When a survey is conducted, the researcher wants the sample to represent with 100% accuracy the population from which it was chosen. Unfortunately, this never happens because there are many uncontrolled variables in the world that influence the sample and lead to biases. Researchers’ greatest concerns are related to nonresponse rates and the biases that come with them. When sampling inanimate objects like, medical records of persons, the researcher will receive 100% response rate. However, this hardly ever is the case when sample surveys involve human respondents.
Nonresponse bias arises when the sample units that the researcher failed to measure, differ from the units that were actually measured in that particular sample. By definition the bias arises from sampled units that do not respond. Nonresponse may occur because respondents are not at home at the time of the survey or because the questionnaires never reach their residence. Some people may be unable to respond because they are too ill, uneducated or do not speak the same language as the researcher. There are also those who relentlessly refuse to take part in surveys.
Nonresponse is classified into unit and item nonresponse. Unit nonresponse appears when the sampling units in the selected sample do not participate in the survey. Item nonresponse arises when for some units who agree to participate information obtained is incomplete, because they refuse to answer specific questions, such as on their income (Korinek, Mistiaen et al. 2005). The response and nonresponse rates, as well as the associated bias, vary depending on the type of the survey such as, telephone survey, mail survey or personal interview. Nonresponse bias is a nonsampling error and thus difficult to estimate.
As mention at the very beginning, nonresponse bias is in a way a researcher’s nightmare. That is because nonresponse is very problematic. ”Low response rates can lead to inaccurate estimates and large standard errors”(Groves 2007). Hence conclusions about the population for which the researcher is interested in will be incorrect. Nonresponse will develop nonsamlping errors but it will also increase sampling error because it reduces the sample size. An example of the impact of nonresponse bias is an article published in December 2006 in ‘The Times’.
The article was about a study of the weight of British children commissioned by the Department of Health, which had been prevented because many families refused to allow their children to participate. The estimates of the weights of the children aged 10 or 11 was underestimated because the parents had the right to deny to allow their children participate. This created a bias towards those who were less heavy because those whose children were heavier were more likely not to take part in the study. Still though, the inference about the bias was made because there was additional available information about the population.
It was noted that, areas with a higher response rate recorded more children as obese than areas with poorer response rate (Floyd J. Fowler 2002). To overcome these types of problems researchers use several different techniques. Well designed surveys focus on minimizing nonresponse prior to implementation. The way the respondents are selected, the objectives and the importance of the survey should be clearly stated so that the respondents can appreciate and hopefully cooperate. A statement assuring the confidentiality of the respondent’s answers, as well as a contact number for help is a good approach to gain their trust.
A well structured questionnaire written in familiar language can help increase response rates. Well trained and motivated interviewers are a necessity to increase responses. Before the survey is implemented a pre-test can reveal unanticipated problems, thus further adjustments can be made before the study is put into action. It is proven that multiple contacts with the respondents enhance response rates. A good approach would be to first inform the selected respondents about the survey by mail, email or a telephone call. Then the survey will take place, either by sending the respondents a questionnaire or by interviewing them.
Then for those who did not respond a reminder will be forwarded. For those who remain nonrespondents a second reminder will follow and for those who still adamantly not respond a different form of contact can be attempted. Of course it is reasonable to expect that respondents will be more willing to comply if they have something to gain from this survey; a reward. Rewards can be intangible or tangible. An intangible reward applies in situations when the survey is very special or important that it may be a privilege to take part in. However the most effective reward is the tangible one.
It can either be monetary or non-monetary; an IPod for instance. Rewards in advance are proved to have a higher response rate. On the other hand, people that will only respond when compensated may not be representative of the selected population and thus produce bias. The way and the timing each step is implemented to reduce nonresponse varies depending on the method of contact. In addition to the above techniques there a few statistical techniques that aim to correct nonresponse bias after the results of the survey have been obtained.
When the direction of the bias is know it is possible to replace some of the obtained figures from trusted sources. For example, if a study to measure the average income of households was underreported, the researcher could use the national accounts to replace average incomes. However, there are limitations in this approach. It is assumed that the trusted source can indeed provide a valid estimate and that the distribution of the two data sources is indifferent. The most popular statistical technique to correct for unit nonresponse bias is weighting adjustments.
When a sample’s data is underrepresented, then the sample is given more weight to become more proportional with the population. For example, a researcher in the United States might be concerned for the rate of personal crimes. The National Crime Victimization Survey for urban areas has an 85% response rate, but the response rate in the rest of the country is known to be 96%. The number of urban persons in the in the respondent data sheet is underrepresented. Therefore, when calculating the mean rate of personal crimes, the weighting adjustment would be to give greater weight to the urban respondents (Groves, Fowler et al. 2004).
In the particular example the identification of the groups was easy; urban and not urban-population. This is not always the case when dealing with more complicated studies. In addition, information is needed about the distribution of the population in order to be used as a benchmark. The weighting technique can become very complicated. There are numeral ways for defining the weighting factors. Thus, there is no clear guidance about weighting and exact knowledge of what the consequences may be. It is possible that weighted estimators are less accurate than the biased estimators. The effects of nonresponse bias are complex.
In theory nonresponse rates can be reduced to zero. The cost however for achieving maximum response rates would be extraordinary and the possible reduction of bias might be modest. Good response rate is vital, but the quality is mostly judged by the true and reliable response of the respondents. No. of words: 1186 (Excluding all References)
1. Armostrong, J. S. and T. s. Overton (1977). “Estimating Nonresponse Bias in Mail Surveys. ” Marketing Research 14: 396-402 2. Banda, J. P. (2003). Nonsampling errors in surveys. ESA/STAT/AC. 93/7, UNITED NATIONS SECRETARIAT, Statistics Division: 19.