S. A. Paranjpe and Anita Shah
The University of Pune, India
Journal of Statistics Education v.8, n.2 (2000)
Copyright (c) 2000 by S. A. Paranjpe and Anita Shah, all rights reserved. This text may be freely shared among individuals, but it may not be republished in any medium without express written consent from the authors and advance notification of the editor.
Key Words: Active learning; Practicals; Sample surveys; Statistics.
In Indian Universities, courses titled 'Statistics Practical' usually involve only numerical evaluation. There is very little scope for independent thinking and decision making on the part of the students. We report here our experience of teaching a practical course on sampling techniques in a different way. On the whole, it was an encouraging exercise.
1 Reorganizing the contents of a statistics course or changing the methods of teaching a course is a continuous process. Fesco et al. (1996) discuss why the typical course on survey sampling should be restructured and how it can be done to suit the needs of the society. Their discussion concerns a heterogeneous group of students. In a university setting, students in a master's degree course may be a homogeneous group. This article discusses how the current syllabus can be taught using a different approach to provide students with practical experience, helping the material to come alive.
2 In Indian Universities, statistics programs include distribution theory, statistical inference, design of experiments, and sampling theory of surveys as the core courses. The classroom teaching of these courses mostly involves derivation of formulae and proofs of theorems, while numerical examples, if any, are on a Lilliputian scale. The scale is small because large datasets cannot be handled in classrooms, and, until recently, computer facilities were very limited.
3 So-called 'practical' courses in statistics are expected to play a role complementary to the theory classes. The goals of an ideal practical course can be set out as follows:
4 At present, probably not many of us go beyond the first two goals listed above. Unfortunately, there is very little scope in the assignments related to goals (1) and (2) for students to think independently. As a result, our practicals become boring and uninteresting. Students just end up plugging numbers into given formulae and writing reject/accept H0 as their 'conclusion' without bothering to think about what H0 is! Hence, alternative ways of conducting the practical courses must be developed. We report here our experience with one such attempt.
5 Any move towards changing this situation requires strong motivation on the part of teachers, reasonable flexibility in the student evaluation system, support from colleagues, and, most importantly, students' cooperation and enthusiasm. We present here our ideas, difficulties encountered by the students, and their responses to a 'practical' on sample surveys. Our M.Sc. program consists of 20 one-semester courses. This study forms one-half of a semester course.
6 Considering the relative weight of this course in the whole program and the time available for this study, a real life problem involving field work was not feasible. Hence a pseudo real life situation was considered.
7 In a routine practical course on sample surveys, we assign a different set of questions for each session devoted to a topic, e.g., cluster sampling or simple random sampling without replacement (SRSWOR). Each session is two and a half hours long. The total time spent on the exercise described below was 8 to 10 sessions of two and half hours each over a period of five weeks. In our changed format, we set a common goal for the entire topic of sample surveys. The goal was estimating the total number of words in the Concise Oxford Dictionary (9th edition).
8 The broad objective to be achieved was as follows: Empirical comparisons of different sampling strategies are to be made, and the sampling variance associated with each strategy is to be studied. Standard textbooks on sampling theory like Murthy (1967), Cochran (1977), and Sukhatme, Sukhatme, Sukhatme, and Asok 1984) discuss situations under which one strategy should do better than another. The exercise gives the students a chance to empirically verify these comments. The question set for this exercise is given as Appendix A.
9 Each student counted the number of words on 50 pages for each of seven sampling strategies. In addition, for the ratio and regression methods of estimation, the student had to record an auxiliary variable. Hence, a student counted words on 450 pages, a substantial amount of manual work. Thus the students generated data instead of passively receiving it. This gave them an opportunity to appreciate difficulties in data generation.
10 For each sample drawn, the student calculated the mean number of words per page and then estimated the population total and its standard error. The formulae were not provided ready-made. Instead, all prescribed textbooks were made available, and students were asked to find and use the appropriate formulae. This forced them to browse through the books and hence increased the group discussions about notational differences. Further, each student had to draw a different sample. This helped in overcoming one more difficulty. In our usual practical setup where each student has the same question and the same dataset, students are tempted to copy the answers. This was not possible in the modified setup.
11 The 'sampling unit' was defined to be one page of the dictionary. Thus pages 1 to 1632 made up the 'sampling frame.' The characteristic under study was a 'word.'
12 The first difficulty we encountered was the definition of 'word.' Various possibilities include the following:
13 It was not obvious which definition one should use. Hence a pilot survey of ten pages was done using the first definition, which led to an estimate of 70 thousand words in the dictionary, approximately half the value claimed. When all bold-type words were counted, the estimate rose to 110 thousand. Hence, the second definition was thought to be more appropriate and was adopted for further work. The pilot study also suggested that there could be some relationship between number of head-words and number of bold-type words on a page. This information was useful in illustrating ratio and regression methods of estimation.
14 The following seven sampling strategies were used in the exercise.
Table 1. Stratum Size and Sample Size for the Five Smallest Strata for Stratified Sampling with a Total Sample Size of 50
|Stratum||Stratum size (Ni)||Sample size (ni)|
Hence, these letters were combined to form one stratum of size 35 with ni = 1, giving a total of 22 strata. Table 2 lists the strata, their sizes, and the corresponding sample sizes. The column of sample sizes contains some entries that are fractional numbers. We can solve this problem by treating each of the two columns on a page as an individual unit. We note that in most cases the observation (i.e., count of words) is recorded only for the whole page. The count is recorded for half pages only when the sample size is fractional. If ni = 1, we cannot estimate the within stratum variation.
The rounding of the sample size was done using the following ad hoc rule. Let be the fractional sample size calculated. Then, if the fractional part of ui is between 0 and 0.25, it is rounded to 0; between 0.25 and 0.75, to .5; and between 0.75 and 1.00, to 1. Using this rule, the estimate of the population total and its standard error required a modification that is discussed in Appendix B.
Table 2. Stratum Sizes and Sample Sizes for Stratified Sampling with a Total Sample Size of 50
|Stratum||Stratum size (Ni)||Sample size (ni)|
|J, Q, X, Y, Z||35||1|
15 For the SRSWOR and systematic sampling strategies and the ratio and regression methods of estimation, standard formulae were available in the textbooks. Since fractional sample sizes were used in stratified sampling, the estimator of the population total and its standard error had to be modified suitably. Such small modifications are inevitable when one drifts away from standard problems. Because the derivation of that formula is not the main goal of this article, we give the details in Appendix B.
16 For cluster sampling, multiple clusters were observed, and an ANOVA table could be prepared to obtain estimates of within and between cluster variation. These were then used to estimate the standard error.
17 For two-stage sampling, the formula for the variance of the estimator was not available because the first stage units were drawn by the Midzuno scheme and the second stage units by SRSWOR. Students therefore recognized the merit of adopting an alternative two-stage sampling scheme.
18 The discussion of results below emphasizes the learning process, rather than just reporting the outcome of the sample survey. All students showed interest in the whole exercise. They came up with interesting questions -- statistical and otherwise -- which does not usually happen in our courses. No textbook solutions were available. An example of a question asked was: "On the first page of the dictionary 'A' appears as a bold-type word 16 times. Why should these be treated as 16 separate words?"
19 Such difficulties were solved by group discussions. Hence it became an interactive process. As teachers, we were looking at the study as a whole. This led to some interesting observations that were also discussed in the group.
20 As this was our first experience with the exercise, our involvement was very high. From results submitted by the students, we did a sample check of two pages per strategy per student. Thus we ended up counting words on 280 pages. In the case of the ratio and regression methods, the population total of X (head-words) was needed, which we estimated by counting head-words on about 500 pages. Thus the physical work for the teachers was considerable. But this involvement enabled informed interaction with students regarding the data collection. Based on all the results, students gave a group seminar at the end of the course in which findings based on calculations done by the teachers were also discussed.
21 Table 3 shows a typical summary prepared by a student. In this particular case, cluster sampling had the smallest standard error, and systematic the largest. The others -- namely, SRSWOR, stratified, ratio, and regression -- appear fairly equivalent in terms of empirical precision.
Table 3. Number of Words (× 10-5) Estimated by Different Sampling Schemes and Estimated Standard Errors (× 10-5)
|Sampling scheme||s.e.||95% CI|
* Standard error and hence CI could not be obtained.
22 Table 4 gives the estimates for the various sampling schemes for all students. From Table 4 the following points can be noted.
Table 4. Estimates of Total Number of Words (× 10-5) in the Concise Oxford Dictionary
23 To check if the definition of 'word' was the cause of the underestimation, the following exercise was done. This also gave an opportunity to discuss ratio and regression methods. A random sample of 160 pages was drawn by SRSWOR (roughly a 10% sample). Each student counted the number of bold-type words (yi) and number of 'meanings' (wi) on 8 pages. After pooling the data for 160 pages, a correction factor was estimated:
All estimates were corrected using this factor to get an estimate of the number of 'meanings.'
24 Averages of these corrected estimates for 20 samples are given in the last row of Table 4. We see that these values (except for systematic and two-stage sampling) are very close to the claim in the dictionary, which means that the third definition of 'word' would have been more appropriate. This helped students realize, somewhat dramatically, how a definitional issue may be far more important than any amount of fine tuning of statistical procedures. Details of how the correction factor may affect the standard error were ignored.
25 In a typical course on sampling theory, we discuss the number of distinct sampling units in a sample drawn with replacement. In the present scheme, individual students sampled without replacement. However, it is possible that the same page could be selected by several students. Hence we thought it would be interesting to ask, "How many distinct pages did the students sample?" We checked this only for the SRSWOR scheme.
26 All students together sampled 20 × 50 = 1000 pages. Of these, 717 were found to be distinct. With n = 50 and N = 1632, the expected number of distinct pages can be obtained as follows:
Probability that a given page will not be selected in a sample of size n is .
Probability that a given page will not be selected in any of the 20 samples is .
Expected number of pages not included in any of 20 samples is .
Expected number of pages included in at least one sample is
Variance of number of pages included in at least one sample is
27 Hence we can test the hypothesis of random selection of pages, using the z test. We obtain . This shows that the observed value is too low. On careful scrutiny, it was suspected that one student might have borrowed observations from other students and put them together. On his own, he counted very few pages.
28 After deleting data from his sample, the observed number of distinct pages for 19 samples turned out to be 705. The expected value 728 and variance 102.94 can be calculated by replacing 20 by 19 in (1) and (2), giving z = -2.27, which is again low. So we still suspect that some students did not choose samples randomly, but rather borrowed some observations from others.
29 Using the distinct pages, the following summary was prepared.
We wish to express our sincere thanks to Drs. A. P. Gore, A. V. Kharshikar, and M. B. Rajarshi for help and encouragement throughout the work. We also thank the referees for their valuable comments which helped in improving this material.
Aim: To estimate the number of words in the dictionary. We will study the following aspects of sampling theory:
Observations and calculations: Use each sampling strategy below to draw 50 random integers between 1 and 1632; these are the page numbers to be sampled. Count the number of words in bold type on each page.
Carry out (a) to (f) for each of the following sampling strategies:
The formulae derived below apply to a single sample of
fractional sample size chosen from the
Let Y1, ..., Yn, Y(h)n+1 be a sample with fractional sample size, i.e., Y1, ..., Yn are observations on n full units, and Y(h)n+1 denotes an observation on a half unit, i.e., one of the two columns on the (n+1)th page. Let Z1 and Z2 represent observations on the two columns of the first unit, Z3 and Z4 observations on the two columns of the second unit, etc. Then, Z2m-1 and Z2m are observations on the two columns of the nth unit, and Z2m+1 = Y(h)n+1. We do not observe Z1 and Z2 separately, but rather Y1 = Z1 + Z2, Y2 = Z3 + Z4, etc.
Assume Z1, Z2, ..., Z2m+1 are a random sample from a normal distribution with mean and variance . Then the population total is given by , where N = 1632. Consider
To obtain an estimator of var, consider
Cochran, W. G. (1977), Sampling Techniques (3rd ed.), New York: John Wiley and Sons.
Fesco, R. S., Kalsbeek, W. D., Lohr, S. L., Scheaffer, R. L., Scheuren, F. J., and Stasny, E. A. (1996), "Teaching Survey Sampling," The American Statistician, 50, 328-340.
Murthy, M. N. (1967), Sampling Theory and Methods, Calcutta: Statistical Publishing Society.
Sukhatme, P. V., Sukhatme, B. V., Sukhatme, S., and Asok C. (1984), Sampling Theory of Surveys with Applications, Iowa State University Press and Indian Society of Agricultural Statistics.
S. A. Paranjpe
Anita Shah email@example.com
Department of Statistics
The University of Pune
Pune - 411007
JSE Homepage | Subscription Information | Current Issue | JSE Archive (1993-1998) | Data Archive | Index | Search JSE | JSE Information Service | Editorial Board | Information for Authors | Contact JSE | ASA Publications