Sampling Methods in Research: Types, Methods and Examples

Muhammad Hassan

Research Sampling

Sampling Methods in Research: Types, Methods and Examples

86,560 Views

Published on: March 22, 2024

by Muhammad Hassan

Updated on: June 29, 2026

Table of Contents

Sampling methods are the procedures researchers use to select people, cases, organisations, documents, events, or other units from a larger population. The two broad families are probability sampling, which uses a known random-selection process, and non-probability sampling, which selects cases through availability, judgement, referrals, quotas, or other non-random criteria.

This guide explains the principal sampling methods used in quantitative, qualitative, and mixed-method research. It shows how each technique works, when it is appropriate, what it allows a researcher to conclude, and what limitations should be reported.

Key Takeaways

Sampling methods determine how units enter a study; sample size determines how many units are selected.

Probability sampling uses known, non-zero inclusion probabilities and supports design-based statistical inference when implemented and analysed correctly.

Non-probability sampling is appropriate for many exploratory, qualitative, specialist, and hard-to-reach population studies, but population generalisation requires additional assumptions.

A large sample cannot repair serious undercoverage, self-selection, or recruitment bias.

Clustered, stratified, multistage, and unequally weighted samples must be analysed using their design information.

Researchers should report the population, frame, selection procedure, recruitment, response, weighting, quality controls, and limitations.

What Are Sampling Methods?

A sampling method is the procedure used to select a subset of units from a population for observation or analysis. Units may be individuals, households, schools, hospitals, businesses, documents, social-media posts, biological specimens, events, or geographical areas.

Researchers use samples because studying every member of a population is often too expensive, slow, intrusive, or logistically impossible. A carefully designed sample can provide useful information about a much larger population, but the conclusions that can be drawn depend on how that sample was selected.

Sampling should not be treated as a final recruitment step. It is part of the research design because it affects:

Whose experiences or measurements are included
Which parts of the population may be missing
Whether population estimates are justified
How uncertainty should be calculated
How widely the findings can be applied

Essential Sampling Terms

Term	Meaning	Example
Target population	The full group to which the research question refers	All registered nurses working in UK hospitals
Accessible population	The portion of the target population the researcher can realistically reach	Nurses employed by hospitals participating in the study
Sampling element	The basic entity about which information is required	One nurse
Sampling unit	The unit selected at a particular stage	A hospital at stage one and a nurse at stage two
Sampling frame	The operational list or system used to identify eligible units	A current staff register
Sample	The units actually selected or recruited	800 nurses
Respondent or participant	A selected person who provides data	A nurse who completes the questionnaire
Parameter	A true but usually unknown population value	Mean weekly working hours of all eligible nurses
Statistic	A value calculated from the sample	Mean weekly working hours among responding nurses
Census	Data collection from every unit in the defined population	Surveying every employee in a small organisation

The target population and sampling frame are not necessarily identical. A register may omit recent entrants, contain duplicate records, include ineligible people, or exclude people without a fixed address or internet access. These differences create coverage error.

Why Are Sampling Methods Important?

A sampling method affects both the credibility of the data and the scope of the conclusions.

An appropriate method can:

Reduce systematic selection differences
Ensure important groups are included
Make fieldwork feasible
Improve the precision of estimates
Support transparent population inference
Produce information-rich qualitative cases
Reduce unnecessary participant burden

An inappropriate method can produce a large but misleading dataset. For example, an online survey advertised through one university’s social-media accounts may collect thousands of responses while still excluding students who do not follow those accounts, use those platforms, or feel motivated to participate.

A sample is therefore not representative merely because it is large or demographically diverse. Representativeness concerns the relationship between the selection process, the achieved sample, the variables being studied, and the target population.

What Are the Two Main Types of Sampling?

The two main types are probability sampling and non-probability sampling.

Probability sampling

Probability sampling uses a defined random-selection mechanism under which each population unit has a known, non-zero chance of inclusion.

The probabilities do not always have to be equal. Simple random sampling gives units equal probabilities, while stratified, oversampled, or probability-proportional-to-size designs may deliberately give some units higher probabilities.

Non-probability sampling

Non-probability sampling does not provide a known random-selection probability for every population unit. Cases may be selected because they are accessible, eligible, knowledgeable, unusual, willing to volunteer, or connected to existing participants.

Non-probability sampling is not automatically poor research. It may be the most appropriate option when the purpose is depth, theory development, case comparison, initial exploration, specialist knowledge, or access to a population for which no usable frame exists.

Probability and Non-Probability Sampling Compared

Feature	Probability sampling	Non-probability sampling
Selection basis	Defined random mechanism	Availability, judgement, quotas, referrals, volunteering, or other criteria
Inclusion probability	Known and greater than zero	Unknown for at least some units
Sampling frame	Usually required, although area and multistage frames may be used	May not be available or required
Statistical generalisation	Supports design-based inference when correctly implemented and analysed	Requires modelling, adjustment, or cautious analytic generalisation
Conventional sampling error	Can be estimated from the design	Cannot automatically be calculated from recruitment counts alone
Typical use	Population surveys, prevalence studies, official statistics, audits	Qualitative studies, pilot research, hidden populations, case studies, rapid research
Principal strength	Transparent basis for population estimation	Feasibility and targeted case selection
Principal limitation	Can be expensive and operationally demanding	Greater vulnerability to unknown selection differences

Probability Sampling Methods

1. Simple Random Sampling

Simple random sampling selects a fixed number of units so that every possible sample of that size has the same chance of selection. Each individual unit consequently has an equal probability of inclusion.

How it works

Define the eligible population.
Create or obtain a complete sampling frame.
assign each unit a unique identifier.
determine the required sample size.
use a reproducible random process to select identifiers.
contact the selected units without replacing nonrespondents with convenient alternatives.

Example

A university has 8,000 enrolled postgraduate students and wants to survey 500 of them. Each student receives a unique number, and software selects 500 numbers using a stored random seed.

Advantages

Conceptually straightforward
Equal inclusion probabilities
Minimal researcher discretion during selection
Supports standard statistical estimation
Easy to reproduce when the frame and random seed are documented

Limitations

Requires a sufficiently complete frame
Selected units may be geographically dispersed
Small subgroups may be selected in insufficient numbers
Nonresponse can still distort the achieved sample
A particular realised sample can differ from the population by chance

Best used when

The population is clearly defined, a reliable list exists, and no subgroup requires guaranteed representation or oversampling.

2. Systematic Sampling

Systematic sampling selects every (k)th unit from an ordered frame after choosing a random starting position. It is a probability method when the starting point is random and the frame is used according to a predetermined rule.

The approximate sampling interval is:

[
k=\frac{N}{n}
]

where:

(N) = number of units on the frame
(n) = desired sample size
(k) = selection interval

Example

A clinic has 2,400 eligible patient records and needs 240 records.

[
k=\frac{2400}{240}=10
]

The researcher randomly selects a starting number from 1 to 10. If the start is 7, records 7, 17, 27, 37, and so forth are selected.

Advantages

Faster than generating a separate random number for every unit
Spreads selections across the frame
Easy to administer in production lines, registers, or ordered files
Can approximate simple random sampling when ordering is unrelated to the outcome

Limitations

Periodicity in the frame can bias selection
The method can fail if records are ordered in a repeating pattern matching (k)
Frames with frequent additions or deletions require careful handling
Circular or fractional interval procedures may be needed when (N/n) is not an integer

Best used when

A reliable ordered list exists and no hidden periodic pattern is likely to coincide with the interval.

3. Stratified Random Sampling

Stratified sampling divides the population into mutually exclusive subgroups called strata and selects a probability sample from every stratum. It is used to guarantee subgroup representation or improve precision.

Strata may be based on:

Region
Institution type
Age group
Occupational category
Programme level
Disease category
Business size

Stratification variables should be known for frame units before selection.

Proportionate allocation

Under proportionate allocation, the sample from stratum (h) is:

[
n_h=n\left(\frac{N_h}{N}\right)
]

where:

(n_h) = sample allocated to stratum (h)
(N_h) = population size of stratum (h)
(N) = total population size
(n) = total sample size

Example

A university population contains:

6,000 undergraduates
3,000 master’s students
1,000 doctoral students

For a proportionate sample of 1,000:

Undergraduates: (1,000 \times 6,000/10,000 = 600)
Master’s students: (1,000 \times 3,000/10,000 = 300)
Doctoral students: (1,000 \times 1,000/10,000 = 100)

A random sample is then selected separately within each group.

Disproportionate stratification

Researchers may deliberately oversample a small or analytically important stratum. For example, doctoral students might be oversampled so that subgroup estimates are sufficiently precise. Population estimates must then use weights reflecting the unequal inclusion probabilities.

Advantages

Guarantees that every defined stratum contributes cases
Can improve precision when units are similar within strata
Supports subgroup estimates
Permits deliberate oversampling
Can reduce the risk that a small subgroup is missed by chance

Limitations

Requires accurate frame information
Strata must be defined before selection
Poorly selected strata may provide little precision benefit
Oversampling creates unequal weights
Analysis must account for strata and weights

Best used when

The population contains important, identifiable subgroups and the researcher needs reliable overall or subgroup estimates.

4. Cluster Sampling

Cluster sampling divides a population into naturally occurring groups, randomly selects some groups, and studies all or a sample of the units within the selected groups.

Common clusters include:

Schools
Hospitals
Villages
Neighbourhoods
Workplaces
Households
Classrooms

One-stage cluster sampling

The researcher selects clusters and includes every eligible unit within them.

Example: Randomly select 20 schools and survey every teacher in those schools.

Two-stage cluster sampling

The researcher selects clusters and then selects a sample of units within each selected cluster.

Example: Select 20 schools, then randomly select 25 teachers from each school.

Advantages

Reduces travel and fieldwork costs
Does not require one national list of all individual units
Practical for geographically dispersed populations
Compatible with multistage national surveys

Limitations

People within the same cluster often resemble one another
Similarity within clusters reduces effective statistical information
Standard errors may be larger than under a simple random sample of the same nominal size
Too few clusters can produce unstable estimates
Analysis must identify clusters or primary sampling units

The efficiency loss is often summarised using the design effect. A design effect greater than 1 indicates that the complex design has a larger variance than a simple random sample of the same nominal size.

Best used when

The population is geographically dispersed, individual-level frames are unavailable, or travel and contact costs make direct random sampling impractical.

5. Multistage Sampling

Multistage sampling selects units through two or more levels, progressively moving from large groups to smaller groups or individuals.

A national education study might:

Stratify the country by region.
select districts within regions.
select schools within districts.
select classrooms within schools.
select students within classrooms.

Different probability techniques can be used at different stages. For example, districts may be selected with probability proportional to size, schools systematically, and students through simple random sampling.

Advantages

Suitable for national or widely dispersed populations
Allows several available frames to be combined
Reduces fieldwork costs
Can oversample particular areas or groups

Limitations

Inclusion probabilities can become difficult to calculate
Weight construction is more complex
Clustering must be included in variance estimation
Errors in one stage can affect all later stages
Detailed documentation is essential

6. Probability-Proportional-to-Size Sampling

Probability-proportional-to-size sampling gives larger clusters a greater chance of selection, usually because their size reflects the number of final elements they contain.

Suppose schools are the first-stage units and students are the final elements. A school with 2,000 students may receive a higher selection probability than a school with 200 students.

When PPS selection is followed by an appropriate fixed number of students per selected school, it can help produce approximately equal final selection probabilities. PPS is common in large household, health, education, and establishment surveys.

It requires a reliable measure of size and careful calculation of inclusion probabilities.

Non-Probability Sampling Methods

1. Convenience Sampling

Convenience sampling recruits the cases that are easiest for the researcher to access.

Examples include:

Surveying students in the researcher’s own class
Posting an open questionnaire to social media
Interviewing visitors at one accessible clinic
Analysing documents found through the first page of a search

Advantages

Fast and inexpensive
Useful for piloting instruments
Helpful for preliminary exploration
May be acceptable for testing processes or technical systems

Limitations

Recruitment pathways may exclude important groups
Inclusion probabilities are unknown
Volunteers may differ systematically from nonparticipants
A large convenience sample can still be severely biased
Population prevalence or percentage estimates may be unjustified

Researchers should identify the accessible population honestly. A sample of students from one class should not be described as a random sample of all university students.

2. Consecutive Sampling

Consecutive sampling includes every accessible case meeting the eligibility criteria during a defined period until the target size is reached.

It is common in clinical and service settings.

Example

A clinic recruits every eligible patient attending between January 1 and June 30, rather than selecting only patients who are especially easy or interesting to approach.

Consecutive sampling is usually stronger than informal convenience selection because the eligibility and time-window rules are explicit. It remains non-probability sampling if attendance at that clinic and period does not arise from a known population-selection process.

3. Voluntary-Response Sampling

Voluntary-response sampling allows individuals to opt into a study after seeing an open invitation.

Examples include:

Website polls
Radio call-in surveys
Public social-media questionnaires
Open customer-feedback forms

People with strong opinions, greater available time, higher digital engagement, or a personal connection to the topic may be more likely to respond. The achieved sample can therefore differ substantially from the target population.

4. Purposive Sampling

Purposive sampling deliberately selects cases because they can provide information relevant to the research question.

It is widely used in qualitative and mixed-method research. The principle is not to recruit whoever is easiest, but to define and justify the characteristics that make a case informative.

Purposeful selection should follow the conceptual framework and research aims and should be ethically and practically feasible (Palinkas et al., 2015).

Common purposive strategies

Strategy	Purpose	Example
Criterion sampling	Include all cases meeting a meaningful condition	Teachers who have used a new curriculum for at least one year
Maximum-variation sampling	Capture a wide range of relevant experiences	Rural and urban schools of different sizes and funding levels
Homogeneous sampling	Examine one narrowly defined group in depth	First-year doctoral students in laboratory sciences
Typical-case sampling	Study cases considered broadly ordinary	A medium-sized public school with average performance
Extreme or deviant-case sampling	Learn from unusual outcomes	Clinics with exceptionally high or low retention
Critical-case sampling	Select a strategically important case	A high-capacity organisation where failure would challenge the intervention
Expert sampling	Recruit people with specialised knowledge	Survey statisticians who design national probability samples
Intensity sampling	Study strong but not extreme examples	Programmes showing substantial, but not exceptional, improvement

Purposive sampling does not automatically mean that all eligible people will be included. Researchers should report who applied the criteria, how potential cases were identified, and how judgement was controlled.

5. Quota Sampling

Quota sampling sets numerical targets for selected subgroups but recruits participants non-randomly within those targets.

A researcher might require:

50 respondents aged 18–29
50 aged 30–49
50 aged 50 or older

Recruiters continue until each quota is filled.

Advantages

Ensures the achieved sample contains selected characteristics
Faster and less expensive than stratified probability sampling
Useful when no complete frame exists
Common in rapid opinion and market research

Limitations

People within each quota are not randomly selected
Matching the population on age or gender does not guarantee similarity on other variables
Recruiter discretion and self-selection remain
Conventional probability-sample margins of error are not automatically valid

6. Snowball Sampling

Snowball sampling recruits initial participants and asks them to refer other eligible people.

It is useful when:

No practical sampling frame exists
Membership is rare or not publicly visible
Trust is necessary before contact
Participants are connected through social networks

Example

A researcher studying undocumented freelance workers begins with several trusted contacts who refer other eligible workers.

Limitations

Initial participants influence subsequent recruitment
Highly connected people may be overrepresented
Separate networks may never enter the sample
Participants may refer people similar to themselves
Confidentiality must be protected during referrals

Researchers should not ask participants to disclose sensitive information about another person without permission. A safer procedure is often to give participants neutral study information that potential recruits can use to contact the research team themselves.

7. Theoretical Sampling

Theoretical sampling is an iterative strategy used principally in grounded theory. Researchers select new cases according to concepts emerging during data collection and analysis.

The sample is not fully specified at the beginning. Early findings indicate which settings, people, events, or documents should be examined next to develop, refine, or challenge emerging theoretical categories.

Theoretical sampling differs from simply continuing recruitment until no new interview topic appears. Its purpose is theory development, not numerical representativeness.

8. Time-Location Sampling

Time-location sampling recruits members of a hard-to-frame population at identifiable venues and times where they gather.

Researchers may construct a frame of venue–day–time units, randomly select those units, and approach eligible people present during each selected period.

Depending on implementation, time-location sampling can incorporate probability elements. Its coverage is nevertheless limited to people who attend the mapped venues and times.

9. Respondent-Driven Sampling

Respondent-driven sampling is a structured form of chain-referral sampling designed for hidden or networked populations.

It normally begins with selected “seeds,” limits how many peers each participant may recruit, records recruitment links, and collects information about network size. Estimation procedures attempt to adjust for network-related selection differences.

RDS was introduced as a way to reduce some weaknesses of ordinary chain-referral samples (Heckathorn, 1997). However, its estimators depend on assumptions about network connections, recruitment, population boundaries, and sampling behaviour. It should not be described as automatically unbiased or equivalent to a conventional simple random sample.

Stratified Sampling Versus Cluster Sampling

Question	Stratified sampling	Cluster sampling
Why divide the population?	To represent or compare important subgroups	To make selection and fieldwork practical
Which groups are sampled?	Every stratum	Only selected clusters
Desired group structure	Units within a stratum are relatively similar on stratification variables	Each cluster ideally contains useful population diversity
Example	Sample students from every degree level	Select several universities and sample students within them
Typical precision effect	Can improve precision	Often reduces precision because of within-cluster similarity
Principal operational benefit	Guaranteed subgroup coverage	Lower travel and listing costs

A study can use both. For example, it may stratify schools by region and then select clusters of schools within every region.

Stratified Sampling Versus Quota Sampling

The methods look similar because both divide a population into categories.

The decisive difference is selection within each category:

Stratified sampling uses a probability-selection procedure within every stratum.
Quota sampling fills subgroup targets using non-random recruitment.

A quota sample with the same age distribution as the population is not automatically a stratified probability sample.

Random Sampling Versus Random Assignment

These concepts address different questions.

Random sampling selects units from a population and concerns population representativeness and external validity.
Random assignment allocates enrolled participants to conditions and concerns causal comparability and internal validity.

A convenience sample can be randomly assigned in an experiment. That random assignment may support a causal comparison within the study while still leaving uncertainty about whether the effect applies to the broader population.

How to Choose a Sampling Method

Step 1: Define the target population

State exactly:

Who or what is included
Geographic boundaries
Relevant dates or periods
Age, role, organisational, or diagnostic criteria
Exclusion criteria
Unit of analysis

“University students” is usually too vague. A more operational definition is “students enrolled in taught postgraduate programmes at publicly funded universities in England during the 2026 spring term.”

Step 2: Decide what conclusion is required

Ask whether the study aims to:

Estimate a population percentage, total, prevalence, or mean
Compare predefined population subgroups
Estimate a causal effect
Explore experiences or mechanisms
Develop theory
Understand exceptional cases
Pilot an instrument or process

Population estimation generally favours probability sampling. In-depth explanation may favour purposive or theoretical sampling.

Step 3: Determine whether a suitable frame exists

Evaluate:

Coverage
Duplicates
Outdated records
Contact information
Eligibility indicators
Legal and ethical access
Availability of stratification variables

When no individual frame exists, area, address, venue, cluster, or multistage approaches may be considered.

Step 4: Examine population structure

Consider whether the population is:

Geographically dispersed
Organised into natural clusters
Highly heterogeneous
Divided into small but important subgroups
Rare, hidden, mobile, or stigmatised
Accessible only through institutions or networks

Step 5: Assess practical and ethical constraints

Review:

Budget
Fieldwork time
Staffing
Travel
Participant risk
Gatekeeper access
Privacy
Language needs
Compensation
Burden on small communities

The most statistically elegant design may not be ethical or feasible.

Step 6: Select the design

A practical rule is:

Research situation	Often suitable
Complete, reliable list and general population estimate	Simple random or systematic
Important known subgroups	Stratified random
Widely dispersed population	Cluster or multistage
Unequal cluster sizes	PPS within a multistage design
Rapid preliminary study	Convenience or voluntary response, with explicit limits
Information-rich specialist cases	Purposive
Fixed subgroup targets without a probability frame	Quota
Hidden or networked population	Snowball, time-location, or carefully designed RDS
Grounded-theory development	Theoretical sampling

Step 7: Plan sample size and allocation

Sample-size planning should match:

The principal outcome
The planned analysis
Required precision or power
Population size where relevant
Number and size of subgroups
Clustering and design effects
Anticipated nonresponse
Expected attrition
Qualitative information needs

Step 8: Predefine implementation rules

Specify:

Randomisation procedure and seed
Replacement policy
Number of contact attempts
Recruitment channels
Quota rules
Referral limits
Eligibility screening
Duplicate prevention
Procedures for unreachable units
Oversampling and weighting plans

Step 9: Monitor the achieved sample

During recruitment, examine:

Response by stratum or subgroup
Frame coverage
Refusal and ineligibility
Duplicate records
Recruitment-source differences
Cluster completion
Quota filling
Interview quality
Fraud or automated responding

Do not quietly replace difficult-to-contact selected participants with easier cases. Such substitution changes the selection mechanism.

Step 10: Match the analysis to the design

Complex-survey analysis may need:

Sampling weights
Stratum identifiers
Cluster or primary sampling unit identifiers
Finite population information
Replicate weights
Taylor-series or replication-based variance estimation

Ignoring clustering can underestimate uncertainty. Ignoring unequal probabilities can produce estimates that do not correctly represent the target population.

Sampling Method Versus Sample Size

A sampling method determines how cases are selected. Sample size determines how many cases are selected or observed.

A large sample selected badly can produce a precise estimate of the wrong value. A smaller, well-designed probability sample may provide more defensible population estimates.

Sample-size calculations do not prove that a study is representative. They assume that the underlying selection, measurement, response, and analysis processes are appropriate.

Basic Quantitative Sample-Size Formula

For estimating a population proportion under simple random sampling, an initial large-population calculation is:

[
n_0=\frac{z^2p(1-p)}{e^2}
]

where:

(n_0) = initial sample size
(z) = critical value for the chosen confidence level
(p) = anticipated population proportion
(e) = desired absolute margin of error

When no defensible estimate of (p) is available, (p=.50) produces the largest variance and therefore a conservative initial size for this formula.

Worked example

For a 95% confidence level, (z=1.96), (p=.50), and (e=.05):

[
n_0=\frac{1.96^2(.50)(.50)}{.05^2}=384.16
]

Round upward to 385 completed observations.

Finite population correction

For a finite population of size (N):

[
n=\frac{n_0}{1+\frac{n_0-1}{N}}
]

For (N=2,000):

[
n=\frac{384.16}{1+\frac{383.16}{2000}}\approx322.4
]

Round upward to 323 completed observations.

Allowing for design effects and nonresponse

A planning approximation is:

[
n_{\text{invited}}=
\frac{n_{\text{required}}\times DEFF}
{\text{expected response proportion}}
]

If the required size is 323, the anticipated design effect is 1.5, and the expected response rate is 60%:

[
n_{\text{invited}}=
\frac{323\times1.5}{.60}
\approx807.5
]

The researcher would plan to invite approximately 808 units.

This is an illustration, not a universal calculator. Studies estimating means, rare outcomes, regression coefficients, diagnostic accuracy, survival differences, treatment effects, or multilevel effects need calculations tailored to those analyses.

Sampling Weights

When units have unequal inclusion probabilities, a basic design weight for unit (i) is:

[
w_i=\frac{1}{\pi_i}
]

where (\pi_i) is the unit’s inclusion probability.

If a person has an inclusion probability of 1 in 200, the base weight is 200. Conceptually, that responding person represents 200 population units under the design.

Final weights may also include adjustments for:

Subsampling
Eligibility
Nonresponse
Coverage
Post-stratification
Raking
Calibration to known population totals
Integration of multiple frames or panels

Weights can reduce some known imbalances, but they do not guarantee removal of selection bias. Adjustment is limited by the quality of the auxiliary variables and the assumptions connecting those variables to participation and outcomes.

Sampling Error and Nonsampling Error

What is sampling error?

Sampling error is the difference that arises because a sample rather than the entire population was observed. Under a probability design, its variability can be estimated from the design.

Different valid probability samples drawn from the same population will usually produce somewhat different estimates.

What is nonsampling error?

Nonsampling error includes errors not caused merely by observing a sample. It can occur in both samples and censuses.

Important forms include:

Error	Description
Undercoverage	Eligible population units are absent from the frame or recruitment channels
Overcoverage	Ineligible or duplicate units appear on the frame
Nonresponse error	Respondents differ meaningfully from nonrespondents
Self-selection bias	Participation is associated with motivation or the study topic
Measurement error	Questions, instruments, interviewers, or respondents produce inaccurate measurements
Processing error	Mistakes occur during coding, entry, cleaning, linkage, or analysis
Attrition bias	Dropout differs according to outcomes or characteristics
Periodicity bias	A systematic interval aligns with a repeating frame pattern
Network bias	Referral recruitment overrepresents particular social connections
Mode-related coverage	A web, telephone, or in-person mode excludes some population groups

Increasing sample size primarily improves random precision. It does not automatically correct these systematic problems.

What Makes a Sample Representative?

A representative sample provides a credible basis for learning about the target population with respect to the variables and claims of interest.

Representativeness cannot be established solely by showing that the sample matches the population on several visible demographics. Two samples may have identical age and sex distributions but differ in health, political interest, digital access, institutional trust, or willingness to participate.

Researchers should examine:

Coverage of the frame
Selection probabilities
Recruitment and response mechanisms
Comparisons with credible population benchmarks
Weighting variables
Sensitivity to alternative adjustments
Subgroup response
Missing data
Whether conclusions extend beyond the observed sample

Sampling in Qualitative Research

Qualitative research normally seeks depth, meaning, process, diversity, or theory rather than a statistically representative estimate of a population percentage.

Frequently used qualitative strategies include:

Purposive sampling
Maximum-variation sampling
Criterion sampling
Homogeneous sampling
Critical-case sampling
Snowball sampling
Theoretical sampling

How many qualitative participants are needed?

There is no universal number of interviews that is appropriate for every qualitative study.

Malterud, Siersma, and Guassora (2016) propose information power as a planning principle. Fewer participants may be sufficient when:

The study aim is narrow
Participants are highly specific to the question
Strong theory guides the inquiry
Interview dialogue is rich
Analysis is focused and in depth

More participants may be needed when:

The aim is broad
The sample is highly diverse
Cases provide limited information
Several subgroup comparisons are planned
Data are thin or inconsistent

“Saturation” should not be used as an unexplained statement. Researchers should identify what they mean—for example, code saturation, meaning saturation, theoretical sufficiency, or no meaningful change to an analytical category—and describe how it was assessed.

Sampling in Mixed-Methods Research

Mixed-method studies may use different samples for different components.

Common designs include:

Identical sampling

The same participants contribute quantitative and qualitative data.

Example: Survey respondents also complete open-ended interviews.

Nested sampling

A qualitative subsample is drawn from a larger quantitative sample.

Example: A survey of 2,000 teachers is followed by purposive interviews with 40 teachers representing different experiences.

Parallel sampling

Separate samples are used for quantitative and qualitative components.

Example: A probability survey estimates prevalence while interviews with service managers explore implementation.

Multilevel sampling

Different components study different levels.

Example: Students complete questionnaires, teachers participate in interviews, and schools provide administrative records.

The sampling relationship should be planned explicitly so that each component answers its intended question and the datasets can be integrated meaningfully.

Sampling in Online Research

Online recruitment can involve either probability or non-probability methods.

Probability-based online panels

Panel members are initially recruited through a probability method, such as address-based sampling, and are subsequently invited to online studies. People without internet access may need equipment or alternative participation modes if the target population includes them.

Opt-in online panels

Members volunteer through advertisements, websites, affiliates, applications, or other recruitment channels. Quotas and statistical adjustments may improve balance on measured characteristics, but recruitment remains non-probability unless inclusion arises from a defined probability design.

Questions to ask an online-panel provider

Researchers should ask:

How were panel members originally recruited?
Is recruitment probability-based, opt-in, or blended?
Which population is covered?
How are people without reliable internet access handled?
How often is the panel refreshed?
How are duplicate accounts detected?
How are bots, identity fraud, and professional respondents screened?
How frequently may one person participate?
How are quotas set and filled?
Which weights and benchmarks are supplied?
Are recruitment-source indicators available?
How are exclusions and quality checks documented?

A completion rate alone does not establish sample quality.

Digital Tools for Sampling

Random selection

Spreadsheets, statistical packages, and programming languages can generate random numbers and select cases. Reproducibility is improved by storing:

The cleaned frame version
Selection code
Random seed
Selection date
Eligibility rules
Final selected identifiers

Survey-design analysis

Researchers analysing complex samples can use:

R packages designed for survey analysis
Stata survey commands
SPSS Complex Samples
SAS survey procedures
Specialist official-statistics software

The chosen software should incorporate the applicable weights, strata, clusters, and variance-estimation method.

Recruitment and data-collection platforms

Survey and research platforms can support:

Eligibility screening
Random invitation lists
Quota controls
Multilingual instruments
Contact scheduling
Duplicate checks
Recruitment-source tracking
Response monitoring
Audit logs

Software does not make a non-random recruitment process random. The sampling classification depends on how units enter the selection process.

Artificial Intelligence and Sampling

Artificial intelligence can assist with some sampling-related tasks, including:

Finding duplicate or inconsistent frame records
Classifying establishments or documents into candidate strata
Predicting contact difficulty
Monitoring differential response
Flagging potentially automated or fraudulent submissions
Supporting adaptive fieldwork
Reviewing whether a sampling description omits key information

Machine-learning methods are also increasingly used in data editing, nonresponse adjustment, model-assisted estimation, data integration, and the analysis of non-probability sources (Rao & Lohr, 2025).

Risks and limitations

AI may reproduce errors or inequalities in the data used to build or apply it. A model used to prioritise recruitment can systematically deprioritise groups that have historically been difficult to contact. Automated exclusions may remove legitimate participants whose response patterns differ from the majority.

Researchers should therefore:

Retain human oversight
Validate automated decisions
Test subgroup error rates
Protect personal data
Document models, inputs, thresholds, and exclusions
Preserve an auditable selection process
Report material AI use
Avoid treating synthetic respondents as substitutes for human population evidence without strong validation

Current AAPOR guidance calls for disclosure when AI is used in data collection or processing and for explanation of how AI-generated or synthetic responses were produced (AAPOR, 2026).

How to Report Sampling in a Research Paper

A transparent sampling section should answer the following questions.

1. What was the target population?

Define population characteristics, location, and time period.

2. What was the sampling frame?

Name the register, address file, institution list, venue frame, recruitment source, or panel. Describe important coverage limitations.

3. Which sampling method was used?

Name the exact method rather than writing only “random sampling” or “purposive sampling.”

4. How was selection implemented?

Report:

Random start or random seed
Sampling interval
Strata and allocation
Cluster stages
PPS size measure
Purposive criteria
Quotas
Initial snowball seeds
Referral limits
Recruitment dates

5. How was sample size determined?

State the outcome, confidence and precision assumptions, power calculation, qualitative rationale, design effect, nonresponse allowance, or information-power considerations.

6. What happened after selection?

Report invitations, eligibility, refusals, unreachable units, exclusions, completions, withdrawals, and missing data.

7. Were weights used?

Explain design, nonresponse, post-stratification, raking, calibration, or other adjustments.

8. How did the analysis reflect the design?

Identify weights, strata, clusters, replicate weights, and variance-estimation procedures.

9. What are the limitations?

Acknowledge frame gaps, nonresponse, self-selection, undercovered groups, small clusters, network dependence, or uncertain transferability.

Sampling Methodology Template

The target population comprised [define population, location, and period]. The sampling frame was [name and describe the frame or recruitment source], which covered [describe coverage] but may have excluded [important limitations]. We used [exact sampling method] because [methodological justification]. Selection was implemented by [randomisation, interval, strata, cluster stages, criteria, quotas, or referrals]. The planned sample size of [number] was based on [precision, power, subgroup, information-power, or feasibility rationale]. Of [number] units approached or selected, [number] were eligible and [number] participated. Analyses incorporated [weights, strata, clusters, or adjustment procedure]. The principal sampling limitations were [limitations], which restrict interpretation by [explain effect on inference or transferability].

Worked Sampling Examples

Example 1: National student survey

Objective: Estimate student satisfaction and compare degree levels.

Suitable design: Stratified multistage probability sampling.

Stratify universities by region and institutional type.
select universities within each stratum.
obtain student registers from selected universities.
stratify students by degree level.
randomly select students.
weight for selection probability and nonresponse.

This design supports population estimates if coverage, implementation, response adjustment, and analysis are adequate.

Example 2: Hospital record review

Objective: Estimate the proportion of records containing a required consent form.

Suitable design: Systematic probability sampling.

A random start is selected, followed by every (k)th eligible record. The team checks whether filing patterns create periodicity.

Example 3: Rural school study

Objective: Estimate access to science laboratories across a large region.

Suitable design: Cluster or multistage sampling.

Districts are selected first, then schools, followed by students or facilities within schools. Cluster identifiers and weights are retained for analysis.

Example 4: Experiences of doctoral isolation

Objective: Understand how isolation differs across study settings.

Suitable design: Maximum-variation purposive sampling.

Participants are selected across disciplines, study stages, domestic and international status, full-time and part-time enrolment, and laboratory- and desk-based programmes.

The goal is to examine variation in experience, not estimate the percentage of all doctoral students experiencing isolation.

Example 5: Hidden occupational population

Objective: Explore working conditions among people whose employment is informal and not listed in official registers.

Suitable design: Snowball, time-location, or respondent-driven sampling, depending on population networks and study aims.

The report should explain initial contacts, recruitment chains, venue coverage, confidentiality protections, and the limits of population estimation.

Example 6: Fast online opinion study

Objective: Obtain rapid feedback on several website designs.

Suitable design: A quota-based opt-in sample may be adequate for exploratory design feedback.

It should not automatically be presented as a probability sample of all internet users. Findings should be described as feedback from the recruited panel unless defensible modelling supports a broader claim.

Common Sampling Mistakes

Calling any selected sample “random”

A sample is random only when a defined chance mechanism determines selection. Approaching whoever happens to be available is convenience sampling.

Confusing a sample-size calculation with a sampling design

A formula may indicate how many observations are desirable, but it does not specify how to recruit them.

Claiming probability sampling eliminates bias

Probability sampling controls the selection mechanism. It does not eliminate frame error, nonresponse, measurement error, attrition, or processing mistakes.

Treating equal and known probabilities as identical

All simple random samples use equal probabilities. Not all probability samples do. Oversampling and PPS designs can use unequal but known probabilities.

Confusing stratified and quota sampling

Both create subgroups, but only stratified probability sampling randomly selects units within those subgroups.

Confusing stratified and cluster sampling

Stratified designs sample from every stratum. Cluster designs select only some clusters.

Ignoring the sampling frame

Researchers sometimes describe the target population but never explain the list or channel through which participants became selectable.

Substituting convenient participants for nonrespondents

Replacing a randomly selected nonrespondent with an easily available person breaks the original selection process unless replacement was built into the design.

Ignoring complex design during analysis

Treating clustered or unequally weighted data as a simple random sample can produce incorrect population estimates and standard errors.

Reporting a conventional margin of error for an opt-in sample

A probability-sample margin of error arises from a probability design. For a non-probability sample, any precision measure must be connected to an explicit model and its assumptions.

Using “saturation” without explaining it

Qualitative researchers should describe what form of adequacy or saturation was evaluated and how the assessment affected recruitment.

Overgeneralising beyond the accessible population

A sample from one institution, platform, clinic, or country should not automatically be described as representing a much broader population.

Conclusion

Sampling methods determine whose data enter a study and what researchers may reasonably conclude from those data. Probability methods are generally preferred when the goal is to estimate characteristics of a defined population with measurable sampling uncertainty. Non-probability methods remain valuable for exploratory, qualitative, specialist, rapid, and hard-to-reach population research.

The best sampling method is not simply the most complex or the easiest to implement. It is the method that aligns with the research question, population, available frame, required inference, ethical obligations, resources, and planned analysis—and that can be described transparently enough for readers to evaluate its strengths and limitations.