Making Surveys Work
for You: Some Important
Design Considerations
By
Don A. Dillman*
For
Net-Conference on Survey Methods and Measurement
June 15,2011
*Regents Professor in the Department of Sociology and Deputy Director of the Social and Economic Sciences Research Center at Washington State University, Pullman, Washington 99164-4014. [email protected] http://www.sesrc.wsu.edu/dillman
1 c Don A. Dillman June 15, 2011
Welcome…
…to this web conference on designing effective
surveys.
I hope that you will find it useful.
My focus today is on how survey methods are,
changing, and to provide helpful suggestions for navigating through these changes.
Let me begin with some of the questions you
asked in the material sent to me, and how I’ll go about responding.
c Don A. Dillman June 15, 2011 2
A few of your questions.
What errors are introduced when modes are combined? Can landline and cell phones be combined effectively?
Is it okay to use mail primarily, with telephone follow-up? Is there bias from internet surveys?y
How can response rates be improved (Incentives?) and how much is enough?
What can we do to get youth and minorities to respond? What is the state of the science on framing opinion questions? Can you give us practical guidelines for urban vs. rural areas?
c Don A. Dillman June 15, 2011 3
Providing answers is a challenge?
Different answers may fit different situations.
When I looked at these (and other) questions, I thought I
could be most useful to focus on criteria for decision-making and give examples of what’s possible or practical…
…and then, responding to questions will allow us to go
deeper into the explanations.
First, some background.
c Don A. Dillman June 15, 2011 4
None of our main data collection
methods are working as well as we
would like when used alone.
Telephone has ~70% coverage; response rates have
declined steady; now as low as10-15% in national surveys.
The Internet has “effective” household coverage < 70%;
(access + skill +frequency of use) and response rates are often low.
In-person methods affordable for only a few surveys, and
coverage is increasing problem.
. c Don A. Dillman June 15, 2011 5
Attempts to circumvent these limitations are prevalent but also have problems
Cell phones added to RDD samples can improve
coverage of the general public, but are also expensive.
In addition, cell phone numbers are often carried from one
geographic area to another when people move.
Effective household Internet access is lower than the
number of connections, because of infrequent use.
We cannot use email to access general public households. We have no sampling algorithm equivalent to RDD.
We can only obtain web response by approaching through postal
mail.
Mail Survey Methods are
experiencing a revival of interest
The U.S. Postal Service makes available (through
vendors) all residential addresses (The Delivery Sequence File, or DSF)
It is now our best household sampling frame and may It is now our best household sampling frame, and may
be our only source for accessing nearly all (~ 95-97%) households.
One can request by mail the return of a mail
questionnaire and/or a response via web.
c Don A. Dillman June 15, 2011 7
Two design considerations now seem
to dominate survey methodology
Tailored Design—matching data collection modes to
populations, topics, and survey task. (Your studies intrigue me in this regard.)
Mixed-mode surveys, using 2y , g nd, 3, rdmode to overcome
problems of single mode, is becoming increasingly important as a way of dealing with the shortcomings of individual modes.
Thus, we need to figure out where mail, web and
telephone can and cannot help us collect better quality survey data in both situations.
c Don A. Dillman June 15, 2011 8
Are Mixed Modes More of a Solution
or More of a Problem?
Perhaps there is no choice!
Response rates are declining (telephone)
Often we cannot reach everyone by single modes. Sample Frames are less adequate (random digit dialing)p q ( g g) Cannot contact random samples of internet addresses
But, are we getting improvement in data quality or perhaps
producing even larger problems by mixing modes?
Do we get the same measurement across modes?
Do we reduce non-response error through adding a new mode
or only increasing measurement problems?
9 c Don A. Dillman June 15, 2011
To talk about these the need for
tailoring and for mixed-mode designs
I need to provide some context:
Doing successful sample surveys means that we must
conduct sample surveys in a way that reduces error to the extent practical
the extent practical.
We also need to recognize that there are multiple
sources of error in sample surveys.
10 c Don A. Dillman June 15, 2011
What a Survey Can Do:
Provide the distribution of a characteristic in a population by surveying only a small portion of the members of that population.
…which allows us to answer questions like these: What percent of Minnesota residents now smoke?
What percent of unemployed adults in Minneapolis use tobacco each day?
Do students in suburban high schools have higher rates of tobacco use than students in inner city schools?
c Don A. Dillman June 15, 2011 11
How do sample surveys allow
us to make precise estimates
for the opinions and
12 c Don A. Dillman June 15, 2011
for the opinions and
behaviors of millions from a
survey of hundreds or a few
Coverage Error
Sampling Error
In Order for Surveys to Make Estimates With Known Precision, We Must Limit Four Kinds of Error:
Measurement Error
Non-response Error
13 c Don A. Dillman June 15, 2011
Coverage
Error
Occurs because all members of the population are not given an equal or known (non-zero) chance of being selected for the survey.
H it h
14 c Don A. Dillman June 15, 2011
How it happens:
Incomplete, poorly constructed and out-of-date lists (e.g. employees, members, and/or customers).
All households do not have telephones.
Many people do not have Internet access or the skills to use it.
For Example
~25%+ of households do not have landlines. Only ~65% of households have internet access from
home (additional 11% from somewhere else).
S i t t ti b d
Some internet connections are by modem.
Households without internet access or landline access
differ from others in significant ways.
15 c Don A. Dillman June 15, 2011
Occurs because only some members of the population are allowed to participate in the survey while others are systematically excluded.
Sampling
Error
The general (but incomplete) indicator of precision is completed sample size.
16 c Don A. Dillman June 15, 2011
Number of completed interviews and precision of estimates at 95% confidence (quadrupling number of completed questionnaires decreases sampling error by one-half).
100 = +/- 10% 1150 = +/- 3% 400 = +/- 5% 2200 = +/- 2%
If 60% of our 400 respondents say they smoke cigarettes each day we can be 95% confident that between 55% and 65% of all people in our population smoke cigarettes each day (ignoring other possible errors).
The Pascal Triangle
and its Application
How is it possible to make estimates with stated confidence (e.g. 95%) that tens of thousands or millions of people have a particular characteristic,
while surveying so few people?
17 c Don A. Dillman June 15, 2011
An Inappropriate Perspective
for an Election Survey
Voters on election day
All Voters are Unique
18 c Don A. Dillman June 15, 2011
A Necessary Perspective in an
Election Survey
Voters on election day
In a two candidate election there are TWO types of voter. Let’s assume 50% are going to vote for each candidate.
19 c Don A. Dillman June 15, 2011
Why Does Random Sampling with
Complete Coverage Work?
• It lets probabilities work, as described in a Pascal Triangle.
• The Pascal Triangle shows the probable outcomes of 1The Pascal Triangle shows the probable outcomes of 1 to 10 random draws of respondents from populations in which one-half of respondents have each characteristic.
20 c Don A. Dillman June 15, 2011
Chances in Ten Consecutive Draws of Randomly Selecting All Combinations of Red and Blue “Voters”
1 2 3 2 4 8 Number of Draws Sum of Chances 252 3 4 5 6 7 8 9 10 8 16 32 64 128 256 512 1024 21 c Don A. Dillman June 15, 2011
Interpretation of the 10 “Draws”
252
Chances of 10 Blue - 1/1024
Chances of 5 red 5 blue - 252/1024
Chances of 4-6 of each color - 672/1024 22 c Don A. Dillman June 15, 2011
Number of Randomly Needed Depends Mostly on Sample Size and Not on Population Size
Sample Size for +/- 3% Precision 92 341 Population of this size 100 500 341 517 880 985 1,057 1,066 1,068 500 1,000 5,000 10,000 100,000 1,000,000 100,000,000 23 c Don A. Dillman June 15, 2011
An Observation
Much of the traditional design work in random sample
surveys has been to assure high coverage so that random sampling will be effective, i.e. all people who vote have an equal chance of being selected in the sample
sample.
But, there are also other sources of error to consider.
24 c Don A. Dillman June 15, 2011
Measurement Error
Occurs because of inadequate question wording, placement in the questionnaire, the interviewer and/or the respondent.
Some known sources of measurement error:
25 c Don A. Dillman June 15, 2011
• Some known sources of measurement error: • Mode effects:
Social desirability Primacy/Recency Acquiescence • Question order effects • Different visual layouts
• Measurement differences can be a significant problem when trying to use more than one mode for the same study.
For example, Interviewer Presence
May Have a Normative Influence:
60
How would you rate your health?
50 50 0 10 20 30 40 50
Excellent Good Fair Poor
Per
cent
Mail Telephone Face-to-face
Respondents more likely to give culturally acceptable answer on telephone or in-person. 30 38 43 50 50 42 18 10 10 1 2 2 26 c Don A. Dillman June 15, 2011
Occurs because some people do not respond when requested to answer the survey questions, and they are different in some way from those who do answer.
Non-response Error
F l
27 c Don A. Dillman June 15, 2011
For example:
Are older people more likely to respond than younger people?
Are employed people more likely to respond than unemployed people in an employment rate survey?
How $2 Token Cash Incentive in Advance Influenced Response Rates Across Age Groups
60 70 80
1993 Survey of Washington State New Drivers License Holders
Without $2 With $2 0 10 20 30 40 50 <35 36-49 50-60 61+ Age Without $2 28 c Don A. Dillman June 15, 2011
To summarize, it’s important to reduce each kind of error. Maximizing one at expense of others is NOT a good idea.
Coverage
Error Sampling Error
c Don A. Dillman June 15, 2011 29 Measurement
Error
Nonresponse Error
A Case Study on seeking
improvement by mixing modes
…the Gallup Organization tried to switch from telephone to web for periodic cross-sectional employee studies in 1999, and noticed significant differences in
measurement.
At the same time, some response rates appeared to be
falling.
These concerns led to an experiment on the effects of
four survey modes on measurement and switching modes as a means of improving response rates.
Case Study: an attempt to improve data quality by offering multiple modes
Four modes were examined:
Telephone, Interactive Voice Response (IVR), web and mail. Can response rates be improved by offering two modes
in sequence?q
What are the measurement consequences?
For details, see: Dillman, Tortora, Phelps, Swift, Kohrell, Berck, and Messer, 2009. “Response rate and measurement differences in mixed-mode surveys using mail, telephone, interactive voice response, and the Internet.” Social Science Research, 38(1), 1-18.
c Don A. Dillman June 15, 2011 31
Case Study Comparison:
Telephone, IVR, Mail, and Web
Questions were comparable across modes, using the preferred format for telephone.
Q2. Overall, how satisfied are you with your long distance company?
1 Not at all satisfied
c Don A. Dillman June 15, 2011 32
1 Not at all satisfied
2
3
4
5 Extremely satisfied
Q3. Considering what you get and what you pay for it, how would you rate the overall value of your long distance company’s products and services?
1 Terrible
2
3
4
5 Outstanding
How the study was implemented
A sample of individuals that submitted warranty
information on new purchases in order to get access by either mail or telephone.
A two-step implementation process; offer initial mode
with multiple contacts and after a pause of 2-3 weeks offer the second mode.
Mail to telephone. Telephone to mail. IVR to telephone. Web to telephone
c Don A. Dillman June 15, 2011 33
Response rates improved
significantly after a mode switch was
made
83% 81% 51% 48% 50 60 70 80 90 100c Don A. Dillman June 15, 2011 34 0 10 20 30 40 50 Additional Response Initial Response
Effects of Initial Modes on choosing
most positive category for five
“long-distance” satisfaction questions
Percent choosing positive labeled end-point
50 43 50 60 ent Telephone IVR Mail Web
c Don A. Dillman June 15, 2011 35 39 26 36 21 39 21 33 22 21 11 29 19 18 9 29 18 16 0 10 20 30 40 Q2 Q3 Q4 Q5 Q6 Perc e Web T I M W T I MW T I M W T I M W T I M W 27
Aural modes produced different
results than visual modes
Responses to aural modes (telephone, IVR) were
similar.
Responses to visual modes modes (web, mail) grouped
together.
Control panels showed that primacy/recency not a cause
of measurement differences;.
Traditional explanations of differences e.g. social
desirability seemed unlikely.
Was the higher response rate worth
the cost?
Probably not, because of the measurement differences
that resulted.
Telephone obtained more extreme answers regardless of
whether first mode or last mode
Additional panels showed no primacy/recency effects. The demographics did not change in a way that
suggested different kinds of respondents answered the second mode.
Either mail or telephone alone would probably have been
adequate, except for the measurement difference. c Don A. Dillman June 15, 2011 37
Why this example?
Mixing survey modes is not always a desirable solution. Doing that requires us to deal with many measurement
issues.
But using mail-alone may also be undesirable But, using mail alone may also be undesirable. To sort out the possibilities and implications, we need to
think about coverage, non-response and measurement.
c Don A. Dillman June 15, 2011 38
Two Ways that Switching Modes
May Change Measurement?
1. Mode choice influences question structure and wording.
2. Presence vs. absence of interviewer (e.g.
l h b) i fl d
telephone vs. web) influences respondent answers.
c Don A. Dillman June 15, 2011 39
1. How Changes in Wording and
Construction Influence Answers
Each mode has evolved to favor certain ways of asking
questions to maximize the potential to obtain quality data.
This tendency comes from the “tradition” of how each
mode was designed to maximize its potential use, regardless of how it might be done in another survey mode.
c Don A. Dillman June 15, 2011 40
Two Examples of Question
Construction Differences:
Not offering the same answer categories in all
modes.
Asking an open-ended question in one mode and a
l d d d ti i th th d
c Don A. Dillman June 15, 2011 41
closed-ended question in the other modes.
Unintentional Construction Effect:
Volunteered Responses on Telephone
To what extent do you favor or oppose expansion of NATO?Telephone: Mail:
1 Strongly favor 1 Strongly favor
c Don A. Dillman June 15, 2011 42 2 Somewhat favor 2 Somewhat favor
3 Somewhat oppose 3 Somewhat oppose 4 Strongly oppose 4 Strongly oppose 5 NO OPINION
6 DON’T KNOW 7 REFUSE
Unintentional Construction Difference:
Open vs. Closed Ended
Marital status was asked differently by one
organization on telephone and web (Tortora, 2004).
c Don A. Dillman June 15, 2011 43
Telephone: What is your marital status?
Web: Which of the following best describes your marital status? Single
Married Separated Divorced Widowed
Unintentional Construction Difference:
Results from a Telephone and Web
Comparison
What is your marital status? Differences were significant and tracked across monthly data collections.
Telephone (n = 10,380)
Web (n = 9,713)
Difference
c Don A. Dillman June 15, 2011 44
Single 26.8 23.5 -3.3 Married 63.1 62.0 -1.1 Separated 7.0 10.5 +3.5 Divorced 1.0 1.5 0.5 Widowed 2.0 2.4 0.4 TOTAL 99.9 99.9
2. How does the presence of interviewer/human interaction influence respondent answers
A substantial literature shows that mail obtains different answers to many questions than for interview surveys.
Among the reasons:
Presence of the interviewer (interaction expectations)
P d h t i ti f i t i d d i t b
Presumed characteristics of interviewer and desire to be seen in a favorable way.
Aural and visual information are perceived using different sensory systems and processed using separate neural centers in the brain.
Aural communication places higher demands on respondent
memory.
c Don A. Dillman June 15, 2011 45
Is the presence of an interviewer
helpful or not?
Interviewers control the delivery of survey information
whereas respondents have control over how they complete the survey in self-administered surveys.
Interviewers can have positive impacts by clarifying
i f ti d ti ti d t
c Don A. Dillman June 15, 2011 46
information and motivating respondents.
Interviewers can have negative effects because
respondents are more likely to give socially desirable answers and acquiesce (rather than disagree) in the presence of an interviewer.
Interview
er
Presence May Have a
Normative Influence:
Social Desirability
63
60
70 How often do you drive a car after drinking alcoholic beverages? Respondents are more likely to give culturally acceptable answers on the telephone.
c Don A. Dillman June 15, 2011 47
1 11 36 0 52 0 2 9 27 0 10 20 30 40 50 Mail Telephone (2= 14.8 p < .001)
Frequently Occasionally Seldom Never Don’t Know
Percent
An Effect of Interviewer Presence:
Acquiescence,
the Tendency to Agree
75 70 70 72 87 87 79 78 81 70 80 90 Mail n=884 Telephone n=400
Percent of respondents who agree with each of 9 opinions on how to encourage seat belt use
c Don A. Dillman June 15, 2011 48 58 39 36 27 19 49 48 35 21 0 10 20 30 40 50 60 1 2 3 4 5 6 7 8 9 p Per cen t
Greater Extremeness on Four-Point
Scales
in General Public Survey
Percent choosing “not a problem” (vs. small, moderate, or serious)
52 49 71 74 60 53 59 60 70 80 T elephone Mail
c Don A. Dillman June 15, 2011 49
49 18 32 40 28 10 12 21 41 24 32 34 0 10 20 30 40 50
Mean for 9 Questions: Telephone, 47.1 Mail, 31.9 Difference, 15.2 (Dillman & Mason, 1984)
Will Scalar Questions With Little or No Social Desirability Produce the Same Results Across Modes?
Telephone respondents tend to give more extreme
positive answers than do web respondents.
In some cases social desirability, acquiescence or
c Don A. Dillman June 15, 2011 50
In some cases social desirability, acquiescence or primacy/recency do not seem to explain why.
Does Presence of an Interviewer Influence Response to Scalar Questions?
A series of identically worded scalar questions was
tested in a web telephone experiment.
Please note at this point that we are starting to bridge
between interviewer presence and visual vs. aural communication. These attributes are intertwined.
c Don A. Dillman June 15, 2011 51
If we use exactly the same wording for scalar questions in aural vs. visual surveys…
Will we get the same answers
A series of evaluations were made for a variety of scalar
questions in random sample student surveys at Washington State University.
c Don A. Dillman June 15, 2011 52
Question formats evaluated in
Telephone vs. Web Experiments
Fully labeled, 1-step.
5-category scales – 9 comparisons. 7-category scales – 6 comparisons.
Polar point labeled 1 step
c Don A. Dillman June 15, 2011 53 Polar-point labeled, 1-step.
5-category scales – 13 comparisons.
7-category scales – 4 comparisons.
11-category scales – 4 comparisons.
Branched 2-step scales.
7-category scales – 10 comparisons.
5-Category Fully Labeled Scales
How satisfied are you with Washington State University as a place to go to school? Would you say you are Very Satisfied, Somewhat Satisfied, Neutral, Somewhat Dissatisfied or Very Dissatisfied? Telephone Survey
c Don A. Dillman June 15, 2011 54
Mean Responses by Mode:
5-Category Fully Labeled Scales
4 5
Telephone Web
*
*
*
* *
*
* *
c Don A. Dillman June 15, 2011 55 1 2 3 Q2 Q16 Q24 Q5 Q21 Q25 Q5 Q21 Q25 Satisfied/ Dissatisfied
Agree/Disagree Construct Specific
*p ≤ .05
5-Category Polar-Point Labeled Scales
On a 5-point scale where 5 means very satisfied and 1 very dissatisfied, how do you feel about the quality of INSTRUCTION in the classes you have taken at WSU?
Web Survey Telephone Survey
c Don A. Dillman June 15, 2011 56
Web Survey 4 5 Telephone Web
* *
*
*
*
* *
*
*
*
*
Mean Responses by Mode: 5-Category Polar-Point
c Don A. Dillman June 15, 2011 57
1 2 3
Q2 Q4 Q6 Q7 Q16 Q24 Q5 Q21 Q25 Q18 Q19 Q9 Q17
Satisfied/Dissatisfied Agree/Disagree Extremely Likely/ Not Likely
Best Possible/ Worst Possible
*p ≤ .05
Summary of Findings on
Mode Effects in Scalar Questions
All six scale formats we tested produced a slightly more positive response on telephone than on web for the vast majority of test items.• 5 and 7 category fully labeled
c Don A. Dillman June 15, 2011 58 • 5, 7, and 11 category polar point labeled • 7 category 2-step scales
See: Christian, Leah Melani, Don A. Dillman, and Jolene D. Smyth. 2008. “The Effects of Mode and Format On Answers to Scalar Questions in Telephone and Web Surveys.” In Lepkowski, et al. Advances in Telephone Survey Methodology. New York: Wiley-Interscience. pp. 250-275
Conclusions:
Interviewer Effects in
Scalar Questions
Respondents answer scales more positively and are
more likely to select the extreme positive category on the telephone compared to web, mail, or face-to-face (when a showcard is provided).
c Don A. Dillman June 15, 2011 59
Finding a solution to scalar questions for which social
desirability seems not to apply may be as difficult as when it does apply.
Implications—Presence of Interviewer
May need to avoid combining telephone and either web
or mail modes, especially for panel surveys, where attempts to measure change from Time 1 to Time 2 are being made.
Statistical adjustment may be a possibility, but I leave
that to others to figure out.
However,
What if we combined mail and web. They are both
visual, and measurement differences should be quite small for most questions.
Also, if we can use Postal Service addresses without ,
names, maybe we can get people to go to the web.
Let’s see what might happen if we do that.
c Don A. Dillman June 15, 2011 61
2005-2009; Mail-only surveys with incentives added achieve reasonable response rates Four general public surveys with 12 page questionnaires, $5 token cash
incentive with request, 3-4 follow-ups, respondent-friendly design. (Stern and Dillman, 2006; Smyth et al. 2010; Messer and Dillman, Forthcoming)
2005 regional Lewiston/Clarkston social
participation survey, Telephone directories 66% 2007 Lewiston/Clarkston Quality of Life Survey
Addressed-based sample (ABS), 71%
2008 Washington state-wide community satisfaction
Survey, ABS, most recent birthday 55%
2009 State of Washington Household Economic Survey
Knowledgeable respondent, ABS 68%
c Don A. Dillman June 15, 2011 62
Study Procedures
(Smyth, Dillman, Christian and O’Neill, 2010, Applied Behavioral Science) 12 page questionnaire, 50 items, up to 80 responses (depending
upon branching); a 20-25 minutes survey
Four contacts. Pre-notice letter.
Questionnaire (or web request).
Thank-you post card.
Replacement questionnaire (adjusted by treatment).
$5 token cash incentive included with initial mail questionnaire or web request.
Data collected November 7, 2007, to January 10, 2008.
c Don A. Dillman June 15, 2011 63
Additional background on building mail
implementation systems. (Results of an Experiment, Contacts x Incentive Sent With the Request vs. Post payment).
Incentive
Incentive 11ststMailing Mailing
(%) (%) 22ndndMailingMailing (%) (%) 33rdrdMailing Mailing (%) (%) 44ththMailing Mailing (%) (%) No incentive 20.7 36.7 46.7 52.0 $1 C h 40 7 52 0 61 3 64 0 $1 Cash 40.7 52.0 61.3 64.0 $5 Cash 48.7 60.7 66.7 71.3 $5 Check 52.0 62.7 66.7 67.3 $10 Check 44.0 56.7 62.0 66.7 $20 Check 54.0 70.7 75.3 79.3 $40 Check 54.0 63.3 66.0 69.3 Promise of $50 23.3 43.3 53.3 56.7
Note: each treatment group contained 150 subjects (James and Bolstein 1992)
64 c Don A. Dillman June 15, 2011
Treatments for Lewston/Clarkston
Study
1. Mail preference: Send mail questionnaire and mention web with initial request.
2. Mail preference: Send mail questionnaire but withhold mention of web for about two weeks. 3. Web preference: Web invitation with no mail
questionnaire, but explain that mail questionnaire will be sent in about two weeks.
4. Equal preference: Encourage response by either mode equally.
c Don A. Dillman June 15, 2011 65
Sample Sizes
Treatment n
A. Mail Preference (mention web) 400
B. Mail Preference (no web mention) 400
C. Web Preference (mail in 2 weeks) 600
D. Equal preference (respondent’s choice) 400
TOTAL 1800
Tailoring to the survey topic and location and creating a unified design between mail and web
In the slides that follow…
Pictures of location to be surveyed. Creating common screens for mail and web. Use of common branding for mail and web.
Choice of stationary, envelopes and content based upon
rethinking of personalization strategies given that “name” could not be used.
c Don A. Dillman June 15, 2011 67
Tailoring/personalizing the survey to
the location and population
Photos taken of local landmarks, artwork, and
symbols to make survey recognizable and visually attractive.
c Don A. Dillman June 15, 2011 68
Mail Questionnaire
Use of “personalized” images to help respondents feel connected to
survey and to place an emphasis on study area instead of on survey source. Also used similar color and design to connect to web survey.
c Don A. Dillman June 15, 2011 69
Unified design between mail (on left) and web (on right)
c Don A. Dillman June 15, 2011 70
The opening web page
Used an entry page similar to front cover of paper survey, still focusing on making the survey recognizable through familiar images.
c Don A. Dillman June 15, 2011 71
Design of the web survey—focus on population not sponsor
Question 2
Similar design format to paper survey, and use of familiar image in upper left-hand corner of the screen.
Personalized Correspondence
All letters used WSU
stationary.
Photo of questionnaire
cover used to tie different elements together.
c Don A. Dillman June 15, 2011 73
Exterior of Envelopes (2ndand 4thContacts) Used WSU address labels.
Used a return label showing the photo from survey cover and the survey title to increase familiarity.
c Don A. Dillman June 15, 2011 74
Initial vs. Final Response Rates
Treatments Initial Response Rate (%) Final Response Rate (%) Difference Mail preference (mention web early)
54 62 +8
Mail preference
(no web in early contacts)
62 71 +9
Web preference
(no mail in early contacts)
38 55 +17
Equal preference (choice) 52 63 +11 c Don A. Dillman June 15, 2011 75
Final response rate, paper preference
vs. web preference
Treatments Web (%) Paper(%) Total (%)
Mail preference (mention web early)
4 58 62
Mail preference 1 70 71
(no mention web)
Web preference 41 14 55
Equal preference (choice) 13 50 63
c Don A. Dillman June 15, 2011 76
Response rate summary
Postal DSF with addresses only (i.e., no names) can
produce reasonably high response rates—71% mail preference and 55% web preference.
Using a postal letter to drive people to the web was quite
effective—41% responded by web.
Adding the mail option after two weeks brought in
considerably more respondents—14%
Adding a mail option to the web was more effective than
adding a web option to the mail –14% vs. 1%.
c Don A. Dillman June 15, 2011 77
Is non-response error a problem?
It does not help much to improve response rates if our
respondents are different from non-respondents on variables important to the study objectives.
Thus we need to compare respondent characteristics on Thus, we need to compare respondent characteristics on
web vs. mail within the different treatment groups.
Lewiston/Clarkston: Demographics in the web preference treatment by mode of response
Responded by Mail (n = 80) Responded by Web (n = 232) Difference Mean age 61.6 51.4 -10.2 < 50 yrs 26.6 41.7 15.1 > 65 yrs 48.1 17.0 -31.1 % Female 68.4 59.2 -9.2 % Female 68.4 59.2 9.2 % White 93.8 89.2 -4.6
% Employed (full time) 34.7 50.7 16.0
% Retired 52.0 25.6 -26.4
% With 4-year degree 13.9 30.8 16.9
% Married 42.9 74.0 31.1
% Widowed 24.7 2.2 -22.5
c Don A. Dillman June 15, 2011 79 Red indicates significant difference at .05 level.
Web respondents were quite different
from mail on 7 of 24 substantive
attitudes/opinions
Mail Web Diff.
% attached to the area 90.0 80.4 -9.6
% think willingness for community involvement
has increased 47.7 31.7 -16.0
% think fish population increased 18.9 38.0 19.1
% more internet use improves quality of life 43.4 62.1 18.7
% think more cell use improves quality of life 26.9 44.1 17.2
% think enviro. protection is too weak 16.3 30.7 14.4
% gray wolves not threat to domestic animals 2.5 9.9 7.4
c Don A. Dillman June 15, 2011 80 Red indicates significant difference at .05 level.
But web + mail (all web preference respondents) quite similar to mail preference group.
Mail Preference (n=261) Web Preference (n=312) Difference Mean age 56.2 54.1 -2.1 < 50 yrs 39.4 37.8 -1.6 > 65 yrs 29.1 25.2 -3.9 % Female 52 7 61 6 8 9 % Female 52.7 61.6 8.9 % White 88.9 90.4 1.5
% Employed (full time) 43.2 46.6 3.4
% Retired 36.1 32.2 -3.9
% With 4-year degree 29.6 26.4 -3.2
% Married 56.0 66.0 10.0
% Widowed 11.7 8.0 -2.2
c Don A. Dillman June 15, 2011 81 Red indicates significant difference at .05 level.
Summary: types of respondents by
mode
Web on its own brings in specific types of
respondents and leaves others out.
Web and mail used together bring in a wider range of
respondents – comparable to mail used alone.p p
Our best chance of reducing nonresponse error is to use
both.
c Don A. Dillman June 15, 2011 82
What happens when respondents are
given a choice of which mode to use?
Treatments Web (%) Mail (%) Total (%)
Mail preference
(mention web early) 4 58 62
Mail preference
(no mention web) 1 70 71
c Don A. Dillman June 15, 2011 83
Web preference 41 14 55
Equal preference (choice) 13 50 63
Response rates for the choice treatments are lower than for the mail preference (no mention) treatment but higher than for the web preference treatment.
Expansion to state-wide data
collection
A similar model was used, i.e. personalize questionnaire
to the state with pictures.
We pushed to the web by withholding mail. Two tests in successive years
Two tests in successive years.
Statewide Mail Questionnaire with map and regional pictures
12” X 8.5” booklet with ‘personalized’ images to help respondents feel connected to survey and to place an emphasis on study area instead of on survey source. Also used similar color and design to connect to web survey. (Messer and Dillman, Forthcoming)
85 c Don A. Dillman June 15, 2011
Web Questionnaire
● Designed to emulate the mailquestionnaire and appeal to the local population.
86 c Don A. Dillman June 15, 2011
Response Rates From Mail only is higher than web + mail in three studies. But, 2/3of respondents were successfully “pushed” to respond via web.
Mail Only Web+ Mail
Lewiston/Clarkston
Rural Region Survey-2007 71% 41%+ 14% = 55% Washington Community
Survey-2008 57% 31%+ 15% = 46%
Washington Economic
Survey- 2009 68% 34%+ 18% = 52%
c Don A. Dillman June 15, 2011 87
A few conclusions
We no longer have a good sampling frame for telephone
surveys
If we mix telephone (and web or mail) we stand a good
chance of obtaining different measurement chance of obtaining different measurement.
Mail-alone may provide a good option, but
implementation methods affect response.
Web + mail has lower response than mail-alone, but may
be a good option for future.
c Don A. Dillman June 15, 2011 88
Your questions
You may have questions about what I presented here, or
you may wish to go beyond what I have done here and ask about things not yet discussed.