Surveys and Questionnaires
I want to accomplish three different things in this module. First, I want to introduce you to the three major kinds of surveys and discuss the tradeoffs in deciding which kind of survey to do. Second, we will talk about response rate, why it is important, and how it is calculated. Third, we will discuss how to minimize three kinds of bias that can creep into surveys: interviewer bias, question bias, and questionnaire bias. (We have already talked about the other kind of bias, sample bias, in the last section.)
Types of Surveys
Although there are practically an infinite number of different variations of different surveys that one can perform, virtually all fall into one of three major categories: face-to-face, telephone, and self-administered. Let us consider each in turn and then do some comparisons.
1. Face-to face. The classic face-to-face survey is the door-to-door survey. However, any survey in which an interviewer asks questions to the respondent in person is a face-to-face survey. This can be done in a work setting, in schools, or in people's homes. A few years ago I helped in a survey of workers who worked in a set of office buildings. We surveyed them in the morning as they entered the buildings (using a combination of a cluster sample and a systematic sample). But the interviews were face-to-face.
2. Telephone surveys. This is the most common form of survey today. It is a compromise between the other two kinds in many ways. But it is getting more difficult to do well, largely because of telemarketing and devices such as caller id and answering machines, and most recently. cell phones. Happily for those of us in the survey business, the "Do Not Call List" will help reduce resistance for legitimate surveys as people get bothered less at home.
3. Self-administered surveys. Any questionnaire that the respondent fills out him or herself is a self-administered survey. Your teacher evaluations are self-administered. It works well when you have a captive audience. One of the most common forms is the mail survey, which has a few advantages but many disadvantages, as we shall see.
Response rate is an important consideration in any survey. It can be measured in several ways. Ultimately, what we want is the percentage of units in the population targeted for selection who actually completed surveys. However, that number is not always easy to find. For example, suppose you call a telephone number several times but never get an answer in a survey you are doing of registered voters. If this is a business number or if no one living there is a registered voter, it is not a member of the population and therefore should not count as a nonresponse in the response rate. Suppose you do reach the household but are unable to determine whether the potential respondent is registered before she refuses the interview? Again, you cannot be sure how to count this person. All this yields several possible estimates of response rates. You could calculate the percentage of those actually contacted and determined to be potential respondents who actually completed the interviews. That would be a liberal estimate of response rate. Or you could use a conservative estimate, the percentage of the units originally chosen that responded, only discounting those which were determined to be ineligible (businesses or people not registered to vote in our example). Otherwise, if we can't reach them, we assume that they were part of the target sample. Generally we should use the conservative estimate.
Response rates are important when certain kinds of units are more likely to be surveyed than other kinds of units. Say for example, wealthy people use answering machines or called identification to screen their calls and are therefore less likely to be interviewed. That creates a bias in which wealthy people are less likely to be chosen. If refusals and non-responses are randomly distributed across all groups, then a low response rate is not a problem. A random sample minus a random sample equals a random sample. Unhappily, random refusals and non-responses are usually not the case.
You can see how much distortion this can create with a little "pretend" example. Suppose we draw a sample of 100 likely voters. Suppose that in the population of all voters 30% will actually vote for Ross Perot, the Reform Party candidate. But suppose that these Perot voters are so distrusting that two thirds of them refuse to talk to interviewers. Suppose that those who will vote for the other candidates do not refuse. So in our initial target sample or 100, 20 of the 30 Perot voters we try to interview refuse to be interviewed. That leaves us with 10 Perot voters. We then replace them with 20 new target voters to be interviewed. Of these, 6 should be Perot voters (.3 x 20 = 6). But 2/3's of these 6, or 4 also refuse to be interviewed. So that gives us 2 more Perot voters in our sample. We replace those 4 and about 1 of these new potential interviews is also a Perot voter. But she refuses and is replaced by a non-Perot voter. So we end up with 12 Perot voters, or 12% of the sample, when an unbiased estimate that would have come with a 100% response rate would have given us 30% Perot voters. That is a pretty big distortion!
Sometimes researchers will talk about refusal rates. This is just the opposite of the response rate. To put it in algebraic form, 100% - Response Rate = Refusal Rate. This may be conceptualized a little differently however. It could be the percentage who actually refused, and would not count those who were never successfully contacted. So make sure you explain exactly how it was calculated or ask exactly how someone else calculated it.
What do you do when you have refusals or units that were target for interviewing but you were unable to reach? You have two choices. You can just use what you have and end up with a smaller sample--this is called sampling without replacement. Or you can replace them with other units--sampling with replacement. Mathematically speaking, as long as the additional units are chosen so that each unit in the population has an equal chance of being chosen, it makes little difference. If you know what the response rate is likely to be, you should account for this in determining the original sample size. For example, if you expect a response rate of 50% in a mail survey and you want a sample of 300, you should select 600 in order to end up with 300 at the end. The same would be true of other kinds of samples, with one interesting exception. In the "plus one" method of telephone survey sample selection, you just go ahead and start with 300. Then for a refusal of a number, you just add another digit. So if 642-3341 refuses or is unable to be reached, you replace it with 642-3342. The only time this does not work is if you get into a group of nonworking numbers. If this happens a number of times, you may have to go back and choose another sample to add on to your original sample. But that is ok so long as it is also a probability sample in which each member of the population has an equal chance of being chosen. A random sample plus a random sample equals a random sample!
Since refusals are bad, how do we maximize response rates? The answer depends on the kind of survey one is doing.
In telephone surveys:
In mail surveys:
In face-to-face surveys:
Tradeoffs in Different Interviewing Formats
The table below compares the three most common types of public opinion
surveys in terms of several important parameters that anyone planning a
survey should consider. As you can see, there is no perfect survey. I should
point out that these are very rough estimates, because the actual numbers
will depend a great deal on the population that is being surveyed and on
the number and complexity of questions.
|Response rate||60-80%||40-65%||50% or much lower|
|Length||up to an hour||up to 20 minutes||2 pages max|
|Interviewer bias||potentially high||moderate||none|
|Speed||several weeks at best||fast--as low as 24 hrs||month at best|
|Question complexity||can be high||moderate at best||must be low|
|Cost||high: $25-100/resp||moderate: $10-30||low: $5-10|
As you can see from the table, telephone interviews are generally the interview of choice. Mail interviews are too slow, must be very short, must only use the simplest of questions that require no more than checking a box, have a very low response rate, and take a long time to do. However, they are cheap. They work well for surveys of the membership of social/business/interest groups in which it is politically important to ask for everyone's opinion. But here you are talking about a census rather than a survey. At the other extreme, face-to-face interviews are generally just too costly, unless you have a captive audience and do not have to do a lot of traveling to get the interviews.
However, things are changing as people become more resistant to telephone calls because of the excesses of tele-marketing and other "junk" calls. Response rates have fallen from the 70 percent range to the 40 percent range in just the last decade. We used to get 70% in our telephone surveys here where we have about everything working in our favor, but recently we have fallen into 50s and low 60s. The "Do Not Call List" should help things here over the next few years, but the growing tendency for people to use cell phones as their only phone may begin to cause significant problems. The General Social Survey (which we use in the workbook for this course) is a door-to-door survey. That is why they are able to ask so many questions. The surveys of the future may involve specially selected Web target populations to represent the general
Bias can result from unreliable and/or invalid questions. Sloppily worded questions that are subject to multiple interpretations are unreliable and an create bias in unknown ways. More typically, invalid questions that push the respondent toward a certain kind of answer creates bias in predictable ways.
Avoid loaded terms or phrases:
Do you support the Reagan administration's request for a major increase in the U.S. defense budget to counter the massive Soviet arms buildup? (Republican National Legislative Action Survey, 1982)
Do you favor or oppose Republican efforts to make huge increases in the current defense budget? (DNC 1996 Presidential Poll)
Avoid "motherhood and apple pie" or silly questions:
Do you feel it is very important to institute common-sense legal reform to stop excessive legal claims, frivolous lawsuits and overzealous lawyers? (National Republican Congressional Committee poll, 1995)
Agree or Disagree? Sex crimes, such as rape and attacks on children, deserve more than mere imprisonment; such criminals ought to be publicly whipped, or worse. (Adorno, The Authoritarian Personality). This particular example was found to be biased toward lower educated people. Well educated people knew that agreeing with outdated modes of punishment was a sign of lack of education, and were therefore more likely to disagree, however they really felt.
Avoid response set bias by reverse wording several agree/disagree format questions when they occur in groups. Resonse set bias can occur in two ways. "Acquiescence" response set bias occurs in lower educated people who tend to feel intimidated and and as a result try to please the interviewer by agreeing with all statements. Occasionally a well-educated person will be a "nay sayer" and disagree with almost everything just for the sake of argument. A good interviewer can tell when this is happening and can terminate the interview or it can be thrown out as not useable.
Avoid touchy matters on which the respondent is unlikely to tell the truth. People want to be liked and generally give the socially desirable response. That is why so many more people say they voted than actually did vote.
How often do you smoke pot?
Did you drink alcohol on a daily basis during your pregnancy?
A closely related problem is questions that call for subjective self-evaluations--avoid them. Again social desirability comes into play here. Make them as objective as possible.
Do you generally vote when there is an election?
Did you vote in the election last month? (better)
Are you a good student? (poor)
What is your grade point average at present? (better)
Are you a racist? (very poor)
Agree or disagree? Blacks don't get ahead in life because they don't work as hard as people in other ethnic groups. (better--this is called "social racism," by the way)
Avoid long questions--they are unreliable--shoot for 15 words or less.
There is a philosophical debate as to whether GAO
should be an organization consisting of many highly specialized employees
with great expertise in narrow issue areas, supplemented by a few generalists
with a broad range of knowledge, or an organization with a few highly specialized
employees in each issue area, working with a large number of generalists
who have a broad knowledge of many issue areas. Proponents of the specialist
approach point to the high level of expertise in very difficult technical
fields that is fosters, providing GAO with an institutional knowledge and
memory unparalleled in government. Proponents of the generalist approach
argue that high degrees of specialization lead to organizational inflexibility,
while a large number of generalists will give GAO flexibility to respond
to changing demands. Which philosophy best reflects you views?
___ many specialists, few generalists
___ many generalists, few specialists
___ equal number of specialists and generalists
___ don't know/no opinion (House Republican Conference member survey, 1996)
Avoid compound questions. A compound measure consisting of several questions is fine, but keep each question separate.
Do you believe that Social Security and Medicare should be targets for major cuts in the battle to balance the nation's budget? (1996 DNC Platform Poll) This should be two separate questions.
Use filter questions or built-in filters to eliminate "door-step' opinions. You only want to ask people about things on which they have an opinion. So eliminate them if they have not "read or thought about the debate on whether patients should have the right to sue HMO's" or give them a graceful way to choose the answer " I have not had time to think about that issue."
When possible, use questions that have been well-tested by other researchers. If this is not possible, then pretest or use a focus group to see how questions are being interpreted.
Just as individual questions can cause problems, so can the overall layout of a questionnaire. Here are a number of areas of concern.
Branching instructions or interviewer instructions. Anytime you have a question that is contingent on another question, you must make sure that the instructions are clear as to when to ask it. Avoid these altogether in self-administered questionnaires if at all possible. I like to place interviewer instructions in brackets and in caps to distinguish them from questions themselves. [IF "YES," GO TO QUESTION 6]
Include codes when possible and use higher numbers for higher amounts of whatever it is that you are measuring, if it is at the ordinal or interval level of measurement.
Stick to close-ended questions where possible--that greatly simplifies coding and analysis later on. if it is open-ended, make sure that the interviewer writes down the answer as completely and accurately as possible.
Multiple response questions should be coded as a series of dichotomous yes/no questions.
What kinds of movies do you like?
horror? __ yes __ no
mystery? __ yes __ no
romance? __ yes __ no
action? __ yes __ no
other? [SPECIFY] ____________________
Separate answer boxes/blanks so that it is clear which one goes with each question.
If the election were held today, for whom would you
vote in the 2000 presidential election? __ Dole __Buchanan
__Bush __ Forbes __ Bradley __ Gore
(This is a very bad layout!)
Question interaction. If you are asking open-ended questions and close-ended questions on the same subject, you should use the open-ended questions first, because the answers to the close-ended questions can suggest answers for the open-ended questions. Separate question that can influence each other, or randomize the order so as to minimize consequences.
Keep questionnaire as short as possible. Tired respondents give unreliable answers. I have found in dealing with many clients who say they want surveys that most don't really know what they need to know. They often use surveys as a defense against decision making. So spend a lot of time forcing them to do some hard thinking about what it is that they really need to decide and how answers to very specific questions will help them decide. The rule I use is "actionability." Don't ask a question unless you know how an answer will help you take some action. Don't just ask it because it is interesting, or it might be nice to know. This is a hard rule for even me to follow in that I am interested in finding out about a lot of things.
When you ask people for their opinions and they let you into their lives, as a professional you have certain obligations to them and to upholding the reputation of your profession. Here is a partial list.
The best way to avoid interviewer bias is to train and properly reward interviewers for their hard work. What follows is an outline of the steps in interviewer training. I will go over the specifics in class.
Steps in Telephone Interviewer Training
I. General Training
A. What makes a good interviewer
1. Use of instruction sheets (call-sheet, fall-back, and questionnaire)
2. Quality over quantity
3. properly completed interviews (4 to 1 ratio)
4. pleasant neutral voice -- not telephone therapy
5. probing without bias to maximize reliability and validity
B. Handling refusals -- being politely persuasive
1. Assume that every respondent needs incentives
2. Good timing -- reschedule if seems bad
3. Pleading -- personalize -- help me as a poor student
4. Individual questions -- allow to skip uncomfortable questions
5. provide reassurance with sponsor's name and number
C. Role playing on minimizing refusals
2. Validate completeness by examining questionnaires
3. Verify by calling selected respondents to ensure interview properly done
E. Productivity (4/1 completion-refusal ratio) and Payment (1.5 x min. wage
or 2/1 base/productivity incentive system)
II. Specific Training
A. Purpose of survey
B. How telephone numbers selected
C. Use of Call-Sheet
2. When to call (7-9:30 pm Sun, Mon-Thu, Sun afternoon)
D. Introduction and respondent selection at each household
E. Fall-back statements
1. Read all questions and answers
2. Review all special instructions
3. Recording answers
4. Role playing practice