Complex Surveys Tuesday, June 09, 2015. Assembling Design Components Building blocks –Probability sampling Simple random sampling (SRS) Unequal probability.

Complex Surveys

Tuesday, April 18, 2023

Assembling Design Components

• Building blocks– Probability sampling

• Simple random sampling (SRS)• Unequal probability sampling

– Stratification• Purpose: to increase the precision of estimates by grouping

similar items together

– Cluster sampling • Purpose: convenience.

– Ratio estimation

Simple Random Sampling(without replacement)

• There are (N choose n) possible samples

• Each with probability 1/(N choose n)

• Point estimate and C.I.

• Sample size (n) calculation

Si

iSi

i yn

NyNty

ny ˆ,

1

n

S

N

nNyVNtV

n

S

N

nyV

222

2

)1()ˆ()ˆ(,)1()ˆ(

Stratification (Ch 3)

• The estimate of the population total• It’s variance• Sample size allocation - proportional,

optimal• In general, stratified sampling with

proportional allocation is more efficient than SRS

• The more unequal the stratum means, the more benefits

H

hhtt

1

ˆˆ

H

hhtVtV

1

)ˆ()ˆ(

Ratio estimation (ch4)

• Biased

• May results in smaller MSE

• Useful when variables are linearly correlated

• Regression estimation

x

y

t

tB

ˆ

ˆˆ )ˆ,ˆ(voc

ˆ2)ˆ(ˆ1

)ˆ(ˆˆ

)ˆ(ˆ222

2

yxx

yx

xx

ttt

BtV

ttV

t

BBV

Cluster Sampling (ch 5&6)

• Usually less efficient than other methods• The relative efficiency of it and SRS depends on

intra-class correlation coefficient– The larger the correlation coefficient, the less efficient

• Can reduce cost and lead to administrative convenience

• One-stage, two-stage, with equal or unequal probs, point estimate, variance, c.i. – Allocation of m and n for two-stage cluster sampling

Cluster Sampling without Replacement

• Select a sample of n clusters with replacement based on

• Estimate cluster total and variance

• Estimate population total

• Variance can be estimated by formulas in ch5,6 or resampling methods

it

i

)ˆ(ˆitV

Si i

iHZ

tt

ˆ

ˆ

Cluster Sampling with Replacement

• Select a sample of n clusters with replacement based on

• Estimate cluster total

• Calculate

• Estimate population total and variance

i

it

iii tu /ˆ

nsun

u u

n

i i /,1 2

1

In practice

• Most of large surveys involves – several ideas of techniques– Different types of estimators

An example: background

• Malaria is a common public health problem in tropical and subtropical regions

• It is infectious. People get it by being bitten by a kind of female mosquito

• Without timely and proper treatment, the death rate can be very high

• Can be prevented by using mosquito nets• The prevention is only affective if the nets

are in widespread use

Summary

• Goal: To estimate the prevalence of bed net use in rural areas

• Sampling frame: all rural villages of <3,000 people in The Gambia

The survey in Gambia (1991) 3000 rural villages

Stage Sampling unit

eastern central western

1 district

Sampling method

•Stratified by region•Prob district size•5 districts per region

PHC Non-PHC

2 village

•Stratified by PHC•Prob village size•4 villages per district

3 compound•SRS•6 compounds / village

Top-down

The survey in Gambia (1991)3000 rural villages

eastern central western

PHC Non-PHC

......)ˆ(ˆ,ˆˆˆ,, districtPHCNondistrictPHCdistrictdistrict tVttt

compoundy

)ˆ(ˆ,ˆregionregion tVt

Top-downBottom-up

n

s

N

nNtVyNt villagecompoundvillage

22 )1()ˆ(ˆ,ˆ

)ˆ(ˆ)ˆ(ˆ)ˆ(ˆ)ˆ(ˆ

ˆˆˆˆ

westerncentraleastern

westerncentraleastern

tVtVtVtV

tttt

SRS

Stratified sampling

district

village

compound

)ˆ(ˆ),ˆ(ˆ,ˆ,ˆ,,,, PHCNondistrictPHCdistrictPHCNondistrictPHCdistrict tVtVtt

Stratified sampling

Sampling with unequal probs,two-stage cluster, Ch 6

Sampling with unequal probs,two-stage cluster, Ch 6

(average number of nets per compound)

The survey in Gambia (1991)

• The way to calculate the estimated total and its variance seems to be complicated– It can be worse if we include ratio estimators

• In practice, we can – Use sampling weights to obtain point

estimates– Use computer intensive methods to obtain

standard error (ch9)• Such as jackknife, bootstrap

Sampling weights• The sampling weight is the reciprocal of

Pr(being selected)

• Each sampled unit “represents” certain number of units in the population

• The whole sample “represents” the whole population

Sampling weights• Weights are used to deal with the effects of

stratification and clustering on point estimate

• Stratified sampling

NNwH

hh

H

h

n

jhj

h

11 1

Sampling weights• Cluster sampling with equal probabilities

Si

iSi Sj

ij MNMn

Nw

i

ijiij

ijiij

www |

|

Sampling weights

pstpsp wwww ,|| p: primarys: secondary t: tertiary

• For three-stage sampling

• Very large weights are often truncated– Biases results– Reduces the mean squared error

Sampling weights• Weights contain the information needed to

construct point estimates • Weights do not contain enough information for

computing variance– Weights can be used to find point estimates

because calculating variance requires prob(pairs of units are selected)

– Computer-intensive methods can be used to find variances

Sampling weights: the malaria example

Pr(a compound in central region PHC villages is selected)=

Self-weighting and Non-self-weighting

• Self-weighting: sampling weights for all observation units are equal

• A self-weighing sampling is representative of the population if nonsampling errors are ignored

• Most large self-weighting samples are not SRS• Standard software with the usual assumption of

iid leads to correct estimate of mean, proportion, percentiles; but erroneous estimation for variance

Ratio Estimation in Complex Surveys

• Ratio estimation is part of the analysis, not the design

• Can be used at any level. Usually used near the top

Ratio Estimation in the Malaria example

• Region level:

• Above the region level

Ratio Estimation in Complex Surveys

• The bias of ratio estimation can be large when sample sizes are small

• Separate ratio estimator for a population total

– Improves efficiency when ratios vary from stratum to stratum; works poorly for small strata sample sizes

• Combined ratio estimator for a population total

– Has less bias when strata sizes are small; works poorly when ratios vary from stratum to stratum

XYX ttt ˆ/ˆ

Estimating a Distribution Function

• Historically, sampling theory was developed to find population means, totals, and ratios.

• Other quantities, such ass,– Pr(Statistics > means or totals)– Median? 95th percentile? Probability mass

function?

• Sampling weights can be used in constructing an empirical distribution of the population

Population quantities and functions

• Probability mass function (pmf)

• Distribution functionN

yyf

is value whoseunits ofnumber )(

yx

xfN

yyF )(

value whoseunits ofnumber )(

)()( yYPyf

}{)( yYPyF

1)()( Fyf

)(yyfyU

]))(()([1

))()((1

)(1

1

22

2

1

22

yy

y xi

N

iUi

yyfyfyN

N

xxfyyfN

N

yyN

S

Empirical Functions

• Empirical probability mass function

• Empirical distribution function

• Empirical functions can be used to estimate population quantities such as mean, median, percentiles, variance, ect.

yyi

yySii

i

i

w

w

yf :)(ˆ

yx

xfyF )(ˆ)(ˆ

Plotting data from a complex survey

• SRS– Histograms/smoothed density estimates– Scatterplots and scatterplot matrices

• In a complex sampling design, simple plots can be missleading

Incorporating weights

Incorporating weights

Plotting data from a complex survey

• The 1987 Survey of Youth in Custody– Family background, previous criminal history, drug

and alcohol use– 206 PSU’s (facilities) were divided by 16 strata– SSUs were the 1985 Children in Custody (CIC)

The 1987 Survey of Youth in Custody

•The two figures are very similar because the survey was aimed to be self-weighting•Youths aged 15 were undersampled due to unequal selection prob and nonresponse•Youths aged 17 were oversampled

The 1987 Survey of Youth in Custody

The 1987 Survey of Youth in Custody

The 1987 Survey of Youth in Custody

The 1987 Survey of Youth in Custody

Design effects

• Cornfield’s ratio (1951) – Measure the efficiency of a sampling plan by

the ratio of the variance that would be obtained from an SRS of k observation units to the variance obtained from the complex sampling plan with k observation units

• The design effect (deff, Kish 1965)– The reciprocal of Cornfield’s ratio

The design effects

• The design effect provides a measure of the precision gained/lost by use of the more complex design instead of SRS

• For estimating a mean

The design effects

• Stratified

• Cluster