Document(s) 6 of 10

4.1 Guiding principles for data collection in the Acacia Initiative

Four main principles guide data collection in the Acacia Initiative. They stem from the goals and structure of the initiative but are also useful objectives for telecentre projects other than those funded by IDRC and its partners. These main principles are as follows:

The information needs of the various telecentre stakeholders should be built into the data-collection design  — These stakeholders are likely to include leaders and institutions at the community level, telecentre owners and operators, private-sector investors, national agencies responsible for telecommunications, and international agencies and donors (see Table 2).
Learning opportunities for stakeholders should be part of the data-collection design — An important way to strengthen learning is to have the stakeholders participate in the data-collection and data-interpretation processes. At the community level, the researchers should use participatory research methods and, at all levels, have an effective stakeholder information process to inform the stakeholders of the results and provide them with opportunities to discuss the significance of the results and have input into the design of future rounds of data collection.
Approaches should facilitate comparisons of results across telecentre projects and between countries — Two objectives of the Acacia Initiative are to better understand the interplay between local telecentre operations and national policies (vertical links) and to identify the determinants of success across telecentres in different situations with different operational histories (cross-sectional comparisons). These comparisons can only be valid if the data collected from one telecentre project are reasonably consistent with those collected from all the others. Without consistency in the definition and selection of samples and in the methods and instruments used to collect data, researchers will have little chance of making meaningful comparisons between pilot countries.
Data sets should be stored in a common database or data repository — The current Acacia telecentre projects will collect baseline data on the pretelecentre situation or in the initial year of telecentre operations, or both. Some projects are designed to collect data to measure changes and impacts within the first 1–3 years of telecentre operations. These data sets have significant value beyond the objectives of the individual projects and should be properly maintained within a common facility so that researchers can use them to answer new and different questions that emerge in the future. Baseline data are also critical to any future longitudinal studies on telecentres and their communities. Consideration of the unanticipated data needs of various stakeholders, including future stakeholders, underscores the importance of such a shared database for African telecentres.

A number of practical implications for data collection flow from these four guiding principles. Although they will be discussed under the various research methods in section 5, they are worth highlighting here:

Data disaggregation — If data are to be combined and reanalyzed in various ways, it is important that they be disaggregated as much as possible when collected and initially recorded. This provides for maximum flexibility in future analysis.
Multiple methods — As we shall see, each method has strengths and limitations, and the variety of information needs of the telecentre stakeholders at various levels is a strong reason for using several methods with each pilot project, rather than relying on any single approach.
Multiple samples — A similar practical consideration favours the use of several sample groups in the telecentre studies: telecentre users provide the most direct and relevant information on telecentre performance, but they will be unable to provide adequate data on the impacts on the community or indicate whether the telecentre is responding to community needs. It will be advisable and probably more cost-effective to have different research instruments for different sample groups. Telecentre operators, for example, can most efficiently monitor certain aspects of use patterns and equipment performance within the telecentre environment, and it would be more costly and less accurate to try to collect these data through a community household survey. However, a telecentre operator would be unsuitable for monitoring the attitudes of nonusers or the social and economic impacts on families or organizations.
Data-collection methods appropriate to data needs — As will be seen in section 5, different methods are most useful and effective for different types of data, and the type of data needed should determine the selection of methods.
Degree of intervention and local participation — Some telecentre-project evaluations are more “external” than others, both in terms of who is undertaking the evaluation and in terms of the role that the evaluation itself is expected to play in community change and development. The active role of the data-collection process as a change agent and the degree to which community members will be collecting the data on their own community will be considerations in selecting the most appropriate method for collecting data.
Methods appropriate to the level of training of the field researchers — Some methods, such as group techniques (including focus groups and Delphi surveys [see section 5.6]) and advanced question techniques (including some attitude surveys or personality tests) require more training and experience than straightforward observation schedules or structured questionnaires do. Ethnographic studies require considerable training and commitment. The methods selected should take into consideration the qualifications and training of the field staff who will be the primary data collectors.
Time and cost implications of data-collection methods — This is perhaps the most obvious of the practical issues of data collection, but it is worth underscoring because almost all data-collection exercises and evaluation studies lack the resources they need or would ideally have. It is particularly important for the Acacia Initiative to keep such costs in mind if future follow-up data collection is anticipated, as the cost of replicating a survey or case study can become an impediment to obtaining valuable longitudinal data.

4.2 Issues related to sampling

This section suggests a number of sampling issues to consider in designing an evaluation study and thus points to some of the strengths and weaknesses of various types of samples for telecentre surveys. It is not intended as a primer on sampling strategies, which is a complex subject. For this, the reader should explore some of the suggested reading in the bibliography.

4.2.1 Sampling frame

The sampling frame is a major determinant of the extent to which a sample is representative of the population under study. A frame is perfect “if every element appears on the list separately, once, only once and nothing else appears on the list” (Kish 1965, p. 53). Sampling frames are of two general types: lists, such as electoral registers or the membership of an organization; and sets of locations on maps (such as townships or rural communities). In most cases, the sampling frame is imperfect: it has missing elements, inappropriate listings, or duplications. Kish (1965) provided a good technical discussion of frame problems.

Researchers conducting the Acacia telecentre studies may have no up-to-date or accurate lists of community members or households for designing a household sample of the community. The best frames available may be lists of school students, utility customers, and members of local organizations. However, each of these lists will have built-in biases or missing elements that may be significant enough to make it unsuitable for sampling the community as a whole.

If no adequate map is available to show locations of houses, the researchers may have to make their own sketch map or see if there are any airphotos that they can use as the basis for one (with field checking to update it).

An accurate sampling frame for use at the community level is probably the most difficult to obtain. It will be easier to sample telecentre users if the telecentre keeps a record of all users, because this becomes the sampling frame. The key issue here for researchers is to know how to recognize errors in the sampling frame they use and to seek to compensate for them, such as by using disproportionate sampling fractions or screening (see below).

4.2.2 Unit of analysis

The key to defining the unit of analysis is to find the locus of decision-making for the behaviour under study. Who makes the decision to use a phone or seek a job with an updated résumé prepared at the telecentre? Is the decision made by the individual user, the household, or an organization? And who pays for the telecentre service: the individual user, the household, or the organization? Clearly, it depends on circumstances.

Do some aspects of telecentres operate at the group level, having impacts on community identity or local innovation, for example? The appropriate unit of analysis is not always the same for every aspect of telecentre behaviour, although in practice the survey will have a uniform unit of analysis, usually an individual or a household. In some situations, however, researchers may find it more appropriate to use a local organization as the unit of analysis.

Another approach to selecting the unit of analysis is to take an event, such as a visit to the telecentre, and analyze the visitors or users and what they do during each visit. This is clearly useful for analyzing the performance of the telecentre and its financial sustainability. The key questions in this case will concern the characteristics of the visits: services used, time spent, revenues gained; and the characteristics of the individuals: satisfaction with service, new or repeat user, etc.

4.2.3 Types of sample

In an ideal world, most studies would aim to obtain probability samples, in which every element (person, household, or event) has a known, nonzero probability of being selected. And most statistical inferences about means and variances and regression coefficients are based on the assumption that the sample is a simple random sample. However, many studies — probably most of those undertaken in African communities — do not obtain probability samples. Of necessity and practicality, they adopt another strategy to achieve acceptable accuracy at an acceptable cost. Researchers have a number of alternatives to probability samples, and these can also be part of a good research design.

One common strategy is to use judgment samples in selecting the first and second stages of a stratified sample, such as communities and sections of communities, and organizations within the community. This procedure has considerable validity. As Kish, the “guru” of survey sampling, said,

If a research project must be confined to a single city … I would rather use my judgement to choose a “typical” city than select one at random. Even for a sample of ten cities … I would rather trust my knowledge. But I would raise the question of enlarging the sample to 30 to 100 cities. For a sample of that size a probability selection should be designed and controlled with stratification.

— Kish (1965, p. 29)

Quota sampling is another approach used when probability sampling is impossible or its cost is too high (several Acacia telecentre evaluations have already used quota sampling). In quota sampling, the number of individuals or households in a set of subclasses is estimated, and field investigators are assigned a quota of interviews or observations to make using controls such as geographic location, age, gender, or group membership. The controls should be manageable by the field worker who has to fill the quota.

Severe problems with bias can occur in quota sampling, both in the selection of the controls (they may not be the relevant ones) and in the freedom given to field workers to select the sample. The more freedom the field worker has, the more likely it is to cut survey costs but also to introduce bias. One type of bias — stemming from the tendency of interviewers to select, within quotas, people who are all similar — can lead to an underestimate of the variability within a population. For example, the interviewers may be asked to select a quota of people in a group of low economic status. Although the group of people selected for the sample may be expected to be representative of the whole group, the interviewers’ tendency to select similar people may lead them to undersample some segments.

Knowing a sample carries a risk of bias is not the same as knowing it has, in fact, a bias. You can, in particular cases, obtain enough information to test for bias, but quota samples are difficult to compare with probability samples, because one usually has no means to test the reliability of a quota sample. Nevertheless, despite the problems inherent in using quotas, it is sometimes better to have a quota sample than, for example, no sample or a sample obtained at an unreasonably high cost.

4.2.4 Stratification and multistage sampling

Stratification involves the division of the population into strata, or subgroups, and you sample separately from each stratum, using if you wish, different sampling weights or even different sampling procedures. A common reason for using stratification in research in developing countries is that maps and list sampling frames are available for urban areas but unavailable for rural ones. So researchers use two sampling procedures (Bilsborrow et al. 1984). Another reason is that the populations in the different strata are of particular interest and are considered for separate analysis. For example, in the Community Information, Empowerment, and Transparency (CIET) telecentre study in South Africa, communities that had a telecentre and those that had one planned for the near future were treated as separate strata (Andersson and Pascual-Salcedo 1998).

The total variation in a population equals the variation across strata plus that within strata. For example, urban and rural areas differ in average household distance from a telephone, and access to a phone also differs within urban and rural areas. Stratified sampling takes the difference between the strata out of the calculation of total variation. Thus, the general objective of stratification is to arrange the strata so that they differ as much as possible from each other but contain populations that are as homogeneous as possible. To achieve this goal, the variables that distinguish the strata are chosen to be closely related to the survey subject, as this eliminates the between-strata variation from the total variation.

Another consideration in using stratified samples is determining the number of strata to select. One should have, at a minimum, two primary sampling units per stratum (for example, residential blocks), and it is generally better to have fewer strata constructed using several variables (for example, rural–urban and distance from a telecentre) than many strata structured according to one variable.

One advantage of having a stratified sample is that the strata can be weighted. Ideally, one selects a higher proportion of units (for example, households) in the strata where the variance is greater or the cost of obtaining the sample is less. For many studies, this means oversampling in urban areas, where variance is usually greater and it costs less to collect the sample.

One of the main advantages of multistage sampling is that it can dramatically reduce the cost of field operations. In the first stage, the researchers can use an existing frame, such as a map or census, to select areal units and then do the more costly mapping operations only for the areas selected as the primary sample units. Most national surveys in developing countries have been multistage ones.

At some point, multistage sampling involves “cluster sampling,” usually at the second-to-last stage. In cluster sampling, the units of analysis are clusters of respondents, such as all households in a city block, a section of a village, all members of a household aged more than 15 years, or all children in a secondary school. Clustering sampling produces huge savings in field costs. Choosing the cluster size involves two types of consideration: it should match the organization of field work and the survey objectives; and the larger the cluster, the larger the sampling error (because the farther the sample is from a random or probability sample). More discussion of these considerations is given in Bilsborrow et al. (1984) and Kish (1965).

4.2.5 Finding the needle in a haystack

One of the challenges in sampling design is to capture “rare elements” in the sample in sufficient numbers and at the lowest possible cost. This is a common problem for researchers in developing countries, where good sampling frames are less commonly available to use in identifying rare elements. For example, a telecentre survey may wish to sample households with mobile phones or computers or those with experience using a telecentre. The sample pool may be fewer than 10% of all households in the survey area. Clearly, evaluators conducting a survey based on probability sampling will spend 90% of their efforts collecting data on households outside the interest of the study.

Kish (1965) identified eight ways to find rare elements. These include stratified sampling with disproportionate sampling fractions and multiphase or sequential sampling. Multiphase sampling involves the selection of elements (respondents) from a larger sample: the first phase acts as a screening process and, in the second, more contained phase, the researchers can use probability sampling at a reasonable cost.

Another approach is to use tracing techniques to locate rare elements or respondents. This is common in migration surveys, in which researchers first identify migrants through their original households. In the case of telecentres, all users over a certain period might be “traced” back to their households, and this would delineate a survey sample.

4.2.6 Sample size

One of the most important decisions in designing a survey is choosing the sample size. Choose too large a sample, and you will spend more money than necessary on data collection and processing; choose too small a sample, and you may end up with inclusive findings and poor credibility. There are statistically valid ways of determining the sample size, depending on whether the analysis will use simple or complex statistics (Kish 1965).

An important consideration is the “crucial subgroup.” This is the group — frequent telecentre users, for example — from which the survey must obtain enough observations to result in reasonably accurate statements, such as “frequent telecentre users have higher incomes and higher education levels than occasional users or nonusers.” If the analysis will come from only a part of the sample, then the sample size has to be increased significantly to maintain the level of accuracy.

Another approach is to consider the sampling error for the difference between two groups on a particular variable that is important. Assuming that each group — nonusers and occasional users of telephone service, for example — constitutes about 30% of the total sample and that about 50% of nonusers and 56% of occasional users are male, then to show that the 6% difference in gender composition between the groups is significant, one would need a total sample size of 2300 (Lansing and Morgan 1980)!

In the end, cost and efficiency determine most sample sizes, and these considerations tend to result in smaller samples, which are less robust when complex statistics are applied to them. The CIET baseline survey in South Africa, in which 14086 adults in 12472 households were interviewed, is one of the few telecentre surveys with samples large enough to withstand major statistical manipulation (Andersson and Pascual-Salcedo 1998).

4.2.7 Sample frequency

Researchers repeat surveys over time to collect longitudinal data, and the length of the interval between surveys will depend on the nature of the data and on the costs and time required for each survey. If the objective is to measure a trend over time, the frequency of repeat surveys may be more than if the objective is to determine overall impact in 5–10 years. Researchers should consider the expected rate of change. For example, evaluators might expect the introduction of a telecentre to lead to changes in travel patterns within 1 year, whereas a change in employment rates might be expected to take 3–5 years. Another reason for resurveying is to gauge the impact of a specific intervention, such as the opening of a telecentre.

4.3 Issues related to surveys

The survey is likely to be the most common method used in the Acacia Initiative to study the use and impacts of telecentres. As discussed above, evaluators can use surveys to measure telecentre performance (and user satisfaction) and to evaluate the broader impacts on the community, depending on the target sample of respondents. Surveys are particularly susceptible to what Kaplan (1964) called “the law of the instrument,” which is illustrated in the story of the child who, given a hammer, will discover that a great many things need pounding.

At the outset of a study, a number of decisions are made about the design of the survey, selection of respondents, and procedures to follow in the field and in the analysis. Some of the issues surrounding these decisions are highlighted here to help the research team think them through. Again, this section is not a textbook on how to carry out surveys (the bibliography gives some suggestions for reading on both the practical and the theoretical aspects of designing and conducting surveys). The purpose of this section is to focus attention on some of the issues researchers need to think through as they design their methodology, so that they do not hammer away at things that don’t need pounding and that, when they do pound on something, they hit the nail.

4.3.1 Surveys for various purposes

The main purpose of most social surveys is to explain (or to contribute to the explanation of) certain social or economic phenomena. In the case of telecentre surveys, the purpose is primarily to explain phenomena relating to patterns of behaviour in the use of information and communications, both in the telecentre and beyond. The explanations sought may fall under the deductive model (a behaviour or event is explained by deduction from other facts) or what Kaplan (1964) called a “pattern model,” in which the reason for a behaviour or an event is known if it fits into a known pattern or system. Researchers use surveys under these models of reasoning to find out why people do something or why something happens. These models are used to

Support predictions about behaviour now and in response to policies, events, and circumstances in the future;
Provide input into simulation models on aggregate behaviour and system changes; and
Evaluate the performance and impact of events, organizations, policies, and technologies.

The use of these methods for any of these purposes raises issues in the collection of survey data: whether controls are needed, at what level the sampling regime will be statistically valid, whether a single survey will suffice, and how the questionnaire or observation schedule is to be administered. For the telecentre studies, the most important purpose for conducting surveys is likely to be evaluation.

4.3.2 When surveys are not useful

It should be recognized that in some situations, surveys are not useful. These situations relate to the purpose of the research, the level of data aggregation, and the nature of the phenomena studied. The most common situation in which surveys are misused is when researchers work without a clear hypothesis or a specific issue to guide and structure their survey, beyond a set of “interesting questions” (the need to identify research questions and define the explanatory system was discussed in section 2.1). However, surveys are also inappropriate for testing single elaborate hypotheses. In general, they are best suited for choosing between alternative hypotheses (Lansing and Morgan 1980).

Surveys should not be undertaken if the interviewers need to deceive the respondents about the purpose of the survey or if the study focuses on illegal behaviour, such as malpractice among telecentre operators (see section 4.3.7). Surveys are not good for estimating aggregated national data, particularly where the distributions may be skewed. And, as indicated in section 4.2.5, they are also ill-adapted to the study of rare phenomena.

4.3.3 Alternatives to community surveys

The Acacia Initiative can obtain some (but not all) of the data of interest by surveying more targeted samples, such as telecentre users; telecentre operators and staff; leaders and staff of other institutions, such as a health clinic or school; and leaders and members of local groups, such as women’s and youth groups, chambers of commerce, and craft cooperatives. These samples are less costly to survey than a representative sample of households in the community or telecentre catchment. Researchers can also more easily trace respondents in these samples for reinterviews. Clearly, it is of both theoretical and practical interest to survey these groups whenever possible and appropriate.

However, if several different subsamples are selected, then some common data should be collected across all sample surveys to measure how the groups differ on key socioeconomic and behavioral dimensions. In addition, it is best if a representative sample of households in the community is also surveyed, which can inter alia provide information on the proportion of the community represented by the subgroups and how far the behaviour of each subgroup influences or explains the patterns found in the broader community.

4.3.4 When and why you need community-level data

Community telecentres are, by definition, a community service. Telecentre users are, by definition, individuals. These individual users may visit the telecentre on behalf of other members of their households (or for the household as a whole) or a group or organization. The decision to use the telecentre is sometimes made by a group. Thus, the individual’s purpose in using the telecentre, the money he or she pays for the service, and the outcomes of the visit are best understood at the level of the household, group, or organization. Consequently, research on the use and impact of telecentres must consider more than one explanatory level in its research hypotheses and instrument design. These levels include that of the individual and those of the household and organization or group to which the individual belongs.

Another important level is that of the community. Studies of various social and economic behaviours (such as decisions to migrate, to invest, or to use services such as family planning; the propensity of farmers to adopt innovations) have shown that community-level and individual-level variables have independent effects (Bilsborrow et al. 1984). It would be surprising if patterns in the use of information and communications were any different. Thus, if the researchers aim to explain the phenomena of information and communications and their impacts on people and the community, they will need community-level data, too.

Communities — their geography, economy, demographics, and services — provide “opportunity structures” for individuals and households and can act as major determinants of social behaviour (Ritchey 1976). Conversely, these opportunity structures are, themselves, altered by the communication behaviour they engender, including the long-term sustainability of the telecentre. Researchers also need to use community-level variables to explain why people do not behave in certain ways (for example, communicate with family or use the Internet).

For all the above reasons, researchers should consider the community in the conceptual model and the data collection for telecentre studies. The necessary data will range from baseline information on the services in the community, to demographic and socioeconomic characteristics of the population, to information on norms and patterns of behaviour.

4.3.5 The theory of interviewing

The purpose of a theory of interviewing is to guide the researcher in developing a technique to obtain high-quality data at the least cost, both to the researcher and to the interviewee. An interview is a social transaction in which information is exchanged between people who differ in their reasons for engaging in the exchange, in their levels of knowledge on the subject, and in their perspectives and biases. A number of studies have shown that the views of the interviewer on the subject matter influence the way she or he records a respondent’s answers and that the interviewer’s perceptions of a respondent can introduce even more interviewer bias (Hyman 1954). Consequently, a body of best practice in survey interviews has evolved to minimize interviewer bias and to measure and control for it (Hauck and Steinkamp 1964; USCB 1968).

A second focus of the theory of interviewing is the respondent’s motivation for participating in a survey. Generally, a respondent’s initial motivation is weak. One way to encourage people to participate is to build into the survey a process for providing feedback to respondents on how the group responded. More generally, motivation is related positively or negatively to any of three factors: the stated purpose of the study, who is sponsoring or carrying out the study, and the social situation that the interview presents.

The last factor is the most influential. It is generally accepted that in most circumstances the interviewer should be as similar as possible to the respondent in race or ethnicity, local language or dialect, gender, age, and status. Also important is how the interviewer conducts the interview to ensure that the respondent understands the questions and answers them fully and to avoid nonresponse (see section 4.3.7). In North America, women and younger people are found to make the best interviewers for most topics. The usual interpretation for this is that the most successful interview situation mirrors that of the experienced teacher (the respondent) teaching the student (the interviewer). The accumulated body of evidence on interviewing shows that survey results are only as good as the interviewers who collect the primary data, so it is worth paying attention to their training and performance.

4.3.6 Dealing with nonresponse

The key issue with nonresponse is whether the nonresponders are similar to the rest of the respondent group or systematically different (for example, from a different ethnic group, illiterate, opposed to the telecentre). One strategy for reducing nonresponse as much as possible is to choose a better survey method. Different methods of surveying similar populations have very different characteristic nonresponse rates. In the United States, for example, the nonresponse rate for census surveys is usually less than 5%; for research surveys that use personal interviews or telephone interviews, 10–25%; and for mail surveys, up to 90% (Lansing and Morgan 1980).

A second strategy is to follow up on nonresponders, either with another personal visit or, where appropriate, with a telephone interview or a reminder in the mail. Or the nonresponders can be replaced with other respondents (the “go-next-door” approach), although this procedure has built-in biases (are people who are likely to be “at home” next door similar in important characteristics, such as employment, to those not at home?). It is important that the instructions to interviewers be very clear about what procedure to follow if a designated address turns out not to be a house or is vacant or the occupants are consistently not at home or refuse to be interviewed and how the response in each case is to be designated, as this affects the measurement of error from nonresponse.

A different strategy is to try to find out whether the nonresponders differ from respondents and allow for that in the analysis. The sampling frame may help here. But, more commonly, we know very little about nonresponders in most surveys, and, for simplicity’s sake, we assume that they are not significantly different from the respondents in the sample. If the nonresponse is limited to one or a few questions, analytical techniques to substitute information from the rest of the sample may be helpful (for example, computing household income on the basis of like responses to other questions, such as ownership of household goods).

4.3.7 Mortality in longitudinal surveys

In surveys that reinterview the same respondents or panels of respondents, attrition (“mortality”) of the sample group is inevitable. Respondents may refuse to be interviewed a second or third time. More likely, they have moved away, are unavailable, or cannot be traced. In the United States, typical panel mortality rate for surveys is 10% for each survey “wave” done 6 months to 1 year apart. This means that after five surveys the sample may be only half as large as it was originally (Kish 1965). The evaluation team needs to take account of these cumulative losses in its initial survey design and decisions about sample sizes. The problem becomes more serious if the respondents who are lost from the study differ systematically from those who remain (and several studies indicate that they seem to).

4.3.8 Ethics of interviewing

Evaluators need to consider two related ethical aspects when designing a survey and elaborating the procedures for the field and for data analysis. The first is to ensure that the interviewers understand the meaning of informed consent and their obligation to obtain it from the respondent before beginning an interview. An interviewer must tell the respondent about the purpose of the survey, who is conducting it and for whom, and how the respondent’s responses will be used. In some situations, it is accepted that informed consent is given by the head of a household, the parent of a child, the teacher of a class, or the leader of a group (for example, “headman”). After requesting and receiving the consent of an authority figure, the interviewer should make every effort to ensure that the individual respondent also understands and gives his or her consent.

Normally, the results will only be reported for groups of respondents; their individual responses will not be identified. If a particularly descriptive or apt response is quoted, it should be done in such a way that an individual respondent cannot be identified. If possible, the interviewer should ask for the respondent’s permission to use the quote.

If interviewers take the names and addresses of respondents, these should be recorded separately from their responses. In practice, this means that the interviewer records a respondent’s personal information on a separate sheet, linked to the responses only through an identifier number, and the evaluation team does all analysis of the responses under the identifier number. Access to the personal information is restricted to supervisors and others who need to know, especially for follow-up interviews or feedback.

4.3.9 From questionnaire to analysis

The research team will design and follow a set of procedures for processing the survey data from the point at which the data are obtained on the questionnaire to the point at which analysis can begin. The steps will include checking that the questionnaires have been properly completed, assigning identifier numbers, dealing with nonresponse, coding responses, cross-checking the coding, data entry, checking for errors and consistency, and generating new variables.

For many researchers, this part of the survey process is usually the one they most dislike and neglect. However, the interview phase is interactive and interesting, despite the practical problems it may present. Analysis is rewarding because patterns begin to emerge in the data, hypotheses are tested, and results begin to make sense. But in between, when the data are being quantified and prepared for analysis, the doldrums can set in. It is also, therefore, a time when errors creep in.

A generally recommended strategy is to postpone to a later stage any work that can be postponed, on the rationale that as the process progresses from respondent to computer, the next stage is more specialized but less expensive than the previous one. As a general rule, the survey design should not have the interviewers trying to code in the field. It is cheaper and more accurate (and verifiable) to have coders in the office and let the interviewer concentrate on the respondent’s answers.

On the other hand, CIET Africa in Johannesburg has found that locally trained interviewers can also be trained to do coding and data entry and that they bring a new enthusiasm to the data-processing tasks. Working in pairs, they act as cross-checkers, and their experience as interviewers gives them valuable insight into the data (Andersson and Pascual-Salcedo 1998). This approach supports the role of Acacia as a learning experience for local participants and is to be recommended. It not only produces good-quality work but also leaves a repository of new skills in local organizations, which will benefit the communities after the project has ended.

Document(s) 6 of 10