist_4 Govt Govt NonGovt Govt Govt Urban Urban Urban Rural Urban Large Large Large Large Large Regular Regular Regular Regular Regular 100 60 70 80 90
Stratum 2.
School_C School_F School_H School_I
Region_1 Small schools
Dist_1 Dist_3 Dist_4 Dist_4 NonGovt Govt NonGovt NonGovt Urban Rural Urban Urban Small Small Small Small Regular Regular Regular Regular 60 40 50 50
Reg_1 Reg_1 Reg_1 Reg_1
Stratum 3. Region_2 Large schools
School_K School_L School_O School_Q Reg_2 Reg_2 Reg_2 Reg_2 Dist_5 Dist_5 Dist_6 Dist_7 Govt NonGovt NonGovt Govt Urban Urban Urban Rural Large Large Large Large Regular Regular Regular Regular 60 40 50 50
Stratum 4. Region_2 Small schools
School_M School_N School_P School_S School_T Reg_2 Reg_2 Reg_2 Reg_2 Reg_2 Dist_6 Dist_6 Dist_7 Dist_8 Dist_9 Govt Govt Govt NonGovt NonGovt Urban Urban Rural Urban Urban Small Small Small Small Small Regular Regular Regular Regular Regular 10 [ 80 30 50 30
© UNESCO
41
Module 3
Sample design for educational survey research
Step 10 For schools with students in the defined target population, assign ‘lottery tickets’ such that each school receives a number of tickets that is equal to the number of students in the defined target population
Note that the pseudoschool made up from School_M and School_N has tickets numbered 1 to 90 because these two schools are treated as a single school for the purposes of sample selection.
Stratum 1. Region_1 Large schools
Lottery tickets
School_A School_B School_E School_G School_J
Reg_1 Dist_1 Govt Reg_1 Dist_1 Govt Reg_1 Dist_2 NonGovt Reg_1 Dist_3 Govt Reg_1 Dist_4 Govt
Urban Urban Urban Rural Urban
Large Large Large Large Large
Regular 100 1100 Regular 60 101160 Regular 70 161230 Regular 80 231310 Regular 90 311400
Stratum 2. Region_1 Small schools
Lottery tickets
School_C School_F School_H School_I
Reg_1 Reg_1 Reg_1 Reg_1
Dist_1 Dist_3 Dist_4 Dist_4
NonGovt Govt NonGovt NonGovt
Urban Rural Urban Urban
Small Small Small Small
Regular Regular Regular Regular
60 40 50 50
160 61100 101150 151200
42
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
Stratum 3. Regional_ 2 Large schools
Lottery tickets
School_K School_L School_O School_Q
Reg_2 Reg_2 Reg_2 Reg_2
Dist_5 Dist_5 Dist_6 Dist_7
Govt NonGovt NonGovt Govt
Urban Urban Urban Rural
Large Large Large Large
Regular Regular Regular Regular
60 40 50 50
160 61100 101150 151200
Stratum 4. Regional_2 Small schools
Lottery tickets
School_M School_N School_P School_S School_T
Reg_2 Reg_2 Reg_2 Reg_2 Reg_2
Dist_6 Dist_6 Dist_7 Dist_8 Dist_9
Govt Govt Govt NonGovt NonGovt
Urban Urban Rural Urban Urban
Small Small Small Small Small
Regular Regular Regular Regular Regular
10 [ 80
190
30 91120 50 121170 30 171200
© UNESCO
43
Module 3
Sample design for educational survey research
Step 11 Select the sample of schools
• Selection of two schools for stratum 1
In the first stratum two schools must be selected with probability proportional to the number of students in the defined target population. This is achieved by allocating a number of ‘lottery tickets’ to each school so that the first school on the list, School_A, with 100 students in the defined target population receives lottery tickets that are numbered from 1 to 100. The second school on the list, School_B, with 60 students receives lottery tickets that are numbered from 101 to 160, and so on – until the final school in the first stratum receives tickets numbered 311 to 400. Since two schools must be selected there is a need to identify two ‘winning tickets’. The ratio of the total number of tickets to the number of winning tickets is known as the ‘sampling interval’. For the first stratum the sampling interval is equal to 400/2 = 200. That, is each ticket in the first stratum should have a 1 in 200 chance of being drawn as a winning ticket. Note that within the other three strata, one winning ticket must be selected from a total of 200 tickets – which gives the same value for the sampling interval (200/1 = 200). The ‘winning tickets’ for the first stratum are drawn by using a ‘random start – constant interval’ procedure whereby a random number in the interval of 1 to 200 is selected as the first winning ticket and the second ticket is selected by adding an increment of 200. Assuming a random start of 105, the winning ticket numbers would be 105 and 305. This results in the selection of School_B (which holds tickets 101 to 160) and School_G (which holds tickets 231 to 310). The probability of selection is proportional to the number of tickets held and therefore each of these schools is selected with probability proportional to the number of students in the defined target population.
44
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
•
Selection of one school each for the other three strata
Only one winning ticket is required for the other strata and therefore one random number must be drawn between 1 and 200 for each stratum. Assuming these numbers are 65, 98, and 176, the selected schools are School_F, School_L, and School_T. • Sample of Schools
Stratum 1. Region_1 Large schools
Lottery tickets
School_B School_G School_F School_L School_T
Reg_1 Reg_1 Reg_1 Reg_2 Reg_2
Dist_1 Dist_3 Dist_3 Dist_5 Dist_9
Govt Govt Govt NonGovt NonGovt
Urban Rural Rural Urban Urban
Large Large Small Large Small
Regular Regular Regular Regular Regular
60 80 40 40 30
105 305 65 98 176
© UNESCO
45
Module 3
Sample design for educational survey research
Step 12 Use a table of random numbers to select a simple random sample of 20 students in each selected school
Within a selected school a table of random numbers is used to identify students from a sequentially numbered roll of students in the defined target population. The application of this procedure has been described in detail by Ross and Postlethwaite (1991). A summary of the procedure has been presented below.
1. Obtain Grade Six register(s) of attendance
These registers are obtained for all students in the defined target population – in this case Grade Six students. In multiple session schools, both morning and afternoon registers are obtained.
2. Place a sequential number beside the name of each Grade Six student
For example, consider a school with one ‘shift’ and a total of 48 students in Grade Six. Commence by placing the number ‘1’ beside the first student on the Register; then place the number ‘2’ beside the second student on the Register; …etc…;finally, place the number ‘48’ beside the last student on the Register. As another example, consider a school with 42 students in the morning ‘shift’ and 48 students in the afternoon session of Grade Six. Commence by placing the number ‘1’ beside the first student on the Morning Register; …etc…; then place a ‘42’ beside the last student on the Morning Register; then place a ‘43’ beside the first student on the Afternoon register; …etc…;finally place a ‘90’ beside the last student on the afternoon register.
46
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of students: a hypothetical example for ‘country x’
3. Locate the appropriate set of selection numbers
In Appendix 6 sets of “random number tables” have been listed for a variety of school sizes. (Note that only the sets relevant for school sizes in the range 21 to 100 have been presented in this Appendix.) For example, if a school has 48 students in Grade Six, the appropriate set of selection numbers is listed under the ‘R48’ heading. Similarly, if a school has 90 Grade Six students then the appropriate set of selection numbers is listed under the ‘R90’ heading.
4. Use the appropriate set of random numbers
After locating the appropriate set of random numbers, use the first random number to locate the Grade Six student with the same sequential number on the Register. Then use the second random number to locate the Grade Six student with the same sequential number on the Register. Continue with this process until the complete sample of students has been selected. For example, in Step 11 the first school selected in Stratum_1 is School_B which has 60 students in the defined target population. Within this school the students selected would be those with the following sequential numbers: 1, 5, 6, 7, 9 …, 52, and 54. (These numbers are obtained from the set of random numbers labelled ‘R60’ in Appendix 6).
© UNESCO
47
Module 3
Sample design for educational survey research
3
How to draw a national sample of schools and students: a real example for Zimbabwe
In Section Two of this module a hypothetical example was presented in order to provide a simple illustration of the 12 main steps involved in selecting a national sample of schools and students. In this section a ‘real’ example has been presented by applying these 12 steps to the selection of a national sample of schools and students in Zimbabwe. For most of these steps, an ‘exercise’ has been given which requires the use of a computer to carry out the kinds of calculations and tabulations that were possible to conduct by hand for the hypothetical example. The example draws upon computerbased techniques that were developed for a largescale study of the quality of education at Grade Six level in Zimbabwe (Ross and Postlethwaite, 1991). The study was undertaken in 1991 by the International Institute for Educational Planning (IIEP, UNESCO) and the Zimbabwe Ministry of Education and Culture. In order to work through the exercises in this section it will be necessary to employ the SAMDEM software (Sylla et al., 2004) in association with a sampling frame that was constructed from the 1991 ‘school forms’ survey of primary schools in Zimbabwe. The database (described in Step 3 of this section, and entitled ‘Zimbabwe.dat’) was developed in association with George Moyo, Patrick Pfukani, and Saul Murimba from the Policy and Planning Unit of the Zimbabwe Ministry of Education and Culture.
48
http://www.sacmeq.org and http://www.unesco.org/iiep
The 12 steps described in this section have been used in many countries to select national samples of schools and students (Ross, 1991). However, it should be emphasized that the general ideas are applicable to samples that might be selected for smaller studies such as surveys of an administrative region or educational district.
Step 1 List the basic characteristics of the sample design
1. Desired Target Population: Grade Six students in Zimbabwe. 2. Stratification Variables: The following two stratification variables were used.
a. ‘Region’: This variable referred to the nine major administrative regions of Zimbabwe and took the following nine values: ‘Harare’, ‘Manica’ (Manicaland), ‘Mascen’ (Mashonaland Central), ‘Masest’ (Mashonaland East), ‘Maswst’ (Mashonaland West), ‘Matnor’ (Matabeleland North), ‘Matsou’ (Matabeleland South), ‘Maving’ (Masvingo), ‘Midlnd’ (Midlands). b. ‘School size’: This variable referred to the size of the school in terms of total school enrolment and took the following two values: ‘L’ (Large) and ‘S’ (Small).
3. Sampling Stages: First Stage: Schools selected within strata with probability proportional to size. Second Stage: Students selected within schools by using simple random sampling. 4. Minimum Cluster Size: A fixed cluster of twenty students per school was accepted because this represented a manageable cluster size for data collection requirements within any single school.
© UNESCO
49
Module 3
Sample design for educational survey research
5. Number of Schools to be Selected: The number of schools required for this example was governed by the requirement that final sample should have an effective sample size for the main criterion variables of at least 400 students. That is, the final sample was required to have sampling accuracy that was equivalent to, or better than, a simple random sample of 400 students.
a. Sample design tables The general sample design framework adopted for the example consisted of a stratified twostage cluster sample design. This permitted the use of sample design tables (Ross, 1987) to provide estimates of the number of schools and students required to obtain a sample with an effective sample size of 400. In order to use the sample design tables, it is necessary to know: the minimum cluster size (the minimum number of students within a school that will be selected for participation in the data collection), and the coefficient of intraclass correlation (a measure of the tendency of student characteristics to be more homogeneous within schools than would be the case if students were assigned to schools at random). From above, it was known that the minimum cluster size adopted for the example was 20 students. Also, from a previous study of Reading Achievement conducted in Zimbabwe (Ross and Postlethwaite, 1991) it was found that the value of roh was around 0.3. In Appendix 2 a set of sample design tables has been presented for various values of the minimum cluster size, and various values of the coefficient of intraclass correlation. The construction of these tables has been described by Ross (1987). In each table, ‘a’ has been used to describe the number of schools, ‘b’ has been used to describe the
50
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
minimum cluster size, and ‘n’ has been used to describe the total sample size. The rows of Table 4 that correspond to a minimum cluster size of one refer to the effective sample size. That is, they describe the size of a simple random sample which has equivalent accuracy. Therefore, the pairs of figures in the fourth and fifth columns in the table all refer to sample designs which have equivalent accuracy to a simple random sample of size 400. The second and third columns refer to an effective sample size of 1600, and the final two pairs of columns refer to effective sample sizes of 178 and 100, respectively. The most important columns of figures in Appendix 2 for this example are the fourth and fifth columns because they list a variety of twostage samples that would result in an effective sample size of 400. To illustrate, consider the intersection of the fourth and fifth columns of figures with the third row of figures in the first page of Appendix 2. The pair of values a=112 and n=560 indicate that if roh is equal to 0.1 and the minimum cluster size, b, is equal to 5, then the twostage cluster sample design required to meet the required sampling standard would be 5 students selected from each of 112 schools – which would result in a total sample size of 560 students. The effect of a different value of roh, for the same minimum cluster size, may be examined by considering the corresponding rows of the table for roh=0.2, 0.3, etc. For example, in the case where roh=0.3, a total sample size of 880 students obtained by selecting 5 students from each of 176 schools would be needed to meet the required sampling standard.
© UNESCO
51
Module 3
Sample design for educational survey research
b. The number of schools required for this example In this study the value selected for the minimum cluster size was 20, and the estimated value of the coefficient of intraclass correlation was 0.3. From the sample design tables in Appendix 8, it may be seen that, in order to obtain a twostage cluster sample with an effective sample size of 400, it is necessary to select a sample of 134 schools – which results in a total sample size of 2680 students.
6. Allocation of the sample across strata: Proportionate allocation across strata. That is, the size of the sample within a stratum should be proportional to the total size of the stratum. 7. Pseudoschools: Each school with less than 20 students in the defined target population is to be combined with another similar (and nearby) school to form a ‘pseudoschool’. 8. Selection Equation
Probability = = where ah Nh Nhi nhi = = = =
ah × ( ah × nhi Nh
Nhi nhi )×( ) Nh nh
the number of schools selected in stratum h, the total number of students in stratum h, the total number of students in school i in stratum h, and the number of students selected from school i.
52
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
Step 2 Prepare brief and accurate written descriptions of the desired target population, the defined target population, and the excluded population
1. Desired target population: “Grade Six students in Zimbabwe”. 2. Defined target population: “All students at the Grade Six level in September 1991 who are attending registered Government or Nongovernment primary schools that offer the ‘regular’ curriculum, that contain at least 10 Grade Six students, and that are located in all nine administrative regions of Zimbabwe”. 3. Excluded Population: “All students at the Grade Six level in September 1991 who are attending (a) special schools for the handicapped in Zimbabwe (these schools do not offer the ‘regular’ curriculum), or (ii) primary schools in Zimbabwe with less than 10 Grade Six students (these schools are smaller than the specified minimum size)”.
© UNESCO
53
Module 3
Sample design for educational survey research
Step 3 Locate a listing of all primary schools in Zimbabwe that includes the following information for each school that has students in the desired target population
E XERCISE B
Use the SAMDEM software to list the following groups of schools in the Zimbabwe.dat file:
1.
the first ten schools in the file,
2. schools that contain students in the ‘excluded population’ (special and small schools) and, 3. schools that must be combined to form `pseudoschools’.
Description of Listing (Stored in ‘Zimababwe.dat’ File)
•
Schlname (Columns 115). This variable describes the official name given to each the school by the Zimbabwe Ministry of Education and Culture. Region (Columns 1722). This variable describes the administrative regions of Zimbabwe. The variable takes nine values: ‘Harare’, ‘Manica’ (Manicaland), ‘Mascen’ (Mashonaland Central), ‘Masest’ (Mashonaland East), ‘Maswst’ (Mashonaland West), ‘Matnor’ (Matabeleland North), ‘Matsou’ (Matabeleland South), ‘Maving’ (Masvingo), ‘Midlnd’ (Midlands). District (Columns 2429). This variable describes the administrative district within each administrative region of Zimbabwe. The variable takes 55 values ranging from ‘Harare’ (which is congruent with the region of Harare) to ‘Zvisha’ (which is located in the region of Midlands).
•
•
54
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
•
Schltype (Columns 3133). This variable describes schools according to the authority responsible for school administration. The variable takes two values: ‘Gov’ (Government Schools), and ‘Non’ (Nongovernment schools administered by District Councils, Rural Councils, Missions, Farm School Councils, Trust Councils, etc.). Schllocn (Column 35). This variable describes the location of the school in terms of the degree of urbanization of the surrounding community. The variable takes two values ‘R’ (rural location) and ‘U’ (urban location).
•
Contents of Listing (First Ten Schools in ‘Zimbabwe.dat’ File)
Alpha Brick Royden St Marnocks Chizungu Nyabira Kubatana Kintyre Gwebi Henderson Beatrice Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Harare Non Non Non Non Non Non Non Non Non Non R R R R R R R R R R L L L L L L L L L L Reg Reg Reg Reg Reg Reg Reg Reg Reg Reg 86 34 125 302 102 43 61 90 36 24
•
Schlsize (Column 37). This variable describes each school in terms of the size of its enrolment at the Grade Six level. The variable takes two values: ‘S’ (a ‘small’ school) and ‘L’ (a ‘large’ school). Schlprog (Columns 3941). This variable differentiates between school programs according to whether schools are regular primary schools or special schools designed for handicapped students. The variable takes two values ‘Reg’ (regular schools) and ‘Spe’ (special schools). Enrolg6 (Columns 4345). This variable describes the Grade Six enrolment for each Zimbabwe primary school in 1991.
•
•
© UNESCO
55
Module 3
Sample design for educational survey research
Step 4 Use the listing of schools in the desired target population to prepare a tabular description of the desired target population, the defined target population, and the excluded population
E XERCISE C
Use the SAMDEM software to prepare frequency distributions for the desired, defined, and excluded populations. Then complete Table 3.1.
Table 7
Schools and students in the desired, defined and excluded populations (Zimbabwe Grade Six 1991)
Defined Excluded
Special Schl Small Stdt Schl Stdt
Desired
Schools Students Schools Students Schl Stdt
4487
283781




65
393
56
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
Step 5 Select the stratification variables
E XERCISE D
Use the names of the stratification variables presented in Step 1 to complete the following list.
Stratum_1: Harare Region Large Schools Stratum_2: Harare Region Small Schools Stratum_3: Manica Region Large Schools Stratum_4: Manica Region Small Schools .................................. .................................. .................................. .................................. .................................. .................................. .................................. .................................. .................................. .................................. .................................. .................................. Stratum_17: Midlnd Region Large Schools Stratum_18: Midlnd Region Small Schools
© UNESCO
57
Module 3
Sample design for educational survey research
Step 6 Apply the stratification variables to the desired, defined, and excluded population
E XERCISE E
Use the SAMDEM software to prepare frequency distributions for the desired, defined, and excluded populations – separately for each stratum. Then complete Table 3.2
Table 8
The desired, defined, and excluded populations (Zimbabwe Grade Six 1991)
Stratum Desired Sch. 140 82 Stu. 25054 3367 Defined Sch. 140 72 4399 Stu. 25054 3298 Special Sch. 0 8 23 Stu. 0 61 275 Excluded Verysmall Sch. 0 2 65 Stu. 0 8 393
No
Region Harare Harare Manica Manica Mascen Mascen Masest Masest Maswst Maswst Matnor Matnor Matsou Matsou Maving Maving Midlnd Midlnd
Sch. size Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small
01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18
Zimbabwe
4487 283781
58
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
Step 7 Establish the allocation of the sample across strata
E XERCISE F
Use the SAMDEM software to establish a proportionate allocation of the sample across the strata. Then complete Table 3.3.
Table 9
The sample allocation for the defined population (Zimbabwe Grade Six 1991)
Stratum Defined Pct. 8.8 1.2 100.0 Schools Exact 11.9 1.6 134.0 Sample Students Pct. 8.9 1.5 100.0
No
Region Harare Harare Manica Manica Mascen Mascen Masest Masest Maswst Maswst Matnor Matnor Matsou Matsou Maving Maving Midlnd Midlnd
Sch. size Number Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small 25054 3298 283113
Planned Number 12 2 135 240 40 2700
01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18
Zimbabwe Total
© UNESCO
59
Module 3
Sample design for educational survey research
Step 8 (an optional step) Prepare tabular displays of students and schools in the defined target population – broken down by School Size, Location, and Type
E XERCISE G
Use the SAMDEM software to prepare ‘marker variable’ information. Complete Tables 3.4 and 3.5.
Table 10
Schools with students in the defined target population listed by school size, school location, and school type (Zimbabwe Grade Six 1991)
Stratum School size Large 140 0 1142 Small 0 72 3257 School location Rural 8 26 4036 Urban 132 46 363 School type Govt 99 17 270 NonGovt 41 55 4129 Stratum Total 140 72 4399
No 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18
Region Harare Harare Manica Manica Mascen Mascen Masest Masest Maswst Maswst Matnor Matnor Matsou Matsou Maving Maving Midlnd Midlnd
Sch. size Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small
Zimbabwe Total
60
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
Table 11
Students in the defined target population listed by school size, school location, and school type (Zimbabwe Grade Six 1991)
Stratum School size Large 25054 0 139642 Small 0 3298 143471 School location Rural 1175 1062 239428 Urban 23879 2236 43685 School type Govt 17798 913 35763 NonGovt 7256 2385 247350 Stratum Total 25054 3298 283113
No 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18
Region Harare Harare Manica Manica Mascen Mascen Masest Masest Maswst Maswst Matnor Matnor Matsou Matsou Maving Maving Midlnd Midlnd
Sch. size Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small Large Small
Zimbabwe Total
© UNESCO
61
Module 3
Sample design for educational survey research
Step 9 For schools with students in the defined target population, prepare a separate list of schools for each stratum with ‘pseudoschools’ identified with a bracket ( [ )
E XERCISE H
Use the SAMDEM software to list the schools having students in the defined target population and located in ‘stratum 2: Harare region small schools!’.
Step 10 For the defined target population, assign ‘lottery tickets’ such that each school receives a number of tickets that is equal to the number of students in the defined target population
E XERCISE I
Use the SAMDEM software to assign ‘lottery tickets’ for schools having students in the defined target population and located in ‘stratum 2: Harare region small schools’. repeat this exercise for one other stratum.
62
http://www.sacmeq.org and http://www.unesco.org/iiep
How to draw a national sample of schools and students: a ‘real’ example for Zimbabwe
Step 11 Select the sample of schools.
E XERCISE J
Use the SAMDEM software to to select the sample of schools from schools having students in the defined target population and located in ‘stratum 2: Harare region small schools’. repeat this exercise for one other stratum.
Step 12 Use a table of random numbers to select a simple random sample of 20 students in each school.
E XERCISE K use the random number tables (see Appendix 1) to identify student selection numbers for the schools that were selected from ‘stratum 2: Harare region small schools’. repeat this exercise for one other stratum.
© UNESCO
63
Module 3
Sample design for educational survey research
4
The estimation of sampling errors
The computational formulae required to estimate the variance of descriptive statistics, such as sample means, are widely available for simple random sample designs and some sample designs that incorporate complexities such as stratification and cluster sampling. However, in the case of more complex analytical statistics, such as correlation coefficients and regression coefficients, the required formulae are not readily available for sample designs which depart from the model of simple random sampling. These formulae are either enormously complicated or, ultimately, they prove resistant to mathematical analysis (Frankel, 1971). In the absence of suitable formulae, a variety of empirical techniques have emerged in recent years which provide «approximate variances that appear satisfactory for practical purposes» (Kish, 1978:20). The most frequently applied empirical techniques may be divided into two broad categories : Subsample Replication and Taylor’s Series Approximation. In Subsample Replication a total sample of data is used to construct two or more subsamples and then a distribution of parameter estimates is generated by using each subsample. The subsample results are analyzed to obtain an estimate of the parameter, as well as a confidence assessment for that estimate (Finifter, 1972). The main approaches in using this technique have been Independent Replication (Deming, 1960), Jackknifing (Tukey, 1958), and Balanced Repeated Replication (McCarthy, 1966).
64
http://www.sacmeq.org and http://www.unesco.org/iiep
Alternatively, Taylor’s Series Approximation may be used to provide a more ‘direct’ method of variance estimation than these three approaches. In the absence of an exact formula for the variance, the Taylor’s Series is used to approximate a numerical value for the first few terms of a series expansion of the variance formula. A number of computer programs have been prepared in order to carry out the extensive numerical calculations required for this approach (Wolter, 1985, Appendix ). In the remainder of this section the Jackknife procedure has been described. This procedure offers the following two important benefits. • Greater Flexibility The Jackknife may be applied to a wide variety of sample designs whereas (i) Balanced Repeated Replication is designed for application to sample designs that have precisely two primary sampling units per stratum, and (ii) Independent Replication requires a large number of selections per stratum so that a reasonably large number of independent replicated samples can be formed Ease of Use The Jackknife does not require specialized software systems whereas (i) Balanced Repeated Replication usually requires the prior establishment by computer of complex Hadamard matrices that are used to ‘balance’ the half samples of data that form replications of the original sample, and (ii) Taylor’s Series methods require specific software routines to be available for each statistic under consideration.
•
© UNESCO
65
Module 3
Sample design for educational survey research
The Jackknife Procedure
The development of the Jackknife procedure may be traced back to a method used by Quenouille (1956) to reduce the bias of estimates. Further refinement of the method (Mosteller and Tukey, 1968) led to its application in a range of social science situations where formulae are not readily available for the calculation of sampling errors. The Jackknife procedure requires that an initial estimate, yall, of a statistic, y, be made on the total sample of data. The total sample is then divided into k subgroups and yi is computed, using the same functional form as yall, but based on the reduced sample of data obtained by omitting the ith subgroup. Then k ‘pseudovalues’ yi* (i=1,2,...,k) can be defined – based on the k reduced samples:
�� � ������������� � � �� �
Quenouille’s estimator (also called the ‘Jackknife value’) is the mean of the k pseudovalues:
�� � ��� ���� �� � �
Quenouille’s contribution was to show that, while yall may have bias of order 1/n as an estimate of y, the Jackknife value, y*, has bias of order 1/n2.
66
http://www.sacmeq.org and http://www.unesco.org/iiep
The estimation of the sampling errors
The variance of y* may be estimated from the pseudovalues.
���� � � � ���� � � � � � � � � � � ��� � � � ���� � � ��������
�
Tukey (1958) set forward the proposal that the pseudovalues could be treated as if they were approximately independent observations and that Student’s t distribution could be applied to these estimates in order to construct confidence intervals for y *. Later empirical work conducted by Frankel (1971) provided support for these proposals when the Jackknife technique was applied to complex sample designs and a variety of simple and complex statistics. Substituting for yi* in the expression for var(y*) permits the variance of y* to be estimated from the k subsample estimates, yi, and their mean – without the need to calculate pseudovalues.
�
���� � � �
�
������� � � � ��� � � ����� � � �
Wolter (1985, p. 156) has shown that replacing y* by yall in the right hand side of the first expression for var(y *) given above provides a conservative estimate of var(y*) – the overestimate being equal to (yally*)2/(k1).
© UNESCO
67
Module 3
Sample design for educational survey research
In practice, these expressions for var(y*) have also been used to estimate the variance not only of Quenouille’s estimator y*, but also of yall (Wolter, 1985, pp. 155, 172). Substituting yall for y* provides the following estimator for the variance of y* (and also a conservative estimator for yall).
� � ���� � � � � ���� � � � ���� ��� � ��������
�
Wolter (1985, p.180) and Rust (1985) have presented an extension of these formulae for complex stratified sample designs in which there are kh primary sampling units in the hth stratum (where h = 1,2,...,H). In this case, the formula for the variance of yall employs yhi to denote the estimator derived from the same functional form as yall – calculated after deleting the ith primary sampling unit from the hth stratum.
���� � ��� � � ���� �
� � ������ ��
����� � � ������ ��� � �
�
where K = Σkh is the total number of samples that are formed
In many educational research studies, schools are used as the primary sampling units, and therefore the K estimates of yhi are obtained by removing one school at a time from the total sample and then applying the same functional form used to estimate yall for each reduced sample. In studies where geographical areas are used
68
http://www.sacmeq.org and http://www.unesco.org/iiep
The estimation of the sampling errors
as the primary sampling units, followed by the selection of more than one school per area, the estimates of yhi are based on reduced samples formed by omitting one geographical area at a time.
© UNESCO
69
Module 3
Sample design for educational survey research
5
References
Brickell, J. L. (1974). “Nominated samples from public schools and statistical bias”. American Educational Research Journal, 11(4), 333341. Deming, W. E. (1960). Sample design in business research. New York: Wiley. Finifter, B. M. (1972). “The generation of confidence: Evaluating research findings by random subsample replication”. In H. L. Costner (Ed.), Sociological Methodology. San Francisco: JosseyBass. Frankel, M. R. (1971). Inference from survey samples. Ann Arbor, Michigan: Institute for Social Research. Haggard, E.A. (1958). Intraclass correlation and the analysis of variance. New York: Dryden. Hansen, M. H., Hurwitz, W. N.; Madow, W. G. (1953). Sample survey methods and theory (Vols. 1 & 2). New York: Wiley. International Association for the Evaluation of Educational Achievement (IEA) (1986). Technical committee meeting: sample design and sample sizes. Paper distributed at the Annual General Meeting of the IEA, Stockholm, Sweden. Jessen, R. L. (1978). “Statistical survey techniques”. New York: Wiley.
70
http://www.sacmeq.org and http://www.unesco.org/iiep
Kalton, G. (1983). “Introduction to survey sampling” (Sage Series in Quantitative Applications in the Social Sciences, No. 35). Beverly Hills, CA: Sage. Kish, L. (1965). Survey sampling. New York: Wiley. Kish, L. (1978). “On the future of survey sampling”. In N. K. Namboordi (Ed.), Survey sampling and measurement. New York: Academic Press. McCarthy, P. J. (1966). Replication: An approach to the analysis of data from complex surveys. Washington: United States National Centre for Health Statistics. Mosteller, F. and Tukey, J. (1968). “Data Analysis Including Statistics”. In G. Lindzey and E. Aronson (Eds.), The handbook of social psychology. (2nd ed.). Reading, Massachusetts: AddisonWesley. Quenouille, M. J. (1956). “Notes on Bias in Estimation”. Biometrica, 43, 353360. Ross, K. N. (1976). Searching for uncertainty: Sampling errors in educational survey research. Hawthorn, Victoria: Australian Council for Educational Research. Ross, K. N. (1978). Sample design for educational survey research (Monograph). Evaluation in Education, 2, 105195. Ross, K. N. (1987). Sample design. International Journal of Educational Research, 11(1), 5775. Ross K. N. (1991). Sampling Manual for the IEA International Study of Reading Literacy. Hamburg: IEA International Study of Reading Literacy Coordinating Centre.
© UNESCO
71
Module 3
Sample design for educational survey research
Ross, K. N. and Postlethwaite, T. N. (1992). “Indicators of the Quality of Education: A National Study of Primary Schools in Zimbabwe”. Paris: International Institute for Educational Planning. Rust, K. (1985). Variance estimation for complex estimators in sample surveys. Journal of Official Statistics, 1(4), 381397. Sylla, K.; Saito, M. and Ross,K. (2004). SAMDEM: Sample Design Manager. Paris. International Institute for Educational Planning. Tukey, J. W. (1958). Bias and confidence in notquite large samples (Abstract). Annals of Mathematical Statistics, 29, 614. Wolf, R. M. (1977). Achievement in America: National report of the United States for the International Educational Achievement Project. New York: Teachers College Press. Wolter, K. M. (1985). Introduction to Variance Estimation. New York: SpringerVerlag.
72
http://www.sacmeq.org and http://www.unesco.org/iiep
Appendix 1 Random number tables for selecting a simple random sample of twenty students from groups of students of size 21 to 100
Size of Group
R21 1 2 3 4 5 6 7 8 9 10 11 12 13 14 16 17 18 19 20 21 R22 1 2 3 5 6 7 8 9 10 11 12 13 14 15 16 17 18 20 21 22 R23 1 2 3 5 6 7 9 11 12 13 14 15 16 17 18 19 20 21 22 23 R24 2 3 4 5 7 8 9 10 11 12 13 15 16 17 18 19 21 22 23 24 R25 1 2 3 4 5 6 7 9 11 12 13 14 15 16 18 20 21 23 24 25 R26 1 2 3 5 6 7 8 9 10 12 13 14 15 16 18 19 21 22 25 26 R27 1 2 3 4 6 7 8 11 12 13 15 17 18 19 20 21 22 25 27 28 R28 1 2 3 4 6 7 8 11 12 13 15 17 18 19 20 21 22 25 27 28 R29 1 2 3 4 5 6 9 10 11 12 14 16 21 22 23 24 25 27 28 29 R30 4 6 7 8 10 12 15 16 17 18 19 20 22 23 24 25 26 27 29 30
© UNESCO
73
Module 3
Sample design for educational survey research
Size of Group
R31 1 4 5 8 10 11 14 15 16 17 19 20 22 23 24 25 27 29 30 31 R32 4 6 7 9 10 11 12 13 14 15 16 17 21 23 24 25 26 28 30 31 R33 1 2 4 5 7 10 12 13 14 16 18 19 21 23 25 26 28 31 32 33 R34 2 3 4 5 6 710 12 13 14 16 18 19 21 23 25 26 28 31 32 33 R35 1 2 7 8 9 11 14 15 16 17 21 22 23 24 25 26 32 33 34 35 R36 4 5 7 8 10 13 16 17 19 21 22 24 25 26 27 28 29 31 33 36 R37 1 3 6 7 8 9 10 11 12 13 17 18 19 24 25 28 29 30 34 36 R38 1 4 7 9 11 12 14 16 17 18 19 22 23 24 25 26 28 31 32 37 R39 1 2 3 5 6 7 8 12 14 15 17 20 24 26 27 28 29 31 36 38 R40 2 3 5 6 8 12 13 16 17 18 23 24 25 31 33 34 37 38 39 40
Size of Group
R41 1 4 5 7 8 10 12 13 15 17 20 21 25 27 28 29 32 34 35 39 R42 2 3 4 5 6 8 10 13 16 20 21 23 24 27 29 31 34 35 36 41 R43 1 2 6 11 13 14 15 17 20 25 26 28 31 32 33 37 38 39 40 42 R44 1 3 4 6 7 8 9 11 13 15 18 19 24 25 32 35 38 40 43 44 R45 3 5 7 8 10 11 12 16 19 23 24 25 32 34 35 37 38 40 43 44 R46 1 11 12 15 17 21 22 23 24 25 26 28 29 32 34 36 40 41 44 45 R47 3 5 10 17 24 25 27 28 30 31 32 33 34 36 37 38 39 42 44 45 R48 1 3 6 15 19 20 24 26 27 29 30 31 32 34 36 37 38 40 44 48 R49 1 2 4 6 8 13 15 17 20 23 24 27 31 32 34 37 41 43 44 47 R50 3 4 5 6 7 8 11 13 14 16 20 22 24 29 32 33 36 37 40 45
74
http://www.sacmeq.org and http://www.unesco.org/iiep
Appendix 1
Size of Group
R51 3 4 8 9 10 12 13 15 17 26 27 29 31 32 36 39 40 44 47 48 R52 4 5 6 10 13 17 18 19 20 22 25 27 28 32 35 40 42 49 51 52 R53 3 6 8 9 11 12 14 15 18 19 23 26 28 33 38 42 43 48 51 53 R54 1 2 4 8 10 11 16 20 24 25 27 28 29 32 37 39 41 46 47 49 R55 1 6 7 9 10 12 14 16 29 31 36 38 40 42 44 47 48 52 53 54 R56 2 5 7 10 12 15 18 24 25 27 32 34 38 43 45 47 49 50 52 55 R57 2 6 7 10 17 19 23 29 34 35 37 38 43 44 46 48 52 53 55 57 R58 3 5 6 8 16 17 21 23 27 32 37 38 42 44 46 48 49 50 52 56 R59 3 4 5 8 10 11 15 16 17 20 24 26 29 31 36 38 42 48 50 53 R60 1 5 6 7 9 11 14 16 19 21 24 28 29 39 40 49 50 51 52 54
Size of Group
R61 2 5 7 8 11 18 19 21 22 27 30 34 35 39 46 48 49 50 53 54 R62 9 12 14 22 23 25 227 28 29 33 41 42 43 46 50 51 53 57 59 62 R63 8 12 14 15 16 18 24 32 33 34 38 39 40 42 45 46 48 54 61 63 R64 5 8 10 17 19 21 25 26 27 29 33 35 37 39 45 48 49 54 58 61 R65 3 4 7 10 14 18 19 20 28 37 40 42 46 47 49 51 58 59 61 65 R66 1 2 8 9 11 19 22 23 25 26 27 30 28 46 50 52 53 54 61 66 R67 4 5 6 10 14 17 18 20 21 25 29 30 34 37 39 41 47 57 58 63 R68 2 4 5 10 16 17 21 22 23 28 31 32 33 37 42 49 55 58 61 67 R69 6 9 11 14 23 24 29 31 36 40 44 48 49 50 52 55 59 60 67 68 R70 5 6 7 9 11 13 17 18 21 22 36 39 41 47 49 50 53 55 61 67
© UNESCO
75
Module 3
Sample design for educational survey research
Size of Group
R71 9 15 17 20 22 25 27 28 34 35 37 39 46 48 50 53 60 61 62 64 R72 3 6 10 12 18 20 22 26 29 32 38 41 43 44 47 48 50 52 61 66 R73 1 5 9 11 12 13 14 21 28 29 34 37 42 48 54 57 62 63 68 72 R74 4 15 19 23 24 27 28 35 38 42 47 49 50 51 52 56 62 64 65 72 R75 2 7 9 15 16 23 29 36 37 41 43 45 46 50 53 60 66 67 69 72 R76 6 20 21 22 24 25 30 31 35 37 39 50 51 58 61 63 65 67 68 73 R77 3 5 6 7 10 16 24 31 32 36 38 45 48 62 64 70 71 72 73 75 R78 12 13 23 24 28 32 46 47 48 49 52 53 57 59 63 64 67 70 75 76 R79 8 10 14 15 21 23 31 32 24 41 46 48 54 58 61 62 68 71 75 77 R80 6 10 17 19 21 23 26 31 32 36 41 42 44 45 46 50 61 71 73 76
Size of Group
R81 9 13 16 33 40 41 42 44 45 46 54 59 64 71 72 73 74 75 76 79 R82 9 15 16 23 27 28 29 33 42 43 48 50 56 57 60 61 66 67 69 71 R83 2 4 14 17 30 35 41 42 49 52 53 58 63 67 69 71 74 76 77 80 R84 4 6 7 13 14 19 25 31 34 39 40 44 59 62 70 73 74 77 83 84 R85 3 6 7 25 27 30 32 36 45 48 50 54 58 63 64 66 68 72 83 85 R86 1 8 13 24 34 35 39 45 47 50 52 61 64 65 66 67 68 75 80 83 R87 1 3 7 14 16 17 19 20 26 28 30 40 53 54 60 66 77 80 83 86 R88 6 8 16 23 26 29 32 35 42 43 48 50 55 59 63 64 70 74 81 86 R89 2 12 14 15 39 48 52 56 57 60 62 64 66 67 68 70 75 76 88 89 R90 1 2 5 10 22 24 25 27 31 35 40 46 51 53 54 73 81 83 85 90
76
http://www.sacmeq.org and http://www.unesco.org/iiep
Appendix 1
Size of Group
R91 7 15 20 29 36 38 42 43 49 50 54 58 59 67 73 80 84 85 87 91 R92 8 15 17 21 25 26 31 34 40 41 49 54 61 69 71 78 79 81 83 84 R93 1 7 9 19 20 21 22 30 31 32 35 36 40 46 51 62 72 73 76 89 R94 1 8 9 10 14 15 20 23 36 46 48 57 58 60 61 72 79 81 87 91 R95 7 9 12 14 18 24 32 35 46 49 54 60 63 64 69 78 87 88 92 95 R96 2 4 6 9 16 18 21 28 32 47 50 51 63 64 70 73 78 80 89 92 R97 3 10 26 33 37 39 40 41 48 50 51 53 61 62 64 65 68 70 82 97 R98 8 11 17 23 25 31 36 38 42 48 54 58 63 70 71 78 89 91 92 93 R99 3 11 20 29 30 34 38 42 50 51 53 54 60 62 63 68 79 92 94 99 R100 7 21 22 25 27 35 40 49 50 52 56 57 75 79 81 86 87 92 93 94
© UNESCO
77
Module 3
Sample design for educational survey research
Appendix 2 Sample design tables
Cluster size b roh = 0.1 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.2 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.3 1 (SRS) 2 5 10 15 20 30 40 50 1600 1040 704 592 555 536 518 508 503 1600 2080 3520 5920 8325 10720 15540 20320 25150 400 260 176 148 139 134 130 127 126 400 520 880 1480 2085 2680 3900 5080 6300 178 116 79 66 62 60 58 57 56 178 232 395 660 930 1200 1740 2280 2800 100 65 44 37 35 34 33 32 32 100 130 220 370 252 680 990 1280 1600 1600 960 576 448 406 384 363 352 346 1600 1920 2880 4480 6090 7680 10890 14080 17300 400 240 144 112 102 96 91 88 87 400 480 720 1120 1530 1920 2730 3520 4350 178 107 65 50 46 43 41 40 39 178 214 325 500 690 860 1230 1600 1950 100 60 36 28 26 24 23 22 22 100 120 180 280 390 480 690 880 1100 1600 880 448 304 256 232 208 196 189 1600 1760 2240 3040 3840 4640 6240 7840 9450 400 220 112 76 64 58 52 49 48 400 440 560 760 960 1160 1560 1960 2400 178 98 50 34 29 26 24 22 21 178 196 250 340 435 520 720 880 1050 100 55 28 19 16 15 13 13 12 100 110 140 190 240 300 390 520 600 ±0.05s/±2.5% a n ±0.1s/±5.0% a n
(for roh values of 0.1 to 0.9)
95% confidence limits for means/percentages ±0.15s/±7.5% a n ±0.2s/±10.0% a n
78
http://www.sacmeq.org and http://www.unesco.org/iiep
Cluster size b roh = 0.4 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.5 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.6 1 (SRS) 2 5 10 15 20 30 40 50
95% confidence limits for means/percentages ±0.05s/±2.5% a 1600 1120 832 736 704 688 672 664 660 1600 1200 960 880 854 840 827 820 816 1600 1280 1088 1024 1003 992 982 976 973 n 1600 2240 4160 7360 10560 13760 20160 26560 33000 1600 2400 4800 8800 12810 16800 24810 32800 40800 1600 2560 5440 10240 15045 19840 29460 39040 48650 ±0.1s/±5.0% a 400 280 208 184 176 172 168 166 165 400 300 240 220 214 210 207 205 204 400 320 272 256 251 248 246 244 244 n 400 560 1040 1840 2640 3440 5040 6640 8250 400 600 1200 2200 3210 4200 6210 8200 10200 400 640 1360 2560 3765 4960 7380 9760 12200 ±0.15s/±7.5% a 178 125 93 82 79 77 75 74 74 178 134 107 98 95 94 92 92 91 178 143 122 114 112 111 110 109 109 n 178 250 465 820 1185 1540 2250 2960 3700 178 268 535 980 1425 1880 2760 3680 4550 178 286 610 1140 1680 2220 3300 4360 5450 ±0.2s/±10.0% a 100 70 52 46 44 43 42 42 42 100 75 60 55 54 53 52 52 51 100 80 68 64 63 62 62 61 61 n 100 140 260 460 660 860 1260 1680 2100 100 150 300 550 810 1060 1560 2080 2550 100 160 340 640 945 1240 1860 2440 3050
© UNESCO
79
Module 3
Sample design for educational survey research
Cluster size b roh = 0.7 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.8 1 (SRS) 2 5 10 15 20 30 40 50 roh = 0.9 1 (SRS) 2 5 10 15 20 30 40 50
95% confidence limits for means/percentages ±0.05s/±2.5% a 1600 1360 1216 1168 1152 1144 1136 1132 1130 1600 1440 1344 1312 1302 1296 1291 1288 1287 1600 1520 1472 1456 1451 1448 1446 1444 1444 n 1600 2720 6080 11680 17280 22880 34080 45280 56500 1600 2880 6720 13120 19530 25920 38730 51520 64350 1600 3040 7360 14560 21765 28960 43380 57760 72200 ±0.1s/±5.0% a 400 340 304 292 288 286 284 283 283 400 360 336 328 326 324 323 322 322 400 380 368 364 363 362 362 361 361 n 400 680 1520 2920 4320 5720 8520 11320 14150 400 720 1680 3280 4890 6480 9690 12880 16100 400 760 1840 3640 5445 7240 10860 14440 18050 ±0.15s/±7.5% a 178 152 136 130 129 128 127 126 126 178 161 150 146 145 145 144 144 144 178 170 164 162 162 162 161 161 161 n 178 304 680 1300 1935 2560 3810 5040 6300 178 322 750 1460 2175 2900 4320 5760 7200 178 340 820 1620 2430 3240 4830 6440 8050 ±0.2s/±10.0% a 100 85 76 73 72 72 71 71 71 100 90 84 82 82 81 81 81 81 100 95 92 91 91 91 91 91 91 n 100 170 380 730 1080 1440 2130 2840 3550 100 180 420 820 1230 1620 2430 3240 4050 100 190 460 910 1365 1820 2730 3640 4550
80
http://www.sacmeq.org and http://www.unesco.org/iiep
Appendix 3 Estimation of the coeff icient of intraclass correlation
The coefficient of intraclass correlation (roh) was developed earlier this century in connection with studies carried out to measure ‘fraternal resemblance’ – such as in the calculation of the correlation between the heights of brothers. To establish this correlation there was generally no reason for ordering the pairs of measurements obtained from any two brothers. Therefore the approach used was to calculate a productmoment correlation coefficient from a symmetrical table of measures consisting of two interchanged entries for each pair of brothers. In the days before computers, these calculations became extremely arduous when large numbers of brothers from large families were being studied. Some computationally simpler formulae for calculating estimates of this coefficient were eventually developed by several statisticians (Haggard, 1958). These formulae were based either on analysis of variance methods, or on the calculation of the variance of elements and the variance of the means of groups, or clusters, of elements. In the field of educational survey research, the clusters of elements which define primary sampling units often refer to students that are grouped into schools. The value of roh then provides a measure of the tendency for students to be more similar within schools (on some given characteristic such as achievement test scores) than would be the case if students had been allocated at random among schools. The estimate of roh is prepared by using schools as ‘experimental treatments’ in calculating the betweenclusters mean square (BCMS) and the withinclusters mean square (WCMS). The term b refers to the number of students per school.
© UNESCO
81
Module 3
Sample design for educational survey research
��������� ��� �
����������� ���� � �� � �� ����
If the numerator and denominator on the right hand side of the above expression are both divided by the value of WCMS, the estimated roh can be expressed in terms of the value of the F statistic and the value of b.
��������� ��� �
� ���� � � �� � ��
The application of the following alternative formula which is based upon variance estimates for elements (students) and cluster means (school means) has been discussed by Kish (1965: 176). The term sc2 is the variance of cluster means, and s2 is the variance of element values.
� �� ������ ��������� ��� � ������� ��
Note that, in situations where the number of elements per cluster varies, the value of b is sometimes replaced by the value of the average cluster size.
82
http://www.sacmeq.org and http://www.unesco.org/iiep