Surveys can easily answer questions about knowledge, and best answer questions about attitudes and beliefs. They may also be useful to gather information on behaviors and practices, although this is dependent on the type of behavior or practice. For sensitive subjects, a survey may be more likely to reveal truth. For other subjects, data sources such as medical records may provide more reliable information.

The first and most important step in survey research is deciding what your research question is. This will allow you to determine whether you can use a previously validated survey or whether you must develop one on your own. Whenever possible, it is preferable to use a previously validated instrument without modification. This ensures that the questionnaire is reliable, valid, and responsive to change. Modifying a validated instrument reduces both its power and validity [1].

If you are unable to use a validated survey to answer your question, plan appropriately so you can spend time developing your survey instrument. This most important step is the aspect of survey research that is most often neglected. The quality of survey data will only be as good as the questions asked on the survey. This is true regardless of the ultimate mode of survey administration, and the techniques described below work for all modes. Often, both qualitative and quantitative techniques must be employed to create the optimal research tool. Clearly, the investigator and the research team have an idea of what questions they would like to include on a particular survey. Employing qualitative techniques using focus groups provides additional information about the topic area being investigated, and helps examine the assumptions brought by the research team during initial drafting of questions. Focus groups also help with how specific terms and vocabulary are understood by the population to be surveyed [4]. As an example, in developing a post-injury quality of life survey, you might ask a group of trauma patients “We are interested in the factors that have affected your quality of life since your injury. Can you tell us some of the things that have affected your quality of life both positively and negatively?” You would also want to ensure that each member of the group has an understanding of the term or concept quality of life.

Each question that you include must be critical. Avoid including questions “just because you’re interested” or “because it would be nice to know”. All questions should be clear and without bias. Avoid questions with two possible answers and leading questions that have a socially desirable answer. Normalizing statements, such as “It can be difficult to…” prior to asking about a sensitive item increase the likelihood of an honest answer. Questions that require a closed response are the easiest to analyze, so if at all possible try to frame the question to require a closed response [4, 8, 11].

Acquiescence is the tendency to endorse any assertion made in a question, regardless of content. Using declarative statements lessens this effect, although when level of agreement is what needs to be measured it is impossible to avoid this effect completely. The magnitude of the acquiescence effect is approximately 10 %; approximately 52 % of respondents agreed with an assertion, while 42 % of respondents disagreed with its opposite [11]. There are two approaches to minimizing the effect of acquiescence—the first is to ask all questions in one direction (usually positive), and the second is to use two questions, one asked positively and one negatively, to check on the magnitude of the effect in a specific survey.

Missing data tends to be more prevalent at the end of surveys, likely related to response burden. Therefore, ask the most important questions at the beginning, leaving the demographic questions to the end. Questions should be grouped thematically, as respondents fatigue if they must keep returning to a particular topic. In addition, questions should progress either from the general to the specific (funneling) or from the specific to the general (reverse funneling). If appropriate to the population, have your survey tested for language level and target the lowest likely educational level in your sample.

Pilot testing your survey will provide valuable information that can help you improve the quality of data in addition to your response rate. There are two options for initial pilot testing, with and without observation. Observing a small number of people answering your survey will give you information about which questions take respondents longest to answer, suggesting that they may be awkwardly or ambiguously worded. You are also able to debrief this small group about areas of concern and get their ideas for improvement. Field testing a pilot group without observation relies on their willingness to answer each question as well as provide written feedback for each question. Both methods should ask this pilot group to give feedback on the aesthetics of the survey, the ease with which it was completed, and their interest in completing the entire questionnaire. The completed questionnaire should then be tested with a final pilot group, which allows calculation of the psychometric properties of the questionnaire. The pilot groups should be similar to your intended population, but should not be included in your final sample.

For very small populations, it is desirable to obtain information from the entire population. However, for most surveys this is both impractical and expensive and therefore a sample of the population must be surveyed. The sampling frame is a list of the entire population. When qualitative surveys, such as focus groups, are performed, non-random sampling may be appropriate. However, for all other surveys, random sampling provides the most generalizable information. This can be accomplished using a random number generator or using a more systematic approach, selecting those to be surveyed from the sampling frame starting from a random point on the list at equal intervals (every fifth person on the list, for instance) [10].

Obtaining the sampling frame can be quite problematic for surveying health professionals, as many organizations have specific policies about handling of membership lists that specifically prohibit contact for survey research. Some are specific for email contact, allowing mail-based survey research to proceed. Many lists contain out of date information, and much effort must be spent cleaning a “dirty” or inaccurate list.

Although data from a simple random sample is optimal, it is often both extremely difficult and expensive to collect. From a practical standpoint, probability samples are often used to obtain data more cost-efficiently using complex sample designs. Probability samples divide the sampling frame into strata, and often to clusters within strata from which the sample is subsequently drawn. Complex sample designs are likely to result in unequal probability of selection for individual units of analysis, lack of independence of individual units within randomly sample clusters, and variable effect on estimates of precision. As a result, this approach which simplifies survey administration and data collection on the front end requires more complex statistical analysis [22].

The size of the sample necessary for a particular survey depends on the statistical analysis that will be performed, highlighting the importance of a well-thought analytic plan. The sample size can be calculated using a number of different computer packages or by consulting a statistician. Estimated non-response rates must be included in the sample size calculations.