Audience Dialogue

Know Your Audience: chapter 7
Telephone surveys

Most of what has been written in the earlier chapters of this book also applies to telephone surveys, but for telephone surveys some things are different: in particular, sampling and some aspects of interviewing. These are covered in this chapter.

1. Sampling telephone numbers

When you are doing a telephone survey of the general population, there are two methods of sampling: from a telephone directory, or by random digit dialling. Both methods work, but both have their problems.

Sampling from telephone directories

A telephone directory is not a complete list of the residential phone numbers in an area. For example, in Australia about 10% to 15% of residential numbers are "silent": unlisted, by the owner's request. Other numbers are unlisted because they have been connected after the directory went to press. Australian directories are already three to six months out of date when issued; as new directories are published about once a year, they are up to 18 months out of date at the end of their issue period. On average, a directory is a year out of date, and between 10% and 20% of residential entries change each year.

Though the names attached to the numbers turn over fairly rapidly, the numbers themselves are much more stable, and generally continue from year to year. When somebody moves to another exchange area, the new resident often takes over the telephone number; when people move only a short distance (within the same exchange area) they usually take their old telephone number to their new address. When a subscriber leaves an address, and the new resident does not take over the phone number, it is usually reallocated to a different address in the same exchange area, after about six months.

Because the numbers are more stable than the names or addresses, telephone surveys normally use only the numbers, regardless of whether the corresponding name or address is the same as it was when the directory was printed.

Unlike door to door surveys, which normally use cluster sampling, clustering is not used in telephone surveys. The main purpose of clustering is to save on interviewer travel expenses, which don't exist in telephone surveys. Because clustering tends to reduce the effective sample size, telephone surveys can use smaller samples than door-to-door surveys.

The first stage in drawing a sample from a telephone directory is to calculate how many telephone numbers are required. Begin with the sample size you want, then enlarge that figure to allow for refusals to be interviewed, ineligible numbers, and numbers which are not answered.

Also decide how many people should be interviewed at each household contacted. Do you want to interview only one person per household, or all adults, or base the number of interviews on the number of people in the household? As interviewing more than one person per household saves very little money, and interviewing more than one per household reduces the effective sample size (qv), it's normal to interview only one person at each phone number.

How many telephone numbers to sample

When you draw a sample of numbers from a telephone directory, many of these don't result in interviews. This experience would be fairly typical:
Begin with 100 entries.
30 are business entries - leaving 70.
Attempt to ring 70 numbers. 5 turn out to be disconnected.
After many attempts, only 40 answer.
10 of the 40 refuse to participate in the survey.
This leaves 30 successful interviews from the 100 numbers.

If the area you want to survey is only a small part of the area covered by the telephone directory, and the directory has a single set of alphabetical entries, you will have to look through many more than 100 entries to find 100 in the survey area.

Another problem arises if you don't want to interview at all households, but only those that meet certain criteria - e.g. listening to your radio station. If only one person in 3 listens to your station, you'd get only about 10 interviews (instead of 30) from the list of 100 numbers.

The percentage of entries that are businesses varies greatly between directories, and is often higher with small directories than those from large cities.

The refusal rate can be anything between 1% and 40%, depending mainly on the skills of the interviewer, and where the respondent lives (refusals are much more common in big cities), but very little on the subject matter. With inexperienced interviewers, to be on the safe side one should allow for a fairly high refusal rate of say 25% in countries where phone surveys have been widely used. In areas where almost no phone surveys have been done before, refusal rates are usually very low, often less than 1%.

The proportion of numbers which are never answered depends mainly on how many times an unanswered number is re-rung. If you ring back up to 10 times, at varying times of day and days of week, leaving at least two weeks between the initial and final attempts, you can reach 90 to 95% of numbers - except in areas which have many holiday homes inhabited only a short part of each year. If you don't have the patience or the resources to keep trying for so long, you should try at least three times to reach each number - varying the time of day and day of week - and you will usually succeed in reaching 85 to 90% of the telephone numbers.

Drawing samples from phone directories

Some of the many pitfalls to be guarded against in selecting telephone numbers from a directory are:

If you are making a quick estimate of the percentage of business entries in a directory, by taking the top entry from each of 100 columns, you may get too high an estimate. But if you take the bottom entry from 100 columns, you'll probably get too low an estimate. This applies to Australian directories, but perhaps not in all countries. The reason for this is that the number of lines in a column is not quite fixed. Entries for middle-sized businesses, taking up two or three lines of the directory, are not split across columns.

The safest way to sample from telephone directories is by taking a fixed number of lines from each Nth column in the directory.

As some directory entries take more than one line, to avoid giving multiple-line entries too high a chance of being chosen, lines which do not contain phone numbers are ignored. For example, if you had selected line 59, and lines 59-61 were a single extra-large entry:

59. Blixenderbloxen, WMP & RQ,
60.  2931Angolelaterana Highway,
61. Mongu Lealui....................581-262

...the number is in line 61, not 59, so you would not use it. You'd take no number from this column, and go straight to the next column. If 10% of lines contain no actual number, you'd have to select 110 lines to get 100 numbers.

Selecting columns

If you need more numbers than there are usable columns in a directory, you'll need to select several line numbers in each column. If you need fewer numbers than columns, you'd select one line number, but not use every column.

For example, if the directory has residential entries starting on page 51 and finishing on page 209, that's 159 pages (from the end of page 50 to the end of page 209). If there are 4 columns on each page, that's 4 x 159 or 636 columns.

Let's assume you want to sample 350 numbers. That's 55% of the number of columns. If you rounded it down to 50%, you could take 2 columns on each page, selecting 2 random numbers between 1 and 4 to work out which columns to include. However if you select too few entries, it's a nuisance to have to create a new sample part-way through the survey, so select slightly more numbers than you think you'll need. In this case, I suggest taking 5 columns out of every 8 (i.e. every 2-page opening). Select 5 different random numbers between 1 and 8 to work out which columns to check on each page opening.

The easiest way to do this is to use the serial number on a banknote. I just happen to have one in my pocket, numbered 95618391. Counting only digits from 1 to 8, and ignoring any repetition, these numbers are usable: 5 6 1 8 3. Luckily there were 5 different numbers. If there were fewer, I'd need a second banknote.

The same principles can be used to select any proportion of columns in a directory.

Selecting lines on a column

Once a random set of columns has been found, a second random number is needed, to work out which line of each column should be sampled. Most Australian directories have about 123 lines per column. This would suggest choosing a random number between 1 and 123, to determine which line should be used in each column. However, because of the fact that the top and bottom few lines in each column may not be typical (as explained above), and the fact that some columns may have fewer lines than others, it's safer to choose a random number between about 4 and 120.

Back to my wallet: I look for a banknote with a serial number ending between 004 and 120 (just as well I have plenty of notes today!)...the first such number is 018.

Counting lines in phone directories can be extremely tedious. The easy way to do this is to count down to line 18 (or whatever) only once, then make a cardboard template. Make it just big enough that if your thumb holds the bottom of the cardboard level with the bottom of the directory pages, the first line fully visible above the top of the cardboard is line 18.

[##drawing of thumb holding template on page]

In fact, phone directories are not always printed accurately, and sometimes the first line fully visible above the template may be line 19 or 17. Too bad. Think of the choice of line 18 as referring to a particular place on the page, rather than a particular line number. It won't upset the results of the survey.

As you look at line 18 (or whatever) in each selected column, copy it out if it is a residential entry with a number on that line, and ignore it if it is:
(a) a business-only entry, or
(b) a line without a telephone number on it (i.e. the first line of a multiple-line entry), or
(c) a residential entry outside the area to be surveyed.

If it is a "after hours" listing as part of a business entry, and a name is given (as in the B. Bluggs example above), see if there is a main entry under that name. If there is, ignore the business "after hours" listing. If not, or if no name is given, accept the "after hours" listing as a valid number to be surveyed. If you're not sure whether a business is also a residence, include the number on the list to be rung, and reject it later if it is only a business.

Avoid biases in unused numbers

In this way, you build up a list of residential numbers to be surveyed. You shouldn't need to use all the numbers, so take steps to avoid any consistent pattern in the numbers which are unused.

For example, if you listed the phone numbers to be rung by working through the directory, taking the 18th line on every 3rd column, the numbers at the end of the list will be those of people whose surnames begin with Z. Now, there is probably something special about people whose surnames begin with Z . In Adelaide, where I live, many of the surnames beginning with Z are Italian and Polish. So if I took the required sample without ringing these numbers, I'd have a sample which may be lacking in people of Italian and Polish origin. Most of these are Roman Catholic, and may have different views on some social issues from the rest of the population.

In New Zealand, most Maori names are in the first half of the alphabet, in Scotland a large proportion of names begin with M, and in Iceland - well, just consult the Icelandic phone book one day, and see for yourself!

To ensure that all sections of the phone directory are represented, call the sampled numbers in a different order from their sequence in the directory. This will mean that the numbers that are not called will come from all parts of the directory.

The easiest way to do this is to copy each number onto a separate call log: a piece of paper which will record the history of calling that number. Before doing any interviews, shuffle the logs into random order. More details on call logs are given below.

Random digit dialling

If the problems of sampling from telephone directories deter you, there is an alternative: dialling numbers at random. The main problem with this method is that you can waste a lot of time dialling nonexistent numbers, so you need to know which ranges of numbers have been allocated in the area you are surveying.

In some parts of Australia, it is not feasible to sample numbers from a telephone directory. For example, if you want to do a telephone survey in Albany, W.A., you'll find that the "local" telephone directory covers the entire southwest of Western Australia, in alphabetical order of surname, and most of these entries are not from Albany. In fact, to get 100 Albany entries, you have to examine and discard about 800 others., Thus the easiest way to create a telephone survey sample in Albany is random digit dialling.

Another situation where RDD is needed is when most numbers are not in the directory - this is the case for mobile or cell phone numbers in many countries.

The problem is to find out which number ranges have been allocated. Some telephone directories list the ranges, others do not. In some places, the local telephone authority is co-operative, and in others it is not. If you only want a small group of local numbers, it's probably best to go through any people you know in the telephone authority in your area.

If all else fails, you can estimate the allocated number ranges by going through a telephone directory - but this is very time-consuming, and you're likely to miss some small ranges. To find what ranges are in use, look up lists of places that are found in even the smallest communities: post offices, banks, schools, and so on. This will show you the prefixes used in each area.

Even if you know the allocated number ranges, you still have to contend with many non-existent numbers. For example, the telephone numbers allocated in Albany (Western Australia) used to be between 41-0000 and 41-8799: that is, 8,800 numbers. However about a quarter of these were not being used.

Sometimes, when you dial a nonexistent number, you get the appropriate "no such number" dial tone. However, in some areas, dialling a nonexistent number will produce the same tone as a telephone that is ringing and not being answered. In this case you can never be sure whether a number is nonexistent, or whether its owners are seldom home. When you suspect that a particular number may not be connected, the telephone authority will usually verify this for you. But when you give the authority a list of 100 possibly non-existed numbers, it's less likely to co-operate.

Another hazard with random digit dialling is business numbers and unlisted business extensions, which you may encounter in large quantities. Though the process of drawing a sample of numbers is much quicker for random digit dialling than for directory sampling (if a computer is used), random digit dialling increases the time you must spend on the telephone, rejecting businesses and nonexistent numbers.

Though it's messy to draw one sample, and use it all up, then have to go back and draw a second sample, this is the best approach for random digit dialling, when you have no idea what proportions of the allocated number ranges are used. You therefore draw an initial sample on the basis that every telephone number in the allocated ranges is in use, and carry out the first part of the survey using this sample. Unconnected numbers found in the first stage are counted, and the total count is used to estimate the size needed for the second sample.

This may be made clearer by taking an example, again using Albany in Western Australia.

Assume that the final required sample size is 100 people. Allowing for a possible high refusal rate brings the number up to 130, and adding another 10 to 15% for connected numbers which are never reached brings it up to approximately 150. Allowing for business numbers (here estimated as 30% of those reached) brings the total up to 200. This means that a likely maximum of 200 connected numbers must be tried, to result in 100 interviews. According to Telecom Australia, these ranges were allocated in the Albany area in 1987:

Albany 41-0000 to 41-8799 8800 numbers
Denmark 48-1100 to 48-1899 800
Mount Barker 51-1000 to 51-1999 1000
Porongorup 53-1000 to 53-1199 200

This is a total of 10,800 numbers, not all of which are necessarily connected to telephones. In the first sample draw, 200 numbers are needed, which is one in every 54. Think of this as 200 groups, each of 54 numbers, and from each group 1 number will be chosen.

We now choose a random number between 1 and 54, and the result is (say) 38. The first group of 54 runs from 41-08000 to 41-0053, and the 38th is 41-0037, which is the first random number to dial.

Using systematic sampling, each successive random number is obtained by adding 54 to the previous number (i.e. getting the 38th number in each group), for example:
+54 = 41-0091
+54 = 41-0145
+54 = 41-0199
and so on, up to 41-8785
when something different happens, as the next number after 41-8799 is 48-1100. So the group of 54 numbers which begins at 8748 does not end at 41-8802 (a nonexistent number), but jumps straight from 41-8799 to 48-1100, and ends with 48-1101. The next group of 54 runs from 48-1102 to 48-1155, and the 38th in this group is 48-1139.

It is tedious to do these calculations by hand, it is normal these days to use a computer to calculate numbers for random digit dialling. This can easily be done with a spreadsheet program.

When you have calculated the 200 numbers, the interviewing can begin. Attempts are made to do interviews at all eligible numbers. When the interviews have been finished, a table is made, like this:

Total numbers = 200
Not connected = 52
Connected = 148
Interviews completed = 87
Refusals = 21
Businesses = 30
Number exists, but no answer after 6 calls = 10

So 87 interviews were completed from 200 numbers, only 148 of which existed. Pro rata, to get the 13 additional interviews needed for the full sample of 100, another 30 numbers will be needed (about 8 of which will turn out not to be connected). Just to be on the safe side, though, it's better to obtain more than 30 extra numbers. Let's take 45.

As there are 10,800 numbers available, we shall get 45 of them by taking one in every 240. Taking a random number between 1 and 240, we get 109, which results in these phone numbers:
+ 240 = 41-0348
+ 240 = 41-0588
and so on.

To complete the survey, you can now ring these additional numbers.

Problems with systematic sampling for random digit dialling

Note that in the above set of phone numbers, every number selected is going to end in 8. This may not be a good idea: it's just possible that numbers ending in 8 may not be typical, having been assigned by the telephone authority in some particular way.

The systematic sampling system illustrated above will often produce such patterns in the numbers selected. Systematic sampling is the simplest possible way of producing a set of phone numbers for random digit dialling, and has been used here to show you how the process of random selection works. For a serious survey it would be dangerous to have a high proportion of the selected phone numbers sharing a common ending, and a more complex sampling system would be used, probably using a spreadsheet to generate random numbers within the number ranges known to be used.

Two ways in which modifications of the above sampling system could overcome the problem of patterns in the endings of numbers are:
(a) Instead of choosing a single random number, and taking the corresponding telephone number in each group of 240, choose a separate random number for each group of 240.
(b) Instead of using a group size of exactly 240 (the result of dividing 10,800 by 45) use a non-integer group size (e.g. divide 10,800 by 46 and get 234.78). Add multiples of this to the first number chosen, and ignore the figures after the decimal point in each telephone number thus selected.

For example:

41-0108 +234.78 = 410342.78 (ring 41-0342)
410342.78 + 234.78 = 410577.56 (ring 41-0577)
410577.56 + 234.78 = 410812.34 (ring 41-0812)

and so on.

Such a system will ensure (as long as the population size is not an exact multiple of the sample) that the telephone numbers selected will be equally likely to end in any digit.

Another problem with random-digit dialling occurs when you know a prefix exists, but have no information on which number ranges within that prefix are actually used.

In Australia, each prefix is followed by four digits, from 0000 to 9999, so there are potentially 10,000 numbers for each prefix. In some country areas, though, as few as 10 numbers of the 10,000 may be allocated. If you don't know which are the 10 numbers allocated, using random digit dialling in these circumstances is like finding a needle in a haystack. By ringing numbers at random, there may be only one chance in a thousand that you'll get a number that is actually connected. It's worth knowing in these cases that (in Australia, at least) it's usually the first 10 of the 10,000.

If the numbers are being dialled in ascending order, it's tempting for interviewers to assume that after about ten or so numbers have been tried, and none found to be connected, that no more numbers have been allocated for that prefix. This is quite often true, but before giving up on a large range of numbers, you should try to check with the telephone authority that the remainder of the range is not actually in use.

When you have chosen a sample using random digit dialling, interviewing should not simply begin with the lowest number selected, and work up to the highest. The danger is that the highest-numbered prefixes, (which probably correspond with particular geographical areas) may not be used, if you have selected more numbers than you'll eventually need. Therefore, the order in which the selected numbers are rung should be randomized, so that no bias is caused by unused numbers being concentrated in a few localities.

If you are conducting a telephone survey over a wide area, it can become quite expensive to make all calls from a central base. On the other hand, if all calls are made from one location, you have much better control over the workload, if any reallocation is needed. Quality control is also better when all the interviewing is done from a single office, as it's easier to make sure that each interviewer is using a standardized approach.

2. Telephone interviewing

Most of the time, telephone interviewing is not very different from face-to-face interviewing - but the lack of vision makes it harder for interviewers and respondents to communicate well, so extra steps have to be taken to compensate for this.

The rest of this chapter describes some of the peculiarities of telephone interviewing: call logs, work flow, and how interviewers need to adapt to invisible respondents.

Call logs

After experimenting with various ways of recording the progress of telephone interviews, we found that the Call Log is the simplest. A call log is a piece of paper: there is one for each telephone number to be called in a survey.

Call logs are an excellent method of managing work flow in a telephone survey. Interviewers have several heaps of call logs on their desks: one heap they are currently working from, and other heaps for each type of call outcome.

Here's an example:

CALL LOG Phone (0. . . .) . . . . . . . . . . . . . Interview. . .

Name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Directory page . . .

Try Day Date Start time Stop time Interviewer Result*



G = Gave up after 10 rings B = Business number

E = Engaged N = No English spoken

F = Fax machine or computer R = Refused immediately

D = Disconnected number S = Spoke to somebody in household

A = Answering machine L = Ring back later: fill in below

"Good [evening], is that [number]? My name is [name] from Bloggly Research. We're doing a survey of radio listening and your phone number has come up. I'd like to ask a few questions about radio listening in your household. I need to speak to the [oldest/youngest] [man/woman] [if youngest:aged 15 or over] who lives there. Is that person there now?"

If the selected person is not home, arrange to ring back later. Also arrange to ring back later if the selected person is home, but too busy to speak now.

Tick one box:
[ ] Interviewed immediately
[ ] Ring back about . . . . . am/pm on . . . . .day and ask for . . . . . . . . . . . . . .
[ ] Refused, because . . . . . . . . . . . . . . . . . . . . . . .

- Call logs should be quite small, because interviewers will have several heaps of them on their desks. A5 (about 15 by 21 centimetres) is a good size.

The above example of a call log is a large one, because it has the entire introductory script for the survey. This is easier for inexperienced interviewers, but it fills a whole A4 sheet of paper. More experienced interviewers can have the script on one piece of paper, and write the call details (day, time, and result of call) on a much smaller piece of paper or a card.

Before the interviewers receive these call logs, several items are written at the top:

There are several ways to find the respondent within a household, as discussed in the Sampling chapter. The above call log is designed to have YM, OM, YW, or OW written in the "Interview" space at the top right, showing whether to try to interview the youngest man in the household, the oldest man, the youngest woman, or the oldest woman.

If a call result is anything except the common listed outcomes (one-letter codes), a brief description is written in the Result column.

The interviewer writes his or her initials in the Interviewer column.

When the interviewer ticks the box labelled Interviewed Immediately, he or she picks up a questionnaire and begins the actual interview.

It's quite possible to have a call log printed as the first page of each questionnaire, but many telephone numbers don't result in interviews, so this method will waste a lot of paper.

If you want to improve your survey methods, the call logs can be entered into a computer (each log as one case) and analysed. Among the most useful items to enter are:

This information can be used to calculate response rates, check telephone charges. and so on.

Smaller call logs

If a questionnaire is very short - such as a screening survey for group discussions, it is best to use a simplified call log. Here's an example, from a screening survey I organized recently. It uses less than a quarter of a page, and can be used when a list of known people is to be contacted.

Ask for . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . at (08) . . . . . . . . . . . .

Call 1 2 3 4 5 6

Date/time . . / . . . . . . / . . . . . . / . . . . . . / . . . . . . / . . . . . . / . . . .

Result . . . . . . . . . . . . . . . . . .

S=spoke, B=busy, N=no answer, D=disconnected, CL=call back later

G=gone from address, AM=answering machine, W=wrong no., NE=no English

There are spaces for up to 6 call attempts here, In each date/time space, the interviewer would put something like 16/4.55 - meaning 4.55pm on the 16th of the month. Leave a blank line after this, for any special instructions if the interviewer needs to ring back - such as what time to ring back, and who to ask for.

The questionnaire continues:

"Hello, is that .... ? I'm ... and I'm ringing on behalf of the Oodnadatta Orchestra. I understand that you were a subscriber to their concerts last year - is that correct?"

1 Correct 2 Not sure -> other person 3 Wrong -> end -> recheck no.

"We're doing some research on the preferences of our audiences. Would you mind answering a few questions, please? It will take about 3 minutes."

1 Will answer now 2 Later -> make appointment 3 Refused -> end

The call log ends here, and the main questions can now begin.

Not sure -> other person means that if the person who answers the phone does not know about any previous contact with your organization, the interviewer needs to speak to somebody else in the household - which may involve an appointment to call back later.

Though this call log format is very simple, it provides almost as much information as the full-page version, and is simpler to handle. However, because it is part of the questionnaire, it will waste a lot of paper unless almost every household contacted is eligible to participate in the survey.

Working with call logs

The work flow with this system is simple. At the beginning of each shift, an interviewer takes a heap of call logs in the "Ring this number again, some time" category, and works through them. When a number is finished with, its log goes into either the "Successful interview" heap or the "No interview" heap.

If a number is engaged, this means that somebody is home, so the number should be tried again soon - normally between 10 minutes and half an hour later.

If there's no answer, nobody is home, so it's no use to try that number again within a few hours - unless the call was made at a time when many people are arriving home. After several successive "no answer" results, try ringing at a different time of day.

Normally we wait for 10 ringing tones before giving up: about 30 seconds. However some people take a long time to reach their phone: specially those living on small farms, and old people who cannot move quickly. If you have given up after 10 rings on 3 successive attempts, try waiting for 20 rings. If there's still no answer, you can record the result as G20 instead of G.

The standard call log has spaces for up to 6 attempts to ring a number. This is usually plenty - few numbers require that many attempts. We usually stop after 5 attempts, and only make a 6th if there's an appointment to be kept.

There are about a dozen possible results from trying to make a phone call, and most of these are shown on the example call log on page ##. All those result codes can be grouped into 4 categories:

These will also correspond with heaps the interviewers have on their desks. Now you can see why it's a good idea to keep the call logs as small as possible, as the interviewers also need enough space for a heap of blank questionnaires, a heap of completed questionnaires, a set of survey instructions, and a phone!

The heap of logs for appointments to call back should be kept in order of the appointment time. It's easy to overlook appointments, so if several interviewers are working together, one should be in charge of appointments, and look after all the call logs with appointments pending.

Introductions in telephone interviews

(1) Avoiding refusals

The first few seconds of a telephone interview are vitally important in communicating the information needed. Imagine you're a respondent. Your telephone rings. You pick it up, and a stranger's voice begins to explain something about asking you questions. Immediately, you are wary. What does this stranger want from you? Now they are asking which person in your household last had a birthday. They are talking in a strange kind of way, as if they are reading aloud. You are suspicious. Why do they want to know that? Is this a new kind of sales pitch? Why do they want to speak to your daughter? Now they are talking about a survey. Last week, a friend told you that somebody rang him up, mentioning a survey, but really trying to sell insurance. Another friend told you about a survey phone call: the interviewer said it would take "a few minutes" but it lasted almost an hour.

You decide you want nothing to do with this survey. "We're not interested," you say. "Goodbye."

On the other end of the telephone line, the interviewer notes down yet another refusal.

In the richer countries, about one telephone call in three ends with a refusal. Most of these refusals seem to arise through suspicion, or bad experiences with similar calls in the past. Commercial firms often telephone people to try to sell things to them. This can become so common, and so annoying, that people defend themselves by having answering machines and unlisted numbers. (They don't know that an unlisted number is no defence against random digit dialling.)

Where people have little experience of sales-related phone calls, there's no problem with refusals. The first telephone survey I organized in the country areas of Australia was in 1985. Most respondents had never had their opinions asked before, and were pleased to be contacted. With over 2,000 interviews, only 6 people refused. In 1998, we did another survey, in the same areas. This time, 30% refused. What was a novelty in 1985 had turned into a nuisance by 1998.

The best way to overcome this high refusal rate is by establishing a genuine dialogue with respondents. You have perhaps 30 seconds to convince them that your survey will benefit them in some way.

Approaches we have found successful are:

1. If the respondent has had any previous contact with you, establish this immediately. For example "Could I speak to [name of person you want]?" When the person you want is there, say something like "I'm ringing on behalf of the Oodnadatta Orchestra. I understand that you were a subscriber last year - is that right?"

When the respondent remembers some prior dealing with your organization, refusal rates are usually negligible.

2. When you are surveying the general population, and your organization is not well known, refusal rates are likely to be high. In this case, two quite different approaches have both worked well for us:

2a.Spend about a minute describing your organization, and the purpose of the survey you are doing. Only then, tell them that you'd like their opinions.

2b. Immediately start to question the respondent, with no preliminary explanation. After the first few questions have aroused their interest, you can explain your purpose. This is almost a way of tricking people into answering questions, It works well, but for ethical reasons, I don't recommend it - specially if any answers could harm the respondent in any possible way, or respondents might regret giving any answers.

3. Assure respondents that you are not trying to sell them anything.

4, In Australia, we found lower refusal rates when we didn't mention the word survey. (There are so many pseudo-surveys around.) Instead, our interviewers said "I'd like to ask you a few questions."

5. Some interviewers have much higher refusal rates than others. Those who speak cheerfully, quickly, and confidently do best, while interviewers who speak slowly, hesitantly, or seem to lack confidence have much higher refusal rates.

6. Most refusals happen in the first minute of a call. Therefore, the first few questions should be interesting to the respondent, not ask for any confidential information, and not be difficult to answer.

Refusal conversion

When refusal rates are high, and are likely to have a serious effect on the accuracy of a survey, refusal conversion can be used. This means that interviewers who have achieved the lowest refusal rates re-telephone people who have refused to participate in the survey, and try to find out why they refused, and convince them to co-operate. The interviewers explain these points:

1. The respondent can nominate a convenient time.

2. The interview will take no longer than X minutes, and the respondent can feel free to stop after that. (This is a good incentive to tell the truth about questionnaire length.)

3. The respondent's individual data will not be used for any other purpose.

4. Co-operation will benefit the society/country/etc, and (indirectly) the respondent.

5. Refusers often say "It's no use talking to me: I don't use that service." The answer can be that "we need opinions from all types of people. As you don't use the service, some questions won't apply, so it won't take long at all."

6. If the budget permits, financial (or other) incentives can be offered.

Using these refusal conversion procedures can halve refusal rates, in many situations.

(2) Informing respondents

In countries where telephone surveys are very frequent, refusals are the main problem. In countries where telephone surveys are rare, the main problem is not willingness to co-operate, but respondents' lack of understanding of the procedure: "Why does this strange man want to talk to the youngest female in my household, not counting any children under 15?"

Respondents should be told that they may decline to answer any question. You should also tell them: