Transcript Slide 1

A Roper Poll on Polling
Sussman, 1985
• “Can interviews with 1,500 or 2,000 people
accurately reflect the views of the nation’s
population or is it not possible with so few
people?”
– Possible: 28%
– Not possible: 56%
– No opinion: 15%
For in-class use
• Write down questions you have about polling
Consult your “neighbors” if you wish
Bring along to the next class
• Any questions except: how is it possible to represent a
big country with only a thousand people (I’ll cover that)
How polling is done (statistical questions)
How polling is carried out (practical questions)
Questions about the questions used in surveys
About the use of polls
About their honesty/meaningfulness
About pres. polls; exit polls; questionable polls
Why take (sample) surveys?
• Can’t afford to get information on entire
population
• Timeliness—survey data can be gathered
quickly
• Survey data is sometimes better than
“hard” data (e.g., crime data)
• Can get survey data on attitudes (and
behavior that would be hard to observe)
• A survey was even proposed to improve
the decennial Census
Survey limited areas for very-hard-toreach individuals, extrapolate to other
areas to arrive at more complete
Census data
• Scuttled by the Reps; it probably would
have increased the count in Dem areas
Why study surveys
(in this course)?
• They are ubiquitous in politics
Over 100 pre-election polls about the
pres election (in one election year)
Many more in gub, Sen, House races
Constant barrage of poll data about
issues
Presidential approval is a standard
news item (has been for 50 years)
Why study surveys (cont.)
• Surveys/polls are not confined to politics
More in marketing than elsewhere
Mall intercept studies
• Often well done; often not. You should be
able to distinguish good from bad.
• Even good surveys are hard to interpret;
you should be able to do to
Basic terminology
• Population: individuals that we wish to
speak about
Adults in the United States
Likely voters in one state
UR graduates, 1995-2004
• Sample: individuals we gather info from
Individuals referred to as respondents
• Basic idea of sampling: We extrapolate
from a sample to a population
We take a sample—usually much
smaller than the population—and make
inferences about the population
• Primary issues:
How small or large a sample do we
need?
How do we sample (what individuals do
we get information from)?
Nonprobability samples
• Haphazard sample
Sometimes called “random”
• Purposive sample
Select “representative” units
• Quota sample
Match population proportions (e.g., half
women, half men in a national sample)
• Problem with nonprobability samples:
Can’t assess accuracy (no statistical basis)
• Lots of bias likely to occur in sampling:
Avoid certain parts of town
Avoid hard-to-reach respondents
British ex—interviews near tube stations
Select cooperative respondents (big no-no)
• Solution: use probability (random) samples
Probability samples
•
•
•
•
•
Simple random samples
Stratified random samples
Clustered random samples
Systematic random samples
Multi-stage random samples (combine
elements of one or more of above)
Simple random samples
• Definition
Every element and every combination of
elements has an equal chance of being
selected; equivalently, every element has
an equal chance of being selected and
the selections are made independently.
• Not true of all random samples We may
have time to discuss; otherwise, see text.
Simple random samples
• How can we represent so many with so
few?
• Intuitive idea (explained in class)
• Formula for proportions (explained in
class)
The great simplification
Confidence intervals,
different sized samples
Confidence interval
Sample size
(+ or – this percent)
2000
2.2
1500
2.5
1000
3.1
500
4.4
200
6.9
Based on the assumption of srs; proportion of .50.
Other random samples
• Most samples of people are not srs
• One example only
Stratify: divide the population in groups
or areas; specify quotas; sample
randomly within
• Major reasons for using more complex samples:
They are feasible, cheaper
Significantly, and surprisingly, they are more
accurate in some respects
A Famous Failure
The 1936 Literary Digest “Straw Poll”
•
Postcard “ballots” succeeded in
predicting the winner in 1920, 24, 28,
and 32
•
10m ballots sent out to addresses
drawn from every phonebook in US,
rosters of clubs (e.g., Auto), city
directories, lists of voters, and mail
order & occupational data.
•
2.2m respondents predicting:
– Alf Landon 57%
– FDR 43%
•
On election day, FDR won by a
landslide (62.5%, all but MN & VT)
What went wrong??
Literary Digest v. George Gallup:
“straw” meets “science”
Some problems with the Literary Digest Straw
Poll
• Sampling bias.
• Response Bias.
• Time. Early September.
• New Deal Coalition: class polarization
• DO NOT LET RESPONDENTS SELFSELECT!
George Gallup got it right in 1936 using “quota
sampling”
– Interview a quota from each demographic
group
Gallup Poll Results
Final poll (for winner) vs election
•
•
•
•
•
•
•
•
•
1936 FDR -4.8
1940 FDR -2.7
1944 FDR -2.1
1948 HST -5.1
1952 IKE
-4.1
1956 IKE
1.9
1960 JFK
1.1
1964 LBJ
2.7
1968 RMN -0.4
1972 RMN 1.3
1976 JC
-2.1
1980 RR
-3.8
1984 RR
0.2
1988 GB
2.6
1992 WJC 6.0
1996 WJC 2.8
2000 GWB 0.1
2004 GWB c2.0
Gallup Poll Results (cont.)
•
•
•
•
Averages
1936-1948: 3.68
1952-1978: 1.94
1980-2004: 2.50
• Source: Election Polls: Elections 1936-2000 (Gallup Poll report)
http://institution.gallup.com/content/default.aspx?ci=9448
Surveys today
• Most are telephone surveys (since c1980)
Use random digit dialing (RDD), not
numbers from phone directories
• Small, but increasing number are internet
surveys
Biggest promoter is Harris Interactive (Roch)
Methods, validity still not fully accepted
• Focus groups supplement surveys
• Some nonscientific polls survive (be wary)
Problems with Public Opinion Polls
• Occasionally people lie
Rs don’t want to admit “unacceptable”
behavior/attitudes
Ex (1989): Gov. Wilder’s race in Virginia
• Attitudes don’t always match behavior
Example: 1930s study of hotels
• Individuals often give ill-informed,
inconsistent responses (next slides)
Ex: R’s are ill-informed
•
•
•
•
•
•
•
•
•
1948 Socialist party’s presidential cand.
1966 Incum. House cand. in their district
1970 Know the pres is limited to two terms
1972 Know China to be communist
1977 Know meaning of no-fault insurance
1987 Can name Chief Justice of the SCt
1987 Know Bill of Rights is part of Const.
The list could go on—and on
New study (1/31/05): 75% of high school
students think flag-burning is illegal
21%
46%
67%
63%
31%
8%
41%
R’s give inconsistent responses
Problems (cont.)
Question wording matters
• Do you think the United States should
allow public speeches against democracy?
• Do you think the United States should
forbid public speeches against
democracy?
One might expect:
% “allow” = % “don’t forbid”
Forbid/Don’t allow example (cont.)
• Results of three surveys using these Q’s
Public speeches against democracy
Don’t
Allow
Forbid
1940
25
46
1974
56
72
1976
55
80
Problems (cont.)
Question order/context can matter
• Survey asked whether Soviet journalists
should be allowed to move freely, report
on U.S. (Result: low % yes)
• Then asked, first, a question about U.S.
journalists in the Soviet Union. (Result:
Much higher % say Soviet journalists
should be free in U.S.
Problems specifically
of election polls
• Predicting turnout
Biggest problem
• Undecideds
A problem for low profile elections
• Timing
Early polls can be quite inaccurate
Pollsters now work until election day.
Why believe polls?
• With all these sources of error, how can we
believe polls actually work?
• With respect to technical matters:
Polls match population data quite well.
Repeated polls give consistent results.
Polls do a good job—generally—of
predicting elections (if not overinterpreted).
Surely better than nonscientific methods
Why believe polls? (cont.)
• With respect to substantive matters:
Aggregate opinion is highly meaningful
even if individual responses are
problematic
We can often identify, overcome biases
Polls do a good job—generally—of
predicting elections
Some “solutions” to
problems with polls
• Look at multiple questions
Attitudes: A chord, not a note
• Be esp careful with controversial issues
• Look at responses over time
Changes are significant even if absolute
level is not fully meaningful
• Don’t over-interpret results (easy with
presidential polls because a few % matter)