RISK OF DENGUE HAEMORRHAGIC FEVERIN BEKASI MUNICIPALITY WITH SMALL AREA APPROACH1
Kismiant ini
Depart ment of Mat hemat ics, Yogyakart a St at e Universit y Anang Kurnia and Khairil A. Not odiput ro
Depart ment of St at ist ics, Bogor Agricult ural Inst it ut e Abst r act
Smal l Ar ea Est i mat i on (SAE) i s a st at i st i cal t echni que t o est i mat e par amet er s of subpopul at i on cont ai ni ng smal l si ze of sampl es. For count dat a, poi sson-gamma model s can be used t o est i mat e par amet er s of t he smal l ar ea. Thi s paper di scusses Bayes est i mat or s and empi r i cal Bayes est i mat or s i n poi sson-gamma model s i n smal l ar ea est i mat i on. We pr ovi de i l l ust r at i on of dengue haemor r hagi c f ever r i sk i n Bekasi Muni ci pal i t y usi ng r eal dat a f r om “ Di nas Kesehat an” and PODES 2003.
Keywords: small area est imat ion, poisson-gamma model, Bayes est imat or, empirical Bayes est imat or.
1. INTRODUCTION
Small Area Est imat ion (SAE) is a st at ist ical t echnique t o est imat e paramet ers of subpopulat ion (domain) cont aining small size of samples. The t erm “ small area” is commonly used t o denot e a small geographical area such as a count y, a municipalit y or census division. Direct est imat ors, based only on t he domain-specif ic sample dat a, are t ypically used t o est imat e paramet ers f or large domains. However, sample sizes in small domains, part icularly small geographical areas, are rarely large enough t o provide accurat e direct est imat es f or specif ic small domains. Theref ore, it is necessary t o f ind indirect est imat ors t o increase t he ef f ect ive sample size and t hus decrease t he st andard error.
There are t hree met hods of est imat ion t o obt ain indirect est imat ions: synt het ic, composit e, and James-St ein met hods. A synt het ic est imat or can brief ly be described as f ollows: “ An unbiased est imat or is obt ained f rom a sample survey f or a large area; and is used t o derive est imat or f or a small area under assumpt ion t hat t he small areas have t he same charact erist ics as t he large area. ” In order t o balance t he pot ent ial bias of a synt het ic est imat or against t he inst abilit y of a direct est imat or, a weight ed average of t he t wo est imat ors need t o be int roduced. This is called composit e est imat ion. James-St ein est imat ion is a composit e approach using common weight .
Indirect est imat ors f rom synt het ic or composit e or James-St ein est imat ion are based on implicit model s. Indirect est imat ors on t he ot her hand are based on explicit models incorporat ing area-specif ic ef f ect s. This includes empirical best linear unbiased predict ion (EBLUP), empirical Bayes (EB), and hierarchical Bayes (HB) est imat ors.
EBLUP met hods are suit able f or cont inuous variables and t hese are not suit able f or binary or count dat a. EB and HB met hods are applicable bot h f or binary and count dat a.
This paper describes a model appropriat e f or analysis of count dat a using small area approach. Bayes est imat or and empirical Bayes (EB) est imat or consider t he est imat ion of small area paramet ers. We use risk of dengue haemorrhagic f ever in Bekasi Municipalit y t o illust rat e t he met hod.
1
Unisba-2. MODEL-BASED INFERENCE FOR COUNT DATA
Empirical Bayes (EB) met hod is more generally valid f or model-based small area est imat ion, especially in handling count dat a.
2. 1 Poisson-Gamma Model
The Poisson model is st andard probabilit y model f or count s
y
i such as number of risk of dengue haemorrhagic f ever. The Poisson will get it s limit in mean and variance when it is used f or est imat ion of single paramet er. Commonly, count (dengue haemorrhagic f ever) dat a exhibit over dispersion (variance is larger t han t he mean) and t hus a more f lexible Poisson f ormulat ion would include an addit ional paramet er t o accommodat e t he ext ra variabilit y observed in t he sample. Based on above explanat ion, t wo-st age model will int roduce f or count dat a, known as a Poisson-Gamma model:( )
e
i
m
2. 3 Empirical Bayes (EB) Est imat or
In t he EB approach may be summarized as f ollows: (i) Obt ain t he post erior densit y of t he small area paramet ers of int erest , (ii) Est imat e t he model paramet ers f rom t he marginal densit y, (iii) Use t he est imat ed post erior densit y f or making inf erences.
Marshall (1991) in Rao (2003) obt ained simple moment est imat ors by equat ing t he weight ed sample mean
where
e
⋅=
∑
i(
e
im
)
and
3. AN ILLUSTRATION USING REAL DATA FROM “DINAS KESEHATAN” AND PODES 2003 We are illust rat ing t he poisson-gamma model in small area est imat ion f or risk of dengue haemorrhagic f ever in Bekasi Municipalit y using real dat a f rom “ Dinas Kesehat an” and PODES 2003.
Bekasi Municipalit y is divided int o m non-overlapping small areas. A small area in here means village, so we get 52 villages. Then, let yi is number of risk of dengue
Table one shows est imat ion f or risk of dengue haemorrhagic f ever in villages in Bekasi Municipalit y based on “Di nas Kesehat an” and PODES dat a in 2003. Generally, t he est imat or f rom Empirical Bayes gives result s of Mean Square Error (MSE) t ends t o small values compare wit h t he ot her est imat or. This can give descript ion of a bet t er accurat eness rat e f or Empirical Bayes est imat or. We can t ake anot her int erpret at ion t hat Bekasi Timur, Bekasi Selat an, and Bekasi Barat sub Municipalit ies have a relat ively high risk of dengue haemorrhagic f ever.
Table 1. The est imat ion f or risk of dengue haemorrhagic f ever based on direct , Bayes, and empirical Bayes est imat or.
4. CONCLUSION
Empirical Bayes met hod is used t o est imat e risk of dengue haemorrhagic f ever, which have a bet t er accuracy rat e wit h relat ively smaller MSE when compare wit h direct or Bayes met hods in t his paper.
An illust rat ion in t his paper used poisson-gamma model wit hout covariat es. The model development include covariat es which relat ed wit h response can be done t o improve est imat ion.
5. REFERENCES
Fay, R. E. and Herriot , R. A. (1979). Est imat es of income f or small places: an applicat ion of James-St ein procedures t o census dat a. Jour nal of t he Amer i can St at i st i cal Associ at i on, Vol. 74, p. 269-277.
Kurnia, A. and Not odiput ro, K. A. (2005). Aplikasi Met ode Bayes Pada Smal l Ar ea Est i mat i on. This paper present ed in St at ist ics Nat ional Seminar 7t h at Surabaya Technology Inst it ut e. [ November 26, 2005] .
Ramsini, B. et . all. (2001). Unisured est imat es by count y: a review of opt ions and issues. <www. odh. ohio. gov/ Dat a/ OFHSurv/ of hsrf q7. pdf >. [ August 23, 2005] .
Rao, J. N. K. and Gosh, M. (1994). Small area est imat ion: an appraisal. St at i st i cal Sci ence, Vol. 9, No. 1, pp. 55-76.
Rao, J. N. K. (1999). Some recent advances in model-based small area est imat ion. Sur vey Met hodol ogy, Vol. 25, No. 2, pp. 175-186.