Skip to main content

• ## Summary

Search form
• 00:11

KEN COPELAND: Hi.I'm Ken Copeland, director of statisticsand methodology for NORC at the Universityof Chicago. [Kennon R. Copeland, PhD, Senior Vice President& Director, Statistics & Methodology,NORC at the University of Chicago]This tutorial provides an introductionto survey weighting.I'll be talking about five major points.First, providing an overview of survey weighting.Second, discussing survey process flow

• 00:34

KEN COPELAND [continued]: and its relationship to survey weighting.I'll review the information needs required for carrying outsurvey weighting.Walk through the weighting steps.And finally, talk about variance estimation.[Survey Weighting Overview]

• 00:56

KEN COPELAND [continued]: A way to think about survey weightsis that we have data from a sample of the population.And we need to derive survey weights thatallow us to calculate appropriate estimates,as well as make inferences for the total population

• 01:17

KEN COPELAND [continued]: of interest.First, we want to look at where does the estimationor weighting methodology fit within the overall surveydesign?The weighting methodology is informedby the survey requirements for the survey,as well as the sample design and data collection methodology.

• 01:39

KEN COPELAND [continued]: The estimation methodology then isused within the data processing to actually calculatesurvey weights, estimates, and variance estimates.The objectives of survey weightingare to carry out weighting so as to reflect the surveysampling, data collection, and processing flow,as well as to derive survey weights that

• 02:01

KEN COPELAND [continued]: yield reliable estimates for the population and any requiredsub-populations.[Survey Process Flow]So let's look at the survey process flow.Here's an illustration which shows the samplingframe from which the sample is selected,

• 02:24

KEN COPELAND [continued]: of which we have a total of capital N unitsfrom which we select little n units for our survey.Then we try to find those units and conduct an interview.In some cases, the units may not belocated or able to be contacted at all.

• 02:44

KEN COPELAND [continued]: For the others, they are located, and at that pointwe determine whether or not the unitis eligible for the sample.So following that step, we then havethe sample as those units were determined as not eligible,those that are determined as eligible,

• 03:05

KEN COPELAND [continued]: and those for which we weren't able to make a determination.For the eligible units, we then attemptto conduct an interview.And from that set, we have some units for whichthe interview is completed.And then we have the sample non-respondents.It is the units for which we are able to collect

• 03:26

KEN COPELAND [continued]: data and complete an interview for which survey weights willbe derived.So now we have superimposed the survey weightingsteps that correspond to this process flow.In selecting the sample, we will use that informationto derive the base weights.In determining eligibility of sample units,

• 03:48

KEN COPELAND [continued]: we would use that for a non-resolution adjustment.Then we have the survey completion, or response,from which we need to apply a nonresponse adjustment.And then finally, we want the survey weightsto correspond to the population total,so we'll do an adjustment to the population totals.

• 04:11

KEN COPELAND [continued]: [Information Needs for Weighting]In order to carry out the survey weightingwe need several pieces of information.First, we need the probabilities of selection,which are to include all stages of selectionwithin the sample design.

• 04:33

KEN COPELAND [continued]: Second, we need sample disposition codeswhich will indicate what the outcome wasfor that particular sample unit at each stage in the surveyprocess flow.For example, was the unit located or not?Was the unit determined to be eligible or not?Was the unit a completed interview?

• 04:54

KEN COPELAND [continued]: Was it a refusal?Other nonresponse?And then finally, we need population data,population counts that we'll use in our final stepsin weighting.Note that the population data isn't referringto the sampling frame but rather to the total populationof interest.

• 05:15

KEN COPELAND [continued]: [Weighting Steps]So now a walk through each of the weighting steps.First, we have the base weight.The base weight is the initial weightapplied to the sample unit that reflectsthe probability of selecting that sampleunit into the sample.

• 05:36

KEN COPELAND [continued]: A base weight is determined for every sampling unit thatwas included in the survey, whether or not that sampleunit was actually interviewed.Another step is the nonresponse adjustment.Not all units that are selected for the survey will respond.In that case, we need to adjust the survey weights

• 05:59

KEN COPELAND [continued]: for the respondents to account for the non-respondentsfor which we weren't able to collect the survey data.We must do this so as to try to minimizethe nonresponse bias associated with notobserving the non-respondents.The two primary approaches in doing this

• 06:19

KEN COPELAND [continued]: are to determine nonresponse adjustment cells whereinevery unit is classified on the basis of characteristicsthat are known for both the respondentsand the non-respondents.The second approach is to predict the response propensityfor each unit, both respondents and non-respondents.

• 06:41

KEN COPELAND [continued]: Within each of the categories thatare determined or within each of the nonresponse adjustmentcells, a nonresponse adjustment is calculated so asto weight up the respondents to represent the non-respondents.Note that this type of approach can alsobe used for adjusting for non-resolution of eligibility.

• 07:07

KEN COPELAND [continued]: The final step in our weighting processis to adjust our survey weights so that weighted counts agreewith known population controls.This is carried out after nonresponse adjustment.The nonresponse adjustment has weighted the survey upto represent the sample frame population

• 07:29

KEN COPELAND [continued]: but may not be representative of the total populationnor of required sub-populations.The adjustment to population controlsis carried out so as to reduce variance, control bias,as well as to ensure that the weighted counts agree

• 07:49

KEN COPELAND [continued]: with selected sub-population totals.Three primary approaches to population control adjustmentare post-stratification ratio adjustment,raking ratio adjustment, and more broadly, calibration.Within post-stratification ratio adjustment,

• 08:10

KEN COPELAND [continued]: the sample is classified within each cellof several characteristics, such as age and gender.And so each sample unit would be classified as to its ageand gender classification.Population controls would be available for each

• 08:31

KEN COPELAND [continued]: of those cells.And the sample weights adjusted upto reflect the total population within each cell.In some cases, the population controlsare not known for each cell, or the sample sizewithin each cell may be too small in orderto derive stable weights.

• 08:53

KEN COPELAND [continued]: In this case, raking ratio adjustment may be applied.In this instance, what happens isthat the weights are controlled to selected marginal totalsrather than to individual cells.One example would be to rake the weights so that they agree

• 09:13

KEN COPELAND [continued]: with marginals in terms of the total populationby age, total population by gender,and, separately, total population by race.The broad class of estimators which both of these fit inare termed calibration estimators.

• 09:33

KEN COPELAND [continued]: Calibration controls the survey weightsto known auxiliary totals, which may include population totalsbut may include other types of estimates.And they're carried out so as to minimize specified lossfunctions.These loss functions are typicallyspecified in terms of minimizing the change in weights

• 09:57

KEN COPELAND [continued]: between the nonresponse adjustment and the final sampleweight so as to control variance.[Variance Estimation]Now let's talk about variance estimation.The purpose of variance estimatesare to provide a measure of uncertainty that

• 10:17

KEN COPELAND [continued]: are associated with the estimates thatare due to our observing only a sample as opposedto the full population.They're typically reported as standard errorsor a coefficient of variation.The common variance estimation approaches.First, design-based methods.

• 10:38

KEN COPELAND [continued]: These methods utilize the sample designthat was applied for the survey and sample survey theoryto determine how to calculate the variance estimates.With complex, multi-stage sample designs,the design-based methods can become very complex.

• 11:00

KEN COPELAND [continued]: So there are other approaches thatcan be used to calculate unbiased estimatesof the variance.One type are replication methods,such as jackknife random groups and balanced half-samples.For some other estimates, such as ratio estimates,linearization methods are used.

• 11:21

KEN COPELAND [continued]: A ratio estimate may look like an unemploymentrate, vaccination rate for a population, and so forth.Linearization methods make use of Taylor series approximationsto derive a linear form for the variance estimate.[Conclusion]

• 11:42

KEN COPELAND [continued]: In summary, we need survey weightsto account for the sample design, data collection,and data processing, as well as to control weighted countsto known population totals in such a manneras to minimize the variance and control the bias,as well as to allow for publication of required survey

• 12:05

KEN COPELAND [continued]: estimates.Several references for further reading.Valliant, Dever, and Kreuter, PracticalTools for Designing and Weighting Survey Samples.Lohr, Sampling: Design and Methodology.Wolter, Introduction to Variance Estimation.And Cochran, Sampling Techniques.

### Video Info

Publisher: SAGE Publications Ltd

Publication Year: 2017

Video Type:Tutorial

### Segment Info

Segment Num.: 1

Persons Discussed:

Events Discussed:

Keywords:

## Abstract

Dr. Kennon R. Copeland discusses survey weighting and variance estimation. Survey weights allow researchers to calculate estimates and make inferences about a population using a sample population. Copeland discusses survey process flow, information needed for weighting, and weighting steps.

## Looks like you do not have access to this content.

An Introduction to Survey Weighting

Dr. Kennon R. Copeland discusses survey weighting and variance estimation. Survey weights allow researchers to calculate estimates and make inferences about a population using a sample population. Copeland discusses survey process flow, information needed for weighting, and weighting steps.

Copy and paste the following HTML into your website