THE SHORTEST CONFIDENCE INTERVAL FOR PROPORTION IN FINITE POPULATIONS

Transcription

1 APPLICATIOES MATHEMATICAE Online First version Wojciech Zieliński (Warszawa THE SHORTEST COFIDECE ITERVAL FOR PROPORTIO I FIITE POPULATIOS Abstract. Consider a finite population. Let θ (0, 1 denote the proportion of units with a given property. The problem is to estimate θ on the basis of a sample drawn according to simple rom sampling without replacement. We are interested in interval estimation of θ. We construct the shortest confidence interval at a given confidence level. Introduction. The problem of the interval estimation of the proportion (fraction θ is very old. The first solution was given by Clopper Pearson (1934 since then many authors have dealt with the problem. Zieliński (2010a, 2012, 2016 considered the problem of constructing the shortest confidence intervals for probability of success in a binomial as well as in a negative binomial model. The solutions are valid for infinite populations. In many applications (economic, social, etc. we deal with a finite population, so we are interested in interval estimation of θ in that case. Remarks on differences in statistical inference in infinite finite populations may be found for example in Yates (1960 or Cochran (1977. The problem of interval estimation of proportion in finite populations was rather rarely considered in literature. This is because the construction of the exact confidence interval is based on the hypergeometric distribution. Different approximations may be found in Buonaccorsi (1987, Wendell Schmee (2001 or Wright (1991. There are at least two approximations commonly used in applications: binomial normal. However, using those approximations may lead to wrong conclusions (Zieliński In what follows, the exact (hypergeometric distribution will be used Mathematics Subject Classification: Primary 62F25; Secondary 62D99. Key words phrases: proportion, confidence interval, shortest confidence intervals, finite population. Received 22 February 2016; revised 30 June Published online *. DOI: /am [1] c Instytut Matematyczny PA, 2016

2 2 W. Zieliński Confidence interval. Consider a population U = {u 1,..., u } containing the finite number of units. Let M denote the unknown number of objects in the population which have an interesting property. We are interested in interval estimation of M, or equivalently, of the fraction θ = M/. A sample of size n is drawn according to simple rom sampling without replacement. Let ξ be the rom variable describing the number of objects with the given property in the sample. On the basis of ξ we want to construct a confidence interval for θ at a confidence level δ. The rom variable ξ has the hypergeometric distribution (Johnson Kotz 1969, Zieliński 2010b P θ,,n {ξ = x} = ( θ ( (1 θ x n x ( n, for integers x from the interval [max{0, n (1 θ}, min{n, θ}]. Let f θ,,n ( be the probability distribution function, i.e. f θ,,n (x { Pθ,,n {ξ = x} for integer x [max{0, n (1 θ}, min{n, θ}], = 0 elsewhere, let F θ,,n (x = t x f θ,,n (t be the cumulative distribution function (CDF of ξ. Then Γ (θ + 1Γ ((1 θ + 1 F θ,,n (x = n! ( n ( n 1 t Γ (θ t + 1Γ ((1 θ n + t + 1 t x where = 1 ( n x+1 ( n θ x 1 ( θ 3F 2 [ {1, x + 1 θ, x + 1 n}, {x + 2, (1 θ + x + 2 n}; 1 ], 3F 2 [{a 1, a 2, a 3 }, {b 1, b 2 }; t] = k=0 (a 1 k (a 2 k (a 3 k t k (b 1 k (b 2 k k! (a k = a(a + 1 (a + k 1. The construction of the confidence interval at a confidence level δ for θ is based on this cumulative distribution function. Let θ(x,, n, α be the solution of the equation F θ,,n (x = α with respect to θ. If ξ = x is observed

3 The shortest confidence interval for proportion 3 then the confidence interval has the form ( (θ L (x 1,, n, δ 1, θ U (x,, n, δ 2. The numbers δ 1 δ 2 are such that δ 1 δ 2 = δ. For ξ = 0 the left end is taken to be 0, for ξ = n the right end is taken to be 1. The analytic solution is unavailable. However, for given x, n, the confidence interval may be found numerically. To be more precise, the confidence interval for θ is in fact a confidence set of the numbers of the form k/ between θ L (x 1,, n, δ 1 / θ U (x,, n, δ / (here t denotes the greatest integer not greater than t. In what follows, we will make use of (. In the stard construction δ 1 = (1 + δ/2 is used. It is of interest to find the shortest confidence interval. Therefore, we want to find δ 1 δ 2 such that the confidence interval is the shortest possible. The shortest confidence interval. Consider the length of the confidence interval when ξ = x is observed: d(δ 1, x = θ U (x,, n, δ 1 θ L (x 1,, n, δ + δ 1. In what follows we consider only the case x n/2. If x n/2, the role of success failure should be interchanged. Let us state the following theorem. Theorem 1. For x > 1 there exists a two-sided shortest confidence interval. For x = 1 the shortest confidence interval is one-sided. To prove the theorem we will prove the following lemma. Define G x,,n (θ = F θ,,n (x. Of course, the domain of this function is [x/, 1 (n x/]. Lemma. For x 1 there exists θ such that the function G x,,n ( is concave for θ < θ, convex for θ > θ. Proof. Consider the second order differences of the function G x,,n (. Some calculations give: H x,,n (θ = Because (G x,,n ( θ + 1 = 1 ( n ( G x,,n (θ (G x,,n (θ G x,,n θ 1 x ( + 1t(t 1 (n 1(2t( + 1 nθ + n(n 1θ 2 (1 θθ(θ + 1 t((1 θ + 1 (n t ( ( θ (1 θ t n t t=0 H x,,n ( x = n!( x 1!!(n x 1! < 0.

4 4 W. Zieliński ( H x,,n 1 n x n!( n + x 1! = ((n x + 1 (n 1(n x > 0,!x! the equation H x,,n (θ = 0 has a solution. The hypergeometric distributions are stochastically ordered, i.e. for θ 1 < θ 2 for any given x we have F θ1,,n(x > F θ2,,n(x. Hence, the equation H x,,n (θ = 0 has only one solution θ. We have { < 0 for θ < θ, H x,,n (θ > 0 for θ > θ. Hence G x,,n ( is concave for θ < θ convex for θ > θ. Proof of Theorem 1. We have to show that for x > 1 there exists 0 < δ 1 < 1 δ such that d(δ 1, x is minimal. The length of the confidence interval equals d(δ 1, x = θ U (x,, n, δ 1 θ L (x 1,, n, δ 1 + δ, where θ U (x,, n, δ 1 is a solution of F θu,,n(x = δ 1 θ L (x 1,, n, δ 1 +δ is a solution of F θl,,n(x 1 = δ 1 + δ. By Lemma, for x > 1 the function (0, 1 δ δ 1 θ U (x,, n, δ 1 is convex (0, 1 δ δ 1 θ L (x 1,, n, δ 1 + δ is concave. Hence, there exists δ1 such that d(δ 1, x = inf{d(δ 1, x : δ < δ 1 < 1}. Consider now the case x = 1. The lower bound of the confidence interval is a solution of F θl,,n(0 = δ 1 + δ. It is easy to check that H 0,,n (θ > 0 for all θ [1/, 1 (n 1/]. Hence the function (0, 1 δ δ 1 θ L (0,, n, δ 1 + δ is concave. The length of the confidence interval reaches its minimum at δ 1 = 1 δ, so the shortest confidence interval for x = 1 is one-sided. It is interesting to note that if n is even x = n/2 then the shortest confidence interval is the stard one. This follows from the fact that for θ < 0.5, G n/2,,n (θ = G n/2,,n (1 θ. In Tables 1 2 some numerical results are given. The stard confidence intervals (θl st, θst U the shortest confidence intervals (θsh L, θsh U for

5 The shortest confidence interval for proportion 5 x Table 1. The shortest confidence intervals (n = 10 θ st L θ st U % % % % % % % % % % % θ sh L θ sh U x Table 2. The shortest confidence intervals (n = 100 θ st L θ st U % % % % % % % % % % % % % % % % % % % θ sh L θ sh U = 1000 δ = 0.95 are shown. In the last column the ratio of the length of the shortest to the length of the stard confidence interval is calculated. In Figure 1 the confidence level of the shortest confidence interval is shown ( = 1000, n = 100, δ = ote that, for some θ, the confidence level is smaller than the nominal one. This is in contradiction to the definition of the confidence interval (eyman 1934, Cramér 1946, Lehmann 1959, Silvey In what follows, a small modification is introduced, after which the shortest confidence interval does not have this disadvantage.

6 6 W. Zieliński Fig. 1. Confidence level of the shortest confidence interval Auxiliary variable. Let η be a rom variable conditionally distributed on the interval [0, 1] with CDF G η ξ=x (. The confidence interval will be constructed on the basis of ζ = ξ+η. The distribution of that r.v. is easy to obtain: P θ,,n {ζ t} 0 if t 0, where α( t, {t}p θ,,n {ξ = t } if t = 0, t 1 = P θ,,n {ξ = k} + α( t, {t}p θ,,n {ξ = t } if 1 t n, k=0 1 if t > n, {t} = t t α( t, {t} = {t} 0 G η ξ= t (du. It is easy to note that the distribution of η may be taken to be uniform U(0, 1 independently of ξ. Let Then G x,y,,n (θ = (1 yf θ,,n (x 1 + yf θ,,n (x. θ L (x, y,, n, δ 1 + δ = G 1 x,y,,n (δ 1 + δ θ U (x, y,, n, δ 1 = G 1 x+1,y,,n (δ 1

7 The shortest confidence interval for proportion 7 are the ends of the confidence interval for θ at the confidence level δ. The length of the confidence interval equals d(δ 1 ; x, y = G 1 x+1,y,,n (δ 1 G 1 x,y,,n (δ 1 + δ. Let ξ = x η = y be observed. We are looking for δ1 such that d(δ 1; x, y = min{d(δ 1 ; x, y : 0 δ 1 1 δ}. Theorem 2. If x 2 then the shortest confidence interval is two-sided. Proof. Consider the second order differences H x,y,,n (θ = (G x,y,,n ( θ + 1 = (1 yh x,,n (θ + yh x+1,,n (θ. ( G x,y,,n (θ (G x,y,,n (θ G x,y,,n θ 1 Analysis similar to the one in the proof of the Lemma shows that for x 2 y (0, 1, ( ( x H x,y,,n < 0 H x,y,,n 1 n x > 0. Hence there exists a two-sided shortest confidence interval. Theorem 3. For x = 1 there exists y such that the shortest confidence interval is one-sided if x + y < 1 + y, two-sided otherwise. Proof. For convenience assume that θ is a continuous parameter. The derivative of d(δ 1 ; x, y with respect to δ 1 equals where We have d(δ 1 ; x, y δ 1 = 1 LHS(δ 1 ; x, y 1 RHS(δ 1 ; x, y, LHS(δ 1 ; x, y = G x+1,y,,n(θ θ RHS(δ 1 ; x, y = G x,y,,n(θ θ θ=g 1 x+1,y,,n (δ 1 θ=g 1 x,y,,n (δ 1+δ G 1,y,,n (θ Γ ((1 θ + 1 = ( θ n Γ (n + 1Γ ((1 θ + 1 n [ny 1 ( + 1 n + θ(ny 1 ( H((1 θ H((1 θ n ].,

8 8 W. Zieliński G 2,y,,n (θ Γ ((1 θ + 1 = θ 2 ( n Γ (n + 1Γ ((1 θ + 3 n [2(n 1( + 2 n 2( + 1 n + n(n 1y ( (n 1nyθ(θ 1 + 2((1 θ + 2 n( + 1 n + (n 1θ (H((1 θ H((1 θ + 2 n ], where H(u = u i=1 1/i. If δ 1 0, then G 1 1,y,,n (δ 1 G 1 1,y,,n (0 = 1 G 1 2,y,,n (δ 1 + δ G 1 2,y,,n (δ > 0, for all y, RHS(δ 1 ; 1, y LHS(δ 1 ; 1, y LHS(0; 1, 0 < 0. Hence d(δ 1;1,y δ 1 < 0 as δ 1 0. If δ 1 1 δ, then G 1 1,y,,n (δ 1 G 1 1,y,,n (1 δ > 0 G 1 2,y,,n (δ 1 + δ G 1 2,y,,n (1 = 1/. Since RHS(1 δ 1 ; 1, 0 < 0 < LHS(1 δ 1 ; 1, 0 (so d(δ 1;1,1 δ 1 RHS(1 δ 1 ; 1, 1 < LHS(1 δ 1 ; 1, 1 < 0 (so d(δ 1;1,0 δ 1 δ1 =0 δ1 =0 > 0 < 0, there exists y such that LHS(1 δ 1 ; 1, y = RHS(1 δ 1 ; 1, y. So, the shortest confidence interval is one-sided for y < y, two-sided otherwise. In Table 3 the values of y for different population sizes sample sizes n (1%, 5% 10% of population size are given. Table 3. Values of y 1% 5% 10% In Figure 2 the confidence level of the romized shortest confidence interval for = 1000, n = 100 δ = 0.95 is shown. Final remarks. 1. The above considerations may be summarized as follows. From a finite population of size we draw a sample of size n (according to simple rom sampling without replacement. Observe the r.v. ξ with

9 The shortest confidence interval for proportion Fig. 2. Confidence level of the romized shortest confidence interval hypergeometric distribution with parameters θ,, n. Draw a r.v. η according to U(0, 1 distribution. For ξ n/2 consider ξ + η. The shortest confidence interval for θ is one-sided if ξ + η 1 + y, two-sided otherwise. For ξ > n/2 consider n ξ 1 η make use of the relationships θ L (ξ, η,, n, γ + θ U (n ξ 1, 1 η,, n, 2δ γ = 1 θ U (ξ, η,, n, γ + θ L (n ξ 1, 1 η,, n, δ + γ = 1. The generated value y of U(0, 1 r.v. must be attached to the final report. Hence, now results are given by four numbers: population size, sample size, number of successes the value y. 2. Because the population is finite, the set of admissible values of θ is also finite. For example, the stard confidence interval (for population of size 1000 when x = 2 is observed is [0.025, 0.555]. The parameter θ may take one of the values {0.025, 0.026,..., 0.554, 0.555} the number of units with a given property is one of {25, 26,..., 554, 555}. 3. We now give an example of application. Let the size of a population be = We took a sample of size n = 100 we observed ξ = 2 objects with a given property. Let the confidence level be δ = The drawn value of the auxiliary variable η is The shortest confidence interval is ( , Its length is

10 10 W. Zieliński The final report may look as follows: = 1000; n = 100, x = 2, y = , δ = 0.95, θ ( ; The stard confidence interval is ( , its length equals References J. P. Buonaccorsi (1987, ote on confidence intervals for proportions in finite populations, Amer. Statistician 41, C. J. Clopper E. S. Pearson (1934, The use of confidence or fiducial limits illustrated in the case of the binomial, Biometrika 26, W. G. Cochran (1977, Sampling Techniques, 3rd ed., Wiley. H. Cramér (1946, Mathematical Methods in Statistics, Princeton Univ. Press (19th printing, L. Johnson S. Kotz (1969, Distributions in Statistics: Discrete Distributions, Houghton Mifflin, Boston. E. L. Lehmann (1959, Testing Statistical Hypothesis, Springer (3rd ed., J. eyman (1934, On the two different aspects of the representative method: The method of stratified sampling the method of purposive selection, J. Roy. Statist. Soc. 97, S. D. Silvey (1970, Statistical Inference, Chapman & Hall (11th ed., J. P. Wendell J. Schmee (2001, Likelihood confidence intervals for proportions in finite populations, Amer. Statistician 55, T. Wright (1991, Exact Confidence Bounds when Sampling from Small Finite Universes; An Easy Reference Based on Hypergeometric Distribution, Lecture otes in Statist. 66, Springer, ew York. F. Yates (1960, Sampling Methods for Censuses Surveys, Hafner, ew York. W. Zieliński (2010a, The shortest Clopper Pearson confidence interval for binomial probability, Comm. Statist. Simulation Comput. 39, W. Zieliński (2010b, Estimation of Proportion, Wydawnictwo SGGW, Warszawa (in Polish. W. Zieliński (2011, Comparison of confidence intervals for fraction in finite populations, in: Quantitative Methods in Economics XII, SGGW, Warszawa, W. Zieliński (2012, The shortest confidence interval for probability of success in a negative binomial model, Appl. Math. (Warsaw 39, W. Zieliński (2014, The shortest romized confidence interval for probability of success in a negative binomial model, Appl. Math. (Warsaw 41, W. Zieliński (2016, The shortest Clopper Pearson romized confidence interval for binomial probability, REVSTAT-Statist. J., to appear.

11 The shortest confidence interval for proportion 11 Wojciech Zieliński Department of Econometrics Statistics Warsaw University of Life Sciences owoursynowska Warszawa, Pol