survey experiment modalities

35
Survey Experiment Survey Experiment Modalities Modalities Some Pros and Cons of Differing Sampling Sources Some Pros and Cons of Differing Sampling Sources and Methods and Methods Matthew A. Baum Harvard University & Leonie Huddy Stonybrook University

Upload: katima

Post on 09-Jan-2016

18 views

Category:

Documents


1 download

DESCRIPTION

Survey Experiment Modalities. Some Pros and Cons of Differing Sampling Sources and Methods. Matthew A. Baum Harvard University & Leonie Huddy Stonybrook University. Outline. Review of prominent Internet players Amazon Mechanical Turk Knowledge Networks Yougov/Polimetrix - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Survey Experiment Modalities

Survey Experiment Survey Experiment ModalitiesModalities

Some Pros and Cons of Differing Sampling Sources and Some Pros and Cons of Differing Sampling Sources and MethodsMethods

Matthew A. BaumHarvard University

&Leonie Huddy

Stonybrook University

Page 2: Survey Experiment Modalities

OutlineOutlineReview of prominent Internet playersReview of prominent Internet players

Amazon Mechanical TurkAmazon Mechanical Turk

Knowledge NetworksKnowledge Networks

Yougov/PolimetrixYougov/Polimetrix

SurveyMonkey AudienceSurveyMonkey Audience

Comparing survey modalitiesComparing survey modalities

USA vs. Swedish Phone/Internet Usage USA vs. Swedish Phone/Internet Usage StatisticsStatistics

ConclusionsConclusions

Page 3: Survey Experiment Modalities

Amazon Mechanical Turk Amazon Mechanical Turk (AMT)(AMT)

OverviewOverviewBegan in 2005Began in 2005““Workers” sign up to participate in tasks for pay Workers” sign up to participate in tasks for pay (called “Human Intelligence Tasks”, or “HITs”)(called “Human Intelligence Tasks”, or “HITs”)

Global workforce (~100,000 workers)Global workforce (~100,000 workers)

Salary: ~$3-$8/hour (maybe $10/hr for high-experience Salary: ~$3-$8/hour (maybe $10/hr for high-experience workers)workers)

““Requestors recruit “workers” on internal message Requestors recruit “workers” on internal message board describing taskboard describing task

Can specify “qualifications” for workers (e.g., Can specify “qualifications” for workers (e.g., experience, quality ratings, nationality, etc.)experience, quality ratings, nationality, etc.)

10% fee to Amazon10% fee to Amazon

AMT provides survey building tools, or Requestors AMT provides survey building tools, or Requestors can include links to external surveyscan include links to external surveysHITs range from 1-second marketing surveys (“Rate HITs range from 1-second marketing surveys (“Rate appeal of this photo on 1-10 scale”) to elaborate 30+ appeal of this photo on 1-10 scale”) to elaborate 30+ minute experimental surveysminute experimental surveys

Page 4: Survey Experiment Modalities

Compensation Amount

Short survey(5 min)

Medium survey (10

min)

Long survey(30 min)

$.02 5.6 5.6 5.3

$.10 25.0 14.3 6.3

$.50 40.5 31.6 16.7

Effects of Compensation Amount and Task Length on Participation Rates(Submitted Surveys per Hour of Posting Time)

Source: Buhrmester, Kwang, and Gosling (2011)

My AMT Experience:•paid $1.00 per completed HIT for ~22 minute survey•Listed as “News and Politics Survey”•Received 1933 completed surveys in ~one month• rapidly diminishing returns• 1825 valid responses

Amazon Mechanical Turk Amazon Mechanical Turk (AMT)(AMT)

Page 5: Survey Experiment Modalities

% Female % Non-White

% Non-American

Age

AMT 55 36 31 32.8

Internet 57 Fewer (?) ~31 24.3

Buhrmester, Kwang, and Gosling (2011)

Amazon Mechanical Turk Amazon Mechanical Turk (AMT)(AMT)

DemographicsDemographics

Paolacci, Chandler, and Ipeirotis (2010)

Category (N=1000 AMT workers)

% Female 64.9

Average Age 36

% Earning below $60k/year

66.7

Education “higher than general pop.”

% Non-American 53

Page 6: Survey Experiment Modalities

Amazon Mechanical Turk Amazon Mechanical Turk (AMT)(AMT)

My AMT Survey (N=1824)

US National Averages (From Pew 2010 Survey)

Average Age 32.6 51.0

% Non-White 12.2 20.3

% Female 58.2 56.9

% Democrats 37.4 31.9

% Republicans 17.7 28.0

% Independents 44.3 40.1

% Liberal 36.7 17.5

% Moderate 46.9 43.9

% Conservative 16.3 38.6

DemographicsDemographics

Page 7: Survey Experiment Modalities

Amazon Mechanical Turk Amazon Mechanical Turk (AMT)(AMT)

Page 8: Survey Experiment Modalities

Follow-Up SurveyFollow-Up SurveyContacted all 1825 workers who completed valid HITs in Contacted all 1825 workers who completed valid HITs in original survey approximately 4 months lateroriginal survey approximately 4 months later

Offered $.50 for ~12 minute surveyOffered $.50 for ~12 minute survey426 valid responses (24%)426 valid responses (24%)

2nd batch of invites

Page 9: Survey Experiment Modalities

TradeoffsTradeoffs

Source: Paolacci, Chandler, & Ipeirotis (2010)

Page 10: Survey Experiment Modalities

Some Conclusions re Some Conclusions re AMTAMT

““Our analyses of demographic characteristics Our analyses of demographic characteristics suggest that MTurk participants are at least as suggest that MTurk participants are at least as diverse and more representative of noncollege diverse and more representative of noncollege populations than those of typical Internet and populations than those of typical Internet and traditional samples. Most important, we found that traditional samples. Most important, we found that the quality of data provided by MTurk met or the quality of data provided by MTurk met or exceeded the psychometric standards associated exceeded the psychometric standards associated with published research.”with published research.”

Buhrmester et al. (2011)Buhrmester et al. (2011)

““Our theoretical discussion and empirical findings Our theoretical discussion and empirical findings suggest that experimenters should consider suggest that experimenters should consider Mechanical Turk as a viable alternative for data Mechanical Turk as a viable alternative for data collection.”collection.”

Paolacci et al. (2010)Paolacci et al. (2010)

Page 11: Survey Experiment Modalities

Knowledge NetworksKnowledge Networks2007 Survey (via TESS)2007 Survey (via TESS)

Participants recruited via residential address Participants recruited via residential address searchessearches

Respond online (Internet access provided to recruits Respond online (Internet access provided to recruits who don’t have it)who don’t have it)

Pretty good samples, but imperfect (selection effects Pretty good samples, but imperfect (selection effects not completely purged) and expensive not completely purged) and expensive

DemographicsDemographicsPost-Survey Matched Sample

(N=1014)

% Female 50

% Non-white 26.7

% Liberal 25

% Moderate 41

% Conservative 33

Average Age 46

Ideological Intensity (0-3 scale)

.99

Political Knowledge (0-1 scale)

.487

Page 12: Survey Experiment Modalities

Recent (2012) KN Recent (2012) KN ProposalProposal

• Design: 4-wave study, longitudinal sample, all waves occurring within 1 yearDesign: 4-wave study, longitudinal sample, all waves occurring within 1 year

• Pretest: N=25 interviews each wavePretest: N=25 interviews each wave

• Sample: General population adults, age 18+, English-language survey-takersSample: General population adults, age 18+, English-language survey-takers

• Number of completed interviews:  N=2,000 wave 1, with about 70%-80% of Number of completed interviews:  N=2,000 wave 1, with about 70%-80% of wave 1 respondents completing each of the later waveswave 1 respondents completing each of the later waves

• Median survey length: 20 minutes each waveMedian survey length: 20 minutes each wave

• Multimedia/graphics: NoneMultimedia/graphics: None

• Incentives: $5 for each of waves 1-3, $10 for wave 4Incentives: $5 for each of waves 1-3, $10 for wave 4

• KN will provide standard deliverables (self-documented data file with all the KN will provide standard deliverables (self-documented data file with all the survey data, general demographic profile data, and field report documenting survey data, general demographic profile data, and field report documenting all sampling and data collection procedures, codebook, and panel recruitment all sampling and data collection procedures, codebook, and panel recruitment methodology)methodology)

• Price: $255,550 (No, that’s not a typo!)Price: $255,550 (No, that’s not a typo!)

Page 13: Survey Experiment Modalities

Yougov/PolimetrixYougov/PolimetrixOpt-in samplingOpt-in sampling

Random draw from target population matched with Random draw from target population matched with most comparable available panelists to create most comparable available panelists to create representative population samplesrepresentative population samples

Demographics + attitudinal/behavioral factorsDemographics + attitudinal/behavioral factors

Post-Survey Matched Sample

My Survey (2007, N=1200)

CCAP (2008, 6-wave panel)*

Average Age 40 44.2

% Female 51 52

% Non-white 31.5 16.7

% Democrat 38.0 36.1

% Republican 39.8 30.6

% Independent 22.2 33.3

% Liberal 26.4 22.8

% Moderate 36.8 34.5

% Conservative 36.8 35.3

*Notes:(1)Overweights battleground states 2-fold.(2)Demo weights based on age, race, gender, educ, marital status, kids, income, state, metro area, employment, citizenship(3)Attitude/Behavior weights based on religion, church attendance, evangelical status, news interest, PID, ideology.

Page 14: Survey Experiment Modalities

Comparing Modalities: KN vs. Comparing Modalities: KN vs. Polimetrix vs. Natural Survey: Polimetrix vs. Natural Survey:

Barabas & Jerit (2010)Barabas & Jerit (2010)

Page 15: Survey Experiment Modalities

Polimetrix vs. Pew vs. NES Polimetrix vs. Pew vs. NES (Partisan Distribution)(Partisan Distribution)

Source: Hill, Lo Vavreck & Zaller (2007)

Page 16: Survey Experiment Modalities

Polimetrix vs. Pew vs. NES Polimetrix vs. Pew vs. NES (Ideologial Distribution)(Ideologial Distribution)

Source: Hill, Lo Vavreck & Zaller (2007)

Page 17: Survey Experiment Modalities

Demographic ComparisonsDemographic ComparisonsSource: Hill, Lo Vavreck & Zaller (2007)

Page 18: Survey Experiment Modalities

Comparing Modalities: KN vs. Comparing Modalities: KN vs. Polimetrix vs. Natural Survey: Polimetrix vs. Natural Survey:

Barabas & Jerit (2010)Barabas & Jerit (2010)

Conclusions:“The results presented here should be encouraging to anyone devoted to the scientific study of politics because they suggest that what occurs in survey experiments resembles what takes place in the real world.”

“Although there was a discrepancy between the size of survey treatment effects and the general population in our natural experiment, we observed correspondence exactly where one would expect to find it—among those who were most likely to be exposed to media messages about the two government announcements.”

Page 19: Survey Experiment Modalities

SurveyMonkey AudienceSurveyMonkey AudienceNew enterprise for SMNew enterprise for SM

still figuring it out!still figuring it out!

Recruit participants from survey respondentsRecruit participants from survey respondents

Currently U.S. only, but likely to expandCurrently U.S. only, but likely to expand

$3 per finished response; $5 for rush project (3 business $3 per finished response; $5 for rush project (3 business days)days)

Custom create demographic profilesCustom create demographic profilesGender, age, income, location, education, race, industry of Gender, age, income, location, education, race, industry of employment, job function, marital status, employment status, home employment, job function, marital status, employment status, home ownership, vehicle ownership, smartphone ownership, exercise habits.ownership, vehicle ownership, smartphone ownership, exercise habits.

No political attitude selectors yet (though I’ve lobbied them!)No political attitude selectors yet (though I’ve lobbied them!)

Similar to Polimetrix, except don’t start with random draw from Similar to Polimetrix, except don’t start with random draw from populationpopulation

Much less expensive!Much less expensive!

Page 20: Survey Experiment Modalities

AAPOR 2010 Task Force on AAPOR 2010 Task Force on Online Panels Online Panels

Focus on nonprobability samplesFocus on nonprobability samples

Informative, but inconclusive…Informative, but inconclusive…

ConclusionsConclusionsIf research objective includes accurate estimate of population If research objective includes accurate estimate of population values, avoid nonprobability online panels.values, avoid nonprobability online panels.

Results differ significantly from probability-based methods (like Results differ significantly from probability-based methods (like RDD telephone) on range of behaviors and attitudes, with latter RDD telephone) on range of behaviors and attitudes, with latter being more accurate.being more accurate.

Nonprobability online panels sometimes appropriate, when precise Nonprobability online panels sometimes appropriate, when precise estimates of population values not criticalestimates of population values not critical

More research needed on evaluating and testing techniques used More research needed on evaluating and testing techniques used across disciplines to make population inferences from across disciplines to make population inferences from nonprobability samples.nonprobability samples.

Page 21: Survey Experiment Modalities

Ansolabehere & Schaffner Ansolabehere & Schaffner (2011)(2011)

Comparison of Survey ModalitiesComparison of Survey Modalities

3-mode study conducted in 20103-mode study conducted in 2010

1.1. Opt-in Internet panelOpt-in Internet panel

2.2. Live telephone interviews (using Live telephone interviews (using national RDD sample of landlines and national RDD sample of landlines and cell phones)cell phones)

3.3. Mail (using national sample of Mail (using national sample of residential addresses)residential addresses)

Page 22: Survey Experiment Modalities

Mode Sample Size

Field Dates Response Rate

Completion Time

Internet 1000 1/15/10-2/11/10

42.9% (RR1)

8.94 mins

Mail 1207 1/30/10-9/30/10

21.1% (RR3)

11.80 mins*

Phone 907 1/28/10-1/30/10

19.5% (RR3)

14.33 mins

Ansolabehere & Schaffner Ansolabehere & Schaffner (2011)(2011)

*mail recruits who took survey online

Page 23: Survey Experiment Modalities

Ansolabehere & Schaffner Ansolabehere & Schaffner (2011)(2011)Item Response Internet Phone Mail Validating

Source

Home Ownership

Own .613 .632 .632 .669 (CPS)

Mobility Moved in past year

At address 5+ years

.152

.555

.155

.609

.162

.519

.154 (ACS)

.588 (ACS)

Smoked 100 Cigarettes

Yes .504 .471 .497 .430 (NHIS)

Smoke Cigarettes now

Every or some days

.259 .242 .241 .203 (NHIS)

Voted in ‘08 (if registered)

Yes .888 .876 .821 .896 (CPS)

Vote Choice in ‘08

Obama

McCain

.482

.469

.454

.505

.553

.431

.529

.456

Avg Diff. .036 .035 .043

Page 24: Survey Experiment Modalities

Non-validatable Political Point Estimate Comparison, by ModeNon-validatable Political Point Estimate Comparison, by ModeIncludes State of Economy, Includes State of Economy, Approval of/Support for Obama, Congress, R’s Member, Abortion, Affirmative Action, Gay Marriage, Investing Social Security, Tax over $200k, Cut Approval of/Support for Obama, Congress, R’s Member, Abortion, Affirmative Action, Gay Marriage, Investing Social Security, Tax over $200k, Cut Spending, GovernmentSpending, GovernmentRight and WrongRight and WrongVoting MethodVoting MethodReligious and/or Political ContributionsReligious and/or Political ContributionsPolitical KnowledgePolitical KnowledgeNews SourceNews Source

Ansolabehere & Schaffner Ansolabehere & Schaffner (2011)(2011)

Internet vs. Phone

Phone vs. Mail

Internet vs. Mail

Avg Diff. (All measures)

.062 .042 .051

Avg Diff. (Attitudinal measures only)

.052 .042 .044

Weighted proportions of respondents in each category, excluding DK.

Page 25: Survey Experiment Modalities

Small (“negligible”) differences across modesSmall (“negligible”) differences across modesExcept…Internet respondents more politically knowledgeable & Except…Internet respondents more politically knowledgeable & made more political contributionsmade more political contributionsMail costs 5 times more than Internet & twice as much as phoneMail costs 5 times more than Internet & twice as much as phoneInternet half as costly as phone and faster turnaroundInternet half as costly as phone and faster turnaroundDifferences from other studies that found Internet samples less Differences from other studies that found Internet samples less valid than phone samples attributed to (1) more Internet users than valid than phone samples attributed to (1) more Internet users than 5+ years ago when prior data samples collected, and (2) advances 5+ years ago when prior data samples collected, and (2) advances in “science of constructing, matching and weighting opt-in Internet in “science of constructing, matching and weighting opt-in Internet panels”panels”Conclusion: “...an opt-in Internet survey produced by a respected firm can produce results that are as accurate as those generated by a quality telephone poll and that these modes will produce few, if any, differences in the types of conclusions researchers and practitioners will draw in the realm of American public opinion.”

Ansolabehere & Schaffner Ansolabehere & Schaffner (2011)(2011)

Page 26: Survey Experiment Modalities

U.S. Telephone OwnershipU.S. Telephone Ownership

Media Access,Media Access,Sweden vs. USASweden vs. USA

Page 27: Survey Experiment Modalities
Page 28: Survey Experiment Modalities

Sweden vs. USASweden vs. USA

Page 29: Survey Experiment Modalities

Sweden vs. USASweden vs. USA

Page 30: Survey Experiment Modalities

Sweden vs. USASweden vs. USA

Page 31: Survey Experiment Modalities

Swedish Phone Use by Swedish Phone Use by AgeAge

Source: Axelsson (2010)

Page 32: Survey Experiment Modalities

Why Persistent Mobile Why Persistent Mobile Phone & Internet Usage Phone & Internet Usage

Gaps?Gaps?

Income Inequality: USA vs. Sweden in 2011

Page 33: Survey Experiment Modalities

Effects of Income on Effects of Income on Internet Usage in U.S.Internet Usage in U.S.

U.S. has more of these folks

Page 34: Survey Experiment Modalities

What Do Data Tell Us?What Do Data Tell Us?Swedes more likely to be online in 2010 (by ~14 Swedes more likely to be online in 2010 (by ~14 percentage points), and make greater use of Internet percentage points), and make greater use of Internet

But, similar in fixed broadband and landline usageBut, similar in fixed broadband and landline usage

More likely to use mobile phonesMore likely to use mobile phones

But similar in volume of mobile calls sent and receivedBut similar in volume of mobile calls sent and received

Moral of story? Infrastructure looks, if anything, LESS Moral of story? Infrastructure looks, if anything, LESS hospitable to probability sampling in Sweden than in USAhospitable to probability sampling in Sweden than in USA

So, if RDD today works better in Sweden for generating So, if RDD today works better in Sweden for generating probability samples, reason seems likely to have more probability samples, reason seems likely to have more to do with attitudes toward surveys than infrastructureto do with attitudes toward surveys than infrastructure

Page 35: Survey Experiment Modalities

ConclusionsConclusionsOpt-in Internet samples here to stayOpt-in Internet samples here to stay

Cheaper (by a lot!)Cheaper (by a lot!)

Faster (by a fair amount…)Faster (by a fair amount…)

Primary competitor (RDD phone surveys) increasingly difficult Primary competitor (RDD phone surveys) increasingly difficult 14% of adults “unreachable” in Sweden?14% of adults “unreachable” in Sweden?

Estimated 25% of U.S. households cell only in 2010.Estimated 25% of U.S. households cell only in 2010.

““Unreachable”: ~13% (AAPOR 2010); others say moreUnreachable”: ~13% (AAPOR 2010); others say more

Open up new possibilitiesOpen up new possibilitiesE.g., cross-national samples/panelsE.g., cross-national samples/panels

Most current evidence suggests that with current matching and Most current evidence suggests that with current matching and weighting techniques, Opt-in Internet samples weighting techniques, Opt-in Internet samples can can be be representative of target populationsrepresentative of target populations