Snowball sampling

1. Draft a participation program (likely to be subject to change, but indicative).
2. Approach stakeholders and ask for contacts.
3. Gain contacts and ask them to participate.
4. Community issues groups may emerge that can be included in the participation program.
5. Continue the snowballing with contacts to gain more stakeholders if necessary.
6. Ensure a diversity of contacts by widening the profile of persons involved in the snowballing exercise.

The participants are likely to know others who share the characteristics that make them eligible for inclusion in the study.[9]

Snowball sampling is quite suitable to use when members of a population are hidden and difficult to locate (e.g. samples of the homeless or users of illegal drugs) and these members are closely connected (e.g. organized crime, sharing similar interests, involvement in the same groups that are relevant to the project at hand).[9]

1. Social computing

Snowball sampling can be perceived as an evaluation sampling in the social computing field. For example, in the interview phase, snowball sampling can be used to reach hard-to-reach populations. Participants or informants with whom contact has already been made can use their social networks to refer the researcher to other people who could potentially participate in or contribute to the study.

2. Conflict environment

It has been observed that conducting research in conflict environment is challenging due to mistrust and suspicion. A conflict environment, where people or groups thinks their needs and goal are contradictory to the goals and or needs of other people or group. These conflicts among groups or people include the differences to claim the area of territory, resources, trade, civil and religious rights that cause considerable misunderstanding and heighten the disagreements that lead to an environment with lack of trust and suspicion. In conflict environment, the entire population is marginalized to some extent rather than a specific group of people and makes it very hard for investigators to reach the study subjects to conduct the research. For example, a threatening political environment under authoritarian regime creates obstacles for the investigators to conduct the research. Snowball sampling has demonstrated as a second best method in conducting research in conflict environments like, in the context of the Israel and Arab Conflict.[10] Snowball sampling allows the investigators to approach the marginalized population at cognitive and emotional level and enroll them in study. Snowball sampling address the conditions of lack of trust that arises due to uncertainty about the future through trace-linking methodology.[11]

3. Expert information collection

Snowball sampling can be used to identify experts in a certain field such as medicine, manufacturing processes, or customer relation methods, and gather professional and valuable knowledge.

For instance, 3M called in specialists from all fields that related to how a surgical drape could be applied to the body using snowball sampling. Every involved expert can suggest another expert who they may know could offer more information.

4. Public and population health research with marginalized and stigmatized populations

Snowball sampling can be used to recruit participants in research in marginalized, criminalized or other stigmatized behaviour, and its consequences. Examples include the use of illegal substances (e.g., unprescribed drugs), collection of illegal materials (e.g., ivory, unlicensed weapons), or stigmatized practices (e.g., support for anorexia, sexual fetish). Exclusion from majority society or fear of exposure or of shaming makes it difficult to contact participants through usual means. However, the nature of many of these behaviours means that people engaging in them have contact with each other. Snowball sampling is used in many studies of street-involved populations.[12]

1. Locate hidden populations: It is possible for the surveyors to include people in the survey that they would not have known but, through the use of social network.
2. Locating people of a specific population: There are no lists or other obvious sources for locating members of the population (e.g. the homeless, users of illegal drugs). The investigators use previous contact and communication with subjects then, the investigators are able to gain access and cooperation from new subjects. The key in gaining access and documenting the cooperation of subjects is trust. This is achieved that investigators act in good faith and establish good working relationship with the subjects.
3. Methodology: As subjects are used to locate the hidden population, the researcher invests less money and time in sampling. Snowball sampling method does not require complex planning and the staffing required is considerably smaller in comparison to other sampling methods.[13]

Snowball sampling can use in both alternative or complementary research methodology. As an alternative methodology, when other research methods can not be employed, due to challenging circumstancing and when random sampling is not possible. As complementary methodology with other research methods to boost the quality and efficiency of research conduct and to minimize the sampling bias like quota sampling.[11][14]

1. Community bias: The first participants will have a strong impact on the sample. Snowball sampling is inexact and can produce varied and inaccurate results. The method is heavily reliant on the skill of the individual conducting the actual sampling, and that individual's ability to vertically network and find an appropriate sample. To be successful requires previous contacts within the target areas, and the ability to keep the information flow going throughout the target group.[15]
2. Non-random: Snowball sampling contravenes many of the assumptions supporting conventional notions of random selection and representativeness[16] However, social systems are beyond researchers' ability to recruit randomly. Snowball sampling is inevitable in social systems.
3. Unknown sampling population size: There is no way to know the total size of the overall population.[9]
4. Anchoring: Another disadvantage of snowball sampling is the lack of definite knowledge as to whether or not the sample is an accurate reading of the target population. By targeting only a few select people, it is not always indicative of the actual trends within the result group. Identifying the appropriate person to conduct the sampling, as well as locating the correct targets is a time-consuming process such that the benefits only slightly outweigh the costs.
5. Lack of control over sampling method: As the subjects locate the hidden population, the research has very little control over the sampling method, which becomes mainly dependent on the original and subsequent subjects, who may add to the known sampling pool using a method outside of the researcher's control.

The best defense against weaknesses is to begin with a set of initial informants that are as diverse as possible.[9] Efforts to improve the main disadvantage of snowball sampling resulted in the respondent-driven sampling (RDS) method.[17] RDS augments the referral method by weighting the sample in order to compensate for the initial non-random selection, which may lead to the reduction of errors occurring in sampling by the referral method.[13]

A total of 214 cases was selected by 45 independent snowball sampling operations. Graphically represented, these sampling ranged in "length" (i.e., the number of cases in the sample) from two cases (1-stage snowball sample) to nine cases(8-stage snowball sample). Samples were collected with different "target" traits to saturate. In one subset, "foreign origin" was the trait to be saturated; in the other, it was "prostitution as occupation". In this example, three samples were selected from the dataset base for analysis to meet a criterion of holding the "length" of the three samples constant. A length of four cases was decided upon because these samples are complex enough to make statistical analysis practical, but short enough to allow clear and simple qualitative comparisons.

A field worker was instructed to start a snowball sample of a particular trait identified as characteristic of the heroin scene. At the zero stage, the subject selected was asked to nominate other heroin users sharing that trait (the maximal number of nominees was 25). From the set of those nominated at each stage, a simple random selection was made of a single individual. The field worker then attempted to make contact with that nominee. The amount of time in days (speed) required to make contact was recorded, as well as other specified traits of the individuals (e.g., sex, age, drug history, and patterns).

Figure 1. The result of one snowball sample

Figure 1 graphically shows the traits in this sample in terms or the days it took to find a specific nominee, the operationalization of speed. Sample one was started with a 27-year-old British heroin user whose referrals included a 19-year-old Italian. After this Italian nominee was selected, it took the field worker 4 days to find him. He then nominated, among others, another Italian, 22 years of age, who was found on the same day. The randomly selected nominee of this third user was a 27-year-old Belgian who could not nominate another.

After subjects were identified, both quantitative and qualitative analyses of the three samples were conducted, and graphic representations of the data were constructed and marked for relevant traits. Descriptive statistical comparisons were made for the entire dataset as well as for the foreign and prostitute subgroups. Inferential statistics were also used to determine whether the distributions for age and the time it took for a field worker to locate a nominee (speed) were significant and whether the respective snowballs were drawn from populations with the same distributions. The second question was seen as especially appropriate for and "ascending" sampling strategy because it cannot be assumed that each snowball is drawn from the same population when only an "imperfect sampling frame" composed of a "special list" compiled by nominees, is available.[19] Because it has been seen as especially appropriate for small samples (as little as three cases) for which population parameters are unknown and cannot be confidently assumed, the nonparametric Kolmogorov-Smirnov (KS) test was used. Two-tailed KS tests were performed on the pooled data of the three samples (one-sample test) and on the between-snowballs (subgroups) data (two-samples test).

• In hard-to-reach and hard-to-involve populations online sampling can better detect individuals of researcher's interest and allows to expand geographical scope of the studies[4]
• Brings the possibility to increase representativeness of the results[4]
• Virtual sampling can increase the number of responses in comparison with traditional snowball sampling. According to Baltar (2012) who used Facebook to search for participants for his study and conduct the research, it was possible to reduce the time necessary for building trust between the participant and researcher. Participants were more likely to share their personal information because the researcher was also sharing personal information on his/her Facebook profile. Increased level of confidence contributed to higher response rate[4]
• Less costly relative to traditional snowball sampling technique[4]

• Even though the virtual sampling method can increase representativeness of the results, sample selection is biased towards the characteristics of online population such as gender, age, education level, socioeconomic level, etc.[4]
• Target population might not always have access to the Internet[4]

Virtual snowball sampling technique was used in order to find participants for the study of a minority group - Argentinian entrepreneurs living in Spain. About 60 percent of this population has double nationality – both Spanish and Argentinian. Spanish national statistics classifies them as European citizens only and there is no information about the place of birth tied to the profiles of entrepreneurs in Spain either. Therefore, referring to national statistics only, made it impossible to build a sample frame for this research. The use of virtual networks in this example of hard to reach population, increased the number of participating subjects and as a consequence, improved the representativeness of results of the study.[4]

Ethical concerns may prevent the research staff from directly contacting many potential respondents. Therefore, program directors or personnel who knew of possible respondents can make initial contacts and then ask those who were willing to cooperate to personally contact the project. In each instance, the newly recruited research participant must be trained to understand and accept the eligibility criteria of the research. For example, in a study on treatment for substance-use disorder which used snowball sampling, it was difficult for many to understand the eligibility criteria because some criteria violated common-sense understandings concerning treatment and non-treatment. For example, many people define themselves as untreated in spite of possible long stays in civil commitment programs because their commitments to these institutions were involuntary and/or because they had become re-addicted upon release and then recovered at a later time.[20] Therefore, the quality of informed consent was in doubt.

In a qualitative research, apprehension around feelings of compulsion are reviewed for potential ethical dilemmas and recommendations for research process are made.[21]

A new approach to the study of hidden populations. It is effectively used to avoid bias in snowball sampling. Respondent-driven sampling involves both a field sampling technique and custom estimation procedures that correct for the presence of homophily on attributes in the population. The respondent-driven sampling method employs a dual system of structured incentives to overcome some of the deficiencies of such samples. Like other chain-referral methods, RDS assumes that those best able to access members of hidden populations are their own peers.[23]

Peer Esteem Snowballing is a variation of snowball sampling, useful for investigating small populations of expert opinion. Its proponents[24] argue that it has a number of advantages relative to other snowballing techniques:

1. reduces the selection bias inherent in initial seed samples for a snowball by advocating for a nominations phase that objectively identifies contact seeds for the first wave;
2. by analysing network data it provides an estimate of the population size, unbiased by any researcher defined population boundary;
3. by reporting the estimate of the sample size vis a vis the population, it provides a measure of relative significance (optimal sampling data can be reported in this context);
4. through a network analysis of referrals it allows for identifying clusters of experts that may be instrumental in explain variations in their response profile;
5. allows for a referrals nominations strategy that, in certain cases, could improve response rates, while the nominations strategy acts as an ultimate validation of expertise for informants and therefore improves content validity.