Presentation on theme: "Sampling Design Questions, questions, questions –Do you support U.S. role in Iraq?"— Presentation transcript:
Sampling Design Questions, questions, questions –Do you support U.S. role in Iraq? –What % of lettuce shipment is bad? –How many children are obese? –What’s the price of gas at the pump across Minnesota? Practically impossible to poll entire population Use a part to make conclusions about the whole Idea #1: Use a SAMPLE to make conclusions about the POPULATION But sample must be representative of population
Polling began in Pennsylvania Harrisburg Pennsylvanian in 1824 predicted Andrew Jackson the victor –He did win the popular vote –But, like Al Gore, he didn’t win the electoral votes and John Quincy Adams took the election Straw polls were convenience samples Solicited opinions of “man on the street” No science of sampling for 100 years Conventional wisdom: bigger the better
1936 election and the Literary Digest survey Magazine had predicted every election since 1916 Sent out 10 million surveys---and 2.4 million responded They said: Landon would win 57% of the vote What happened: 62% Roosevelt landslide
What went wrong? Sample not representative Lists came from subscriptions, phone directories, club members Phones were a luxury in 1936 Selection Bias toward the rich Voluntary response: Republicans were angry and more likely to respond Context: Great Depression – 9 million unemployed –Real income down 33% –Massive discontent, strike waves Economy was main issue in the election
Idea #2: Randomize Randomization insures sample is representative of population Randomization protects against bias Simple Random Sample (SRS): every combination of people has equal chance to be selected How to do it right
Some examples of non-random, biased samples 100 people at the Mall of America 100 people in front of the Metrodome after a Twins game 100 friends, family and relatives 100 people who volunteered to answer a survey question on your web site 100 people who answered their phone during supper time The first 100 people you see after you wake up in the morning
Is blind chance better than careful planning and selection? Another classic fiasco 1948 Election: Truman versus Dewey Ever major poll predicted Dewey would win by 5 percentage points Truman showing the Chicago Daily Tribune headline the morning after the 1948 election.
What went wrong? Pollsters tried to design a representative sample Quota Sampling Each interviewer assigned a fixed quota of subjects in numerous categories (race, sex, age) In each category, interviewers free to choose Left room for human choice and inevitable bias Republicans were wealthier, better educated, and easier to reach –Had telephones, permanent addresses, “nicer” neighborhoods Interviewers chose too many Republicans
Quota Sampling biased Republican bias in Gallup Poll Quota sampling eventually abandoned for random sampling Repeated evidence points to superiority of random sampling YearPrediction of GOP vote Actual GOP vote Error in favor of GOP
How large a sample? Not 10 million, not even 10,000! Remarkably it doesn’t depend on size of population, as long as population is at least 100 times larger than sample Idea #3: Validity of the sample depends on the sample size, not population size Like tasting a flavor at the ice cream shop SRS of 100 will be as accurate on Carleton College as in New York City! Most polls today rely on 1,000-2,000 people
Gallup Poll record in presidential elections since 1948 Year Sample Size Winning candidate Gallup prediction Election result Error 19525,385Eisenhower51.0%55.4%4.4% 19568,144Eisenhower59.5%57.8%1.7% 19608,015Kennedy51.0%50.1%0.9% 19646,625Johnson64.0%61.3%2.7% 19684,414Nixon43.0%43.5%0.5% 19723,689Nixon62.0%61.8%0.2% 19763,439Carter49.5%51.1%1.6% 19803,500Reagan51.6%55.3%3.7% 19843,456Reagan59.0%59.2%0.2% 19884,089Bush56.0%53.9%2.1% 19922,019Clinton49.0%43.2%5.8% 1996Clinton52.0%50.1%1.9% 2000Bush48.0%47.9%0.1%
A peek ahead... A good rule of thumb is that the margin of error in a sample is, where n is the sample size. For n = 1,600, that’s 2.5%. Most political polls report margins of error between 2-3%. The rule of thumb margin of error doesn’t depend on population size, only on sample size
Other sampling schemes Stratified sampling Goal: Random sample of 240 Carleton students To insure representation across disciplines, divide population into strata –Arts and Literature 20%- Humanities 15% –Social Sciences 30%- Math/Natural Sciences 35% Choose 240 x.20 = 48 Arts and Literature 240 x.15 = 36 Humanities 240 x.30 = 72 Social Sciences 240 x.35 = 84 Math/Natural Sciences Within strata, choose a simple random sample
Stratified sampling Advantages: Sample will be representative for the strata; Can gain precision of estimate Disadvantages: Logistically difficult; must know about the population; May not be possible Note Stratified sample is not a simple random sample Every possible group of 240 students is not equally likely to be selected
Cluster sampling – an example Warehouse contains 10,000 window frames stored on pallets Goal: Estimate how many frames have wood rot Determining if a frame has wood rot is costly Sample 500 window frames Pallets numbered 1 to 400 Each pallet contains 20 to 30 window frames Sample pallets, not windows. Pick SRS of 20 pallets from population of 400. Cluster sample consists of all frames on each pallet
Cluster sampling Door-to-door surveys – City blocks are the clusters Airlines get customer opinions –Individual flights are the clusters Advantage: Much easier to implement depending on context Disadvantage: Greater sampling variability; less statistical accuracy
Who likes Statistics?
Most common forms of bias Response bias Anything that biases/influences responses Non-response bias When a large fraction of those sampled don’t respond, such as Voluntary response bias Most common source of bias in polls
Sampling badly: Convenience sampling Sample individuals who are at hand Survey students on the Quad or in Sayles or in Stats class Internet polls are prime suspects American Family Association online poll on gay marriage
You critique it ► Before 2000 election: What to do with large government surplus ► (1) “Should the money be used for a tax cut, or should it be used to fund new government programs?” ► (2) “Should the money be used for a tax cut, or should it be spent on programs for education, the environment, health care, crime-fighting, and military defense?” ► (1): 60% for tax cut; (2): 22% for tax cut
Another type of response bias “Some “Some people say that the 1975 Public Affairs Act should be repealed. Do you agree or disagree that it should be repealed.” Washington Post, Post, Feb Results: Results: For repeal: 24%, Against repeal: 19%, No opinion: 57% No No such thing thing as the Public Affairs Act!
Non-response Non-respondents can be very different from respondents Student surveys at end of term had about 20% response rate General Social Survey (www.norc.org) has % response rate, with 90 minute survey! Huge variability in media and government response rates Typically, media rates at about 25%; government at about 50%. Takes large amount of money, time, and training to insure good response.
Do you believe the poll? What questions should you ask? Who carried out survey? Who carried out survey? What is the population? What is the population? How was sample selected? How was sample selected? How large was the sample? How large was the sample? What was the response rate? What was the response rate? How were subjects contacted? How were subjects contacted? When was the survey conducted? When was the survey conducted? What are the exact questions asked? What are the exact questions asked?