Intended for healthcare professionals

Endgames Statistical Question

Snowball sampling

BMJ 2013; 347 doi: (Published 20 December 2013) Cite this as: BMJ 2013;347:f7511
  1. Philip Sedgwick, reader in medical statistics and medical education1
  1. 1Centre for Medical and Healthcare Education, St George’s, University of London, London, UK
  1. p.sedgwick{at}

Researchers examined the association between plasma HIV-1 RNA concentrations and the incidence of HIV in injecting drug users. A prospective cohort study design was used. Participants were injecting drug users, with or without HIV, followed up every six months between 1 May 1996 and 30 June 2007. Cohort members were recruited from an inner city community in Vancouver, Canada, using the method of snowball sampling. In total, 622 injecting drug users with HIV and 1429 injecting drug users without HIV were recruited.1

The researchers reported that in the community of injecting drug users, a longitudinal measure of plasma HIV-1 RNA concentration was positively correlated with the HIV incidence rate, independently of unsafe sexual behaviours and sharing used syringes. It was concluded the results could help inform HIV prevention and treatment interventions.

Which of the following statements, if any, are true?

  • a) Snowball sampling constitutes probability sampling

  • b) Snowball sampling requires the construction of a sampling frame

  • c) Snowball sampling is prone to selection bias


Statement c is true, whereas a and b are false.

Two types of sampling methods can be used to recruit participants to a study—random sampling (sometimes called probability sampling) and non-random sampling (sometimes called non-probability sampling) methods. Snowball sampling is a type of non-random (non-probability) sampling method (a is false).

Random sampling methods involve some form of random selection of the members of the population. Each member of the population has a known and typically equal probability of being selected for the sample. The most straightforward type of random sampling method is simple random sampling (sometimes referred to as random sampling). Random sampling requires knowledge of exactly who is in the population, with construction of a sampling frame—that is, a list of everyone in the population. A sample of a fixed size is selected at random from the list, with all members of the population having the same probability of being selected independently of all others. The probability that a population member will be chosen is known in advance.

In the above study, snowball sampling initially involved identifying an easily accessible small number of injecting drug users for recruitment to the cohort. These cohort members then recommended other injecting drug users as potential cohort members, who in turn suggested others. The method is so called because the number of sample members increases with time, analogous to a snowball accumulating snow as it rolls down a hill. Snowball sampling is used to recruit samples when members of the desired population are hard to reach or access because they feel disempowered, socially excluded, or vulnerable.

Snowball sampling was probably the only sampling method that could have been used to obtain a reasonably representative sample of injecting drug users. Unlike random sampling methods, it did not require the construction of a sampling frame (b is false). However, it would probably not have been possible to construct a list of injecting drug users in the community because they would have been hard to access. Even if they were identified, they might have shown a certain amount of distrust and been reluctant to be recruited if approached at random. The use of existing cohort members to recommend other potential members probably encouraged recruitment through an element of trust in the community. To increase recruitment and ensure that loss to follow-up was minimised, cohort members were offered $C20 (£11.5; €13.7; $18.8) at each study visit.

The cohort that resulted from snowball sampling was prone to selection bias (c is true). Selection bias, described in a previous question,2 would have occurred if the injecting drug users recruited to the cohort were systematically different from those who were not recruited—resulting in the cohort not being representative of the population. Recruitment of injecting drug users depended on the recommendations of existing cohort members. Therefore, the cohort was unlikely to be representative of the population owing to biases in the injecting drug users who were recommended. For example, injecting drug users who did not have many friends were less likely to be recruited. Typically, any sample resulting from non-probability sampling methods is considered not to be representative of the population. Little is known about the properties of samples resulting from snowball sampling.


Cite this as: BMJ 2013;347:f7511


  • Competing interests: None declared.


View Abstract