Research Methodology ยท Faculty Teaching Resource
Sampling &
Its Types
A complete visual framework for teaching sampling theory, methods, selection logic, bias, and application in health research โ designed for Community Medicine faculty.
2
Major Categories
10+
Sampling Methods
7
Teaching Modules
โ
Application Contexts
Core Concepts in Sampling
Before teaching any sampling method, students must master the foundational vocabulary. These are the building blocks upon which all sampling theory rests.
๐ก
The Central Problem of Sampling
We can almost never study an entire population. Sampling is the science of selecting a subset (sample) from the population in a way that allows us to make valid inferences about the whole. The quality of your research is only as good as the quality of your sample.
๐ Master Mnemonic โ Elements of Sampling
Pop
Study
Sampling
Sample
Statistic
Target Population
The entire group about which you want to draw conclusions.
Example: All adults with hypertension in India
Often too large to study directly โ need a sample
Example: All adults with hypertension in India
Often too large to study directly โ need a sample
Study Population
The accessible subset of the target population from which you will actually sample.
Example: Adults with hypertension in Cuttack district
Must be clearly defined with inclusion/exclusion criteria
Example: Adults with hypertension in Cuttack district
Must be clearly defined with inclusion/exclusion criteria
Sampling Frame
A complete list of all sampling units in the study population. The foundation of probability sampling.
Example: Voter list, hospital register, school roll
If the frame is flawed, the sample is flawed
Example: Voter list, hospital register, school roll
If the frame is flawed, the sample is flawed
Sampling Unit
The element selected at each stage of sampling. May be individual, household, village, clinic, etc.
Example: Individual (simple), Household (cluster), School (two-stage)
Must match your research question
Example: Individual (simple), Household (cluster), School (two-stage)
Must match your research question
Parameter vs Statistic
Parameter: True value in the population (unknown; what we want)
Statistic: Value calculated from the sample (known; our estimate)
Example: True prevalence of TB (parameter) vs. prevalence in our sample (statistic)
Statistic: Value calculated from the sample (known; our estimate)
Example: True prevalence of TB (parameter) vs. prevalence in our sample (statistic)
Sampling Error vs Bias
Sampling Error: Random variation; unavoidable; decreases with larger sample size
Sampling Bias: Systematic error; does NOT decrease with larger sample; caused by flawed method
A large biased sample is WORSE than a small unbiased one
Sampling Bias: Systematic error; does NOT decrease with larger sample; caused by flawed method
A large biased sample is WORSE than a small unbiased one
Key Terminology Reference
| Term | Definition | Symbol | Example in Health Research | Common Confusion |
|---|---|---|---|---|
| Population (N) | All individuals about whom inference is to be made | N | All TB patients in India | Confused with study population (accessible group) |
| Sample (n) | Subset of population actually studied | n | 200 TB patients in Odisha | Students often confuse sample size with sample method |
| Sampling Fraction | n/N โ proportion of population included | f = n/N | 200/20,000 = 1% of TB patients studied | Larger fraction โ better sample if method is biased |
| Representativeness | Degree to which sample reflects population characteristics | โ | Sample has same age/sex distribution as source | Large sample โ representative sample |
| Precision | Narrowness of confidence interval; repeatability | 1/SE | Prevalence: 12% ยฑ 2% vs 12% ยฑ 8% | Precision โ accuracy (can be precise but biased) |
| Confidence Interval | Range likely to contain the true parameter | CI | 95% CI: 10%โ14% for TB prevalence | NOT "95% chance the true value is in this range" |
| Design Effect (DEFF) | Ratio of variance with complex design vs SRS | DEFF | DEFF=2 means cluster sampling needs 2ร sample size | Students forget to account for DEFF in cluster studies |
| Intraclass Correlation | Similarity of units within a cluster | ICC/ฯ | Children in same school more similar in vaccination status | Higher ICC โ bigger DEFF โ need more clusters |
Sampling Taxonomy โ The Complete Tree
The entire universe of sampling methods, organised hierarchically. Click any method to see its full profile.
๐๏ธ Sampling Methods in Research
โฌ
๐ฒ Probability Sampling
Every unit has a known, non-zero chance of selection
Every unit has a known, non-zero chance of selection
Simple Random Sampling (SRS)
Equal probability for each unit; lottery or random number table
Systematic Random Sampling
Every kth unit from a list; k = N/n (sampling interval)
Stratified Random Sampling
Divide into strata, then SRS within each stratum
Cluster Sampling
Randomly select groups (clusters); study all units in selected clusters
Multi-stage Sampling
Two or more levels of random selection; used in national surveys
PPS Sampling
Probability Proportional to Size; clusters selected by size weight
๐ Non-Probability Sampling
Selection based on judgement, availability, or convenience
Selection based on judgement, availability, or convenience
Convenience Sampling
Whoever is available and willing; quickest but most biased
Purposive (Judgement) Sampling
Deliberate selection based on researcher's judgement of typicality
Quota Sampling
Pre-set quotas for subgroups; no random selection within quotas
Snowball Sampling
Existing participants recruit future participants; hidden populations
Volunteer / Self-selection
Participants come forward on their own; extreme volunteer bias
Consecutive Sampling
All eligible patients in sequence over time; common in clinical studies
๐ Mixed / Special Methods
Specialised designs for specific research contexts
Specialised designs for specific research contexts
Lot Quality Assurance Sampling (LQAS)
WHO method for monitoring health program coverage; accept/reject lots
Respondent-Driven Sampling (RDS)
Advanced snowball with mathematical weights; drug users, sex workers
Adaptive Cluster Sampling
Add neighbours when rare event found; wildlife, rare disease research
Theoretical Sampling
Grounded theory; sample until theoretical saturation is reached
EPI 30ร7 Cluster Sampling
WHO's immunisation coverage method; 30 clusters ร 7 children
๐ Teaching Tip: After showing this tree, ask students: "Which method would you use to study TB prevalence in a district of 500 villages?" โ Let them debate before teaching. The correct answer involves multi-stage cluster sampling. The debate reveals their conceptual gaps.
Probability Sampling โ Methods in Depth
Every unit has a known, non-zero probability of being selected. This is the gold standard for quantitative health research as it allows generalisability.
โ
Why Probability Sampling is Preferred
Allows statistical inference โ you can calculate confidence intervals, test hypotheses, and generalise to the population. The mathematical theory of sampling is built on probability. Without probability sampling, you cannot legitimately extrapolate beyond your sample.
| Method | Mechanism | Sampling Frame Needed? | Best Used When | Advantages | Disadvantages | Real-World Example |
|---|---|---|---|---|---|---|
| Simple Random SRS |
Lottery / Random number table / Computer random; each unit has equal probability = n/N | Yes (complete) | Small, homogeneous, well-listed populations | Unbiased; easy to understand; forms basis of statistical theory | Requires complete sampling frame; not practical for large dispersed populations; may miss minorities | Selecting 200 patients from hospital register of 2000 using random numbers |
| Systematic Every kth unit |
k = N/n; random start between 1 and k; select every kth unit thereafter | Yes (ordered list) | Large populations with sequential lists (registers, records) | Easy to execute; spread across list; no need to number all units if list exists | Periodicity bias if list has periodic pattern (e.g., every 7th patient is a Monday case) | Antenatal clinic: k=10; randomly start at 4, then select 4, 14, 24, 34... |
| Stratified SRS within strata |
Divide population into homogeneous subgroups (strata); SRS within each stratum; proportional or disproportionate allocation | Yes (per stratum) | Heterogeneous populations; need subgroup estimates; want to ensure minority representation | Ensures representation; reduces variance vs SRS; permits subgroup analysis | Complex; must know stratum sizes; disproportionate allocation requires weighting | NFS surveys: stratify by urban/rural, then state, then household type |
| Cluster Whole clusters selected |
Divide population into clusters; randomly select clusters; study ALL units in selected clusters | No (only cluster list) | Geographically dispersed; no individual sampling frame; field surveys | Practical; cost-efficient; no complete frame needed; feasible in field | Less precise than SRS (clustering effect); DEFF >1; need more subjects | District nutritional survey: randomly select 30 villages; study all children in selected villages |
| Multi-stage Nested random |
Two or more stages of random selection; each stage uses a sampling frame for that level | Yes (at each stage) | Large national studies; hierarchical populations (districtsโblocksโvillagesโhouseholds) | Practical for national surveys; economical; flexible design | Complex; errors compound across stages; needs lists at each level | NFHS: State โ District โ PSU (village/ward) โ Household โ Individual |
| PPS Sampling Size-weighted |
Probability of selecting a cluster proportional to its size; ensures equal probability of individual selection | Yes (with size data) | Clusters of unequal size; want equal individual probability | Equal probability for all individuals; no weighting needed in analysis | Need size information for all clusters; complex to implement | EPI cluster sampling: villages selected proportional to their population size |
โ๏ธ Key Formulas โ Probability Sampling
Sampling Interval (k) = N รท n
N = Population size ยท n = Required sample size ยท k = Sampling interval (Systematic sampling)
Proportional Allocation: nแตข = n ร (Nแตข / N) where Nแตข = stratum size
Design Effect (DEFF) = 1 + (mโ1) ร ICC ยท m = cluster size, ICC = intraclass correlation
Effective sample size = n รท DEFF ยท Required n (cluster) = SRS sample ร DEFF
Proportional Allocation: nแตข = n ร (Nแตข / N) where Nแตข = stratum size
Design Effect (DEFF) = 1 + (mโ1) ร ICC ยท m = cluster size, ICC = intraclass correlation
Effective sample size = n รท DEFF ยท Required n (cluster) = SRS sample ร DEFF
Non-Probability Sampling โ Methods in Depth
Selection is not based on random chance. Not all units have a known probability of selection. Used in qualitative research, exploratory studies, and when probability sampling is impossible.
โ ๏ธ
When is Non-Probability Sampling Justified?
Non-probability sampling is not inherently wrong โ it is appropriate for qualitative research, pilot studies, exploratory work, hidden populations, and situations where probability sampling is logistically impossible. The error is using it when probability sampling was feasible and then claiming generalisability.
| Method | Mechanism | Bias Risk | Best Used For | Strengths | Limitations | Health Research Example |
|---|---|---|---|---|---|---|
| Convenience Accidental |
Select whoever is readily available; patients in OPD, students in class, mall visitors | HIGH | Pilot studies; feasibility testing; quick surveys | Cheapest; fastest; easy to execute; good for hypothesis generation | Highly biased; not representative; cannot generalise; selection entirely determined by convenience | Exit interviews with OPD patients to pilot a questionnaire on patient satisfaction |
| Purposive Judgement |
Researcher deliberately selects "information-rich" cases based on specific characteristics | MODERATE | Qualitative research; case studies; key informant interviews | Focused on relevant cases; efficient for specific objectives; expert knowledge used | Researcher bias in selection; not generalisable; dependent on researcher's judgement | Selecting ASHA workers with โฅ5 years experience for in-depth interviews on community health |
| Quota Non-random strata |
Set fixed quotas for subgroups (age, sex, etc.); fill quotas by convenience within each subgroup | MODERATE | Market research; large surveys where frame unavailable; needs subgroup balance | Ensures proportional representation of subgroups; faster than stratified random; no frame needed | Selection within quota is non-random; interviewer bias; cannot calculate sampling error | Community survey: quota of 50 males and 50 females in each age group, recruited at convenience |
| Snowball Chain referral |
Initial seeds recruited; each participant refers others; chain grows like a snowball | MODERATE | Hidden or hard-to-reach populations; stigmatised groups | Only practical method for some populations; builds trust networks; reaches hidden groups | Selection bias towards well-connected individuals; network clustering; non-representative | Studying risk behaviour in IV drug users; FSW health surveys; undocumented migrants |
| Volunteer Self-selection |
Individuals volunteer in response to advertisement or invitation | VERY HIGH | Clinical trials (with random allocation after recruitment); experimental studies | Motivated participants; good compliance; ethical (consent built in) | Healthy worker effect; volunteers are atypical; extreme self-selection bias | Vaccine trial: volunteers respond to ad; then randomised to vaccine vs placebo |
| Consecutive Sequential |
Every eligible patient presenting over a defined time period is recruited; no random selection | LOWโMOD | Hospital-based clinical studies; OPD-based research | Simple; minimises selection bias within available patients; complete capture of eligible cases | Limited to patients attending that facility; selection bias due to healthcare-seeking behaviour | All newly diagnosed diabetic patients in medicine OPD over 6 months included in study |
Quota vs Stratified โ The Classic Confusion
Quota: Same subgroups; NO random selection within groups; interviewer picks convenient subjects to fill quota; cannot calculate sampling error
Stratified: Same subgroups; YES random selection within strata; unbiased within stratum; can calculate sampling error
They look similar in design but differ fundamentally in how units within groups are selected.
Stratified: Same subgroups; YES random selection within strata; unbiased within stratum; can calculate sampling error
They look similar in design but differ fundamentally in how units within groups are selected.
Snowball vs RDS โ Important Distinction
Snowball: Simple chain referral; no mathematical weights; biased towards well-connected; qualitative research
RDS: Advanced snowball with mathematical corrections for network effects; gives population estimates; used in HIV research with FSW, MSM
RDS can produce valid prevalence estimates where Snowball cannot.
RDS: Advanced snowball with mathematical corrections for network effects; gives population estimates; used in HIV research with FSW, MSM
RDS can produce valid prevalence estimates where Snowball cannot.
Comparison โ Side by Side
The master comparison table โ use this for teaching contrasts, exam preparation, and decision-making in research design.
| Criterion | SRS | Systematic | Stratified | Cluster | Multi-stage | Convenience | Purposive | Snowball |
|---|---|---|---|---|---|---|---|---|
| Type | Probability | Probability | Probability | Probability | Probability | Non-Prob | Non-Prob | Non-Prob |
| Sampling Frame Required | Yes (complete) | Yes (ordered) | Yes (per strata) | No (cluster list only) | Partial (at each level) | No | No | No |
| Representativeness | High | High (if no periodicity) | Very High | Moderate | ModerateโHigh | Low | Low | Low |
| Statistical Inference | Yes | Yes | Yes | Yes (with DEFF) | Yes (with weights) | No | No | No (RDS: limited) |
| Cost / Complexity | LowโModerate | Low | Moderate | LowโModerate | High | Very Low | Low | Low |
| Variance / Precision | Benchmark | Equal or better than SRS | Better than SRS | Worse than SRS (DEFF >1) | Variable | Cannot estimate | Cannot estimate | Cannot estimate |
| Bias Risk | Very Low | Low (periodicity risk) | Very Low | Moderate (homogeneity) | LowโModerate | High | Moderate | ModerateโHigh |
| Best Research Type | Prevalence studies; RCTs | Hospital-based; sequential lists | Comparative studies; surveys | Field surveys; national studies | NFHS; DHS; large national surveys | Pilot; qualitative | Qualitative; key informants | Hidden populations |
| Indian Health Example | PHC patient study using OPD register | Every 5th antenatal visit to clinic | Urban/rural stratified TB survey | Village-based NCD survey | NFHS-5; DLHS; AHS | Questionnaire pilot in OPD | ASHA worker interviews | IDU risk behaviour survey |
Interactive Decision Guide โ Which Sampling Method?
๐งญ Choose Your Sampling Method โ Answer These Questions
Q1. Do you need to generalise findings to a larger population? (Quantitative study?)
Yes โ Need probability sampling โ
No โ Qualitative/exploratory โ Non-probability is acceptable โ
โ Use Probability Sampling
Now consider: Do you have a complete sampling frame (list of all units)?
Yes + Small population: Use SRS or Systematic Random
Yes + Heterogeneous population: Use Stratified Random
No complete frame + Field survey: Use Cluster or Multi-stage
National scale survey: Multi-stage with PPS (like NFHS)
Yes + Small population: Use SRS or Systematic Random
Yes + Heterogeneous population: Use Stratified Random
No complete frame + Field survey: Use Cluster or Multi-stage
National scale survey: Multi-stage with PPS (like NFHS)
โ Non-Probability Sampling โ Choose Type
Pilot/quick survey: Convenience sampling
In-depth qualitative: Purposive sampling
Need subgroup balance (no frame): Quota sampling
Hidden population (IDU, FSW): Snowball or RDS
Hospital-based clinical study: Consecutive sampling
In-depth qualitative: Purposive sampling
Need subgroup balance (no frame): Quota sampling
Hidden population (IDU, FSW): Snowball or RDS
Hospital-based clinical study: Consecutive sampling
Q2. Is your population homogeneous or heterogeneous in the outcome of interest?
Homogeneous โ SRS or Systematic gives good results โ
Heterogeneous โ Stratified sampling reduces variance significantly โ
โ SRS or Systematic Sampling
When population is homogeneous, SRS gives efficient, unbiased results. Systematic is simpler operationally if a sequential list exists. Watch for periodicity in systematic sampling.
โ Stratified Random Sampling
Divide into strata that are internally homogeneous but differ from each other. Use proportional allocation for overall estimates. Use disproportionate allocation if you need equal precision for each stratum (e.g., rare ethnic minority).
Q3. Is your target population geographically dispersed and costly to reach individually?
Yes โ Cluster or Multi-stage is the practical choice โ
No โ Concentrated, accessible โ Use SRS or Systematic โ
โ Cluster or Multi-stage Sampling
Cluster: Randomly select groups (villages/schools/wards), study all units within. Increases efficiency but reduces precision (DEFF > 1). Multi-stage: Add another layer of random selection at each level โ best for national surveys like NFHS.
โ SRS or Systematic Sampling
When population is accessible and a sampling frame exists, simple approaches work best. Systematic is easiest operationally. SRS gives the theoretical gold standard.
Sample Size โ The Science of How Many
One of the most commonly asked questions in research: "How many subjects do I need?" Sample size is determined by statistical requirements, not budget or convenience.
๐
The 4 Determinants of Sample Size
Sample size is determined by: (1) Expected prevalence/effect size โ the more extreme (near 50% for prevalence), the more subjects needed. (2) Precision desired โ narrower CI needs more subjects. (3) Confidence level โ 99% CI needs more than 95% CI. (4) Power โ analytical studies need power to detect differences.
๐ Core Sample Size Formulas
Cross-sectional (Prevalence): n = Zยฒ ร p ร q / dยฒ
Z = Z-value for confidence level (1.96 for 95% CI; 2.576 for 99%) ยท p = expected prevalence ยท q = 1โp ยท d = allowable error (absolute precision)
Analytical (Comparison): n = Zยฒ ร 2pq / dยฒ (each group) or use Kelsey formula for OR/RR
Cluster adjustment: n_cluster = n_SRS ร DEFF ยท DEFF = 1 + (mโ1) ร ICC
Finite population correction: n_final = n / [1 + (nโ1)/N] (when sampling fraction >5%)
Analytical (Comparison): n = Zยฒ ร 2pq / dยฒ (each group) or use Kelsey formula for OR/RR
Cluster adjustment: n_cluster = n_SRS ร DEFF ยท DEFF = 1 + (mโ1) ร ICC
Finite population correction: n_final = n / [1 + (nโ1)/N] (when sampling fraction >5%)
| Study Type | Formula | Key Inputs | Example Calculation | Software Tool |
|---|---|---|---|---|
| Cross-sectional Prevalence estimation |
n = Zยฒpq/dยฒ | p = expected prevalence; d = allowable error; confidence level | p=0.20, d=0.05, 95% CI: n = (1.96)ยฒร0.20ร0.80/(0.05)ยฒ = 246 | OpenEpi; EpiInfo; G*Power |
| Case-Control Odds Ratio |
Kelsey / Schlesselman formula; based on OR and pโ | Expected OR; exposure prevalence in controls; ฮฑ; power (1โฮฒ) | OR=2.0, pโ=0.30, ฮฑ=0.05, power=80%: nโ133 per group | OpenEpi; EpiInfo; PASS |
| Cohort / RCT Risk difference or RR |
n = Zยฒ(pโqโ+pโqโ)/(pโโpโ)ยฒ | Incidence in exposed/unexposed; ฮฑ; power; dropout rate | pโ=0.15, pโ=0.30, ฮฑ=0.05, power=80%: nโ130/group (add 10โ20% attrition) | OpenEpi; G*Power; Stata |
| Cluster Sampling Design effect adjustment |
n_cluster = n_SRS ร DEFF | n_SRS; DEFF (assume 1.5โ2.0 if ICC unknown); cluster size m | n_SRS=246, DEFF=1.5: n_cluster = 246ร1.5 = 369; if 30 clusters: 369/30 = 13/cluster | EPI cluster; WHO LQAS tables |
| Qualitative Research Saturation-based |
No formula; theoretical saturation | Research question complexity; homogeneity of group; data richness | Typical: 15โ30 in-depth interviews; 3โ5 focus group discussions of 6โ10 participants | Not applicable; literature guidance |
Factors That Increase Sample Size โ Teaching Matrix
| Factor | Change | Effect on Sample Size | Reason | Teaching Analogy |
|---|---|---|---|---|
| Prevalence (p) | p โ 50% | โ Increases | Maximum variance at p=0.5 (pq is maximised) | If you don't know heads vs tails, you need more tosses |
| Confidence Level | 95% โ 99% | โ Increases (Z: 1.96 โ 2.576) | More certainty requires wider margin coverage | More certain = more evidence needed |
| Allowable Error (d) | 5% โ 2% | โ Greatly Increases (ร6.25) | d is squared in denominator; halving d quadruples n | Smaller target needs more shots to hit it reliably |
| Desired Power | 80% โ 90% | โ Increases | Higher power reduces Type II error; needs more data | Better detection = larger radar screen |
| Dropout/Non-response | Add 10โ20% | โ Adds buffer | Some subjects will drop out; need reserves | Order extra food in case guests bring friends |
| Cluster Effect (DEFF) | DEFF = 2 | โ Doubles n | Clustering reduces effective information per subject | Asking one village = not same as asking 30 individuals from 30 villages |
| Population Size (N) | N โ greatly | โ Little effect above N=10,000 | Large populations: FPC correction negligible | A teaspoon from the ocean gives same info as from a pool |
| Effect Size | Small effect | โ Greatly Increases | Harder to detect smaller differences | Need more tests to find a faint signal in noise |
โ ๏ธ Common Exam Trap: Increasing sample size does NOT reduce sampling bias โ it only reduces sampling error. A biased large sample is NOT better than a small unbiased sample. Students frequently confuse precision (sample size dependent) with accuracy (method dependent).
Faculty Teaching Guide
Proven strategies to make sampling genuinely understood โ not just memorised. Based on active learning and conceptual contrast teaching.
๐
The Core Teaching Challenge
Students memorise sampling method names and definitions but cannot choose the right method when faced with a real research scenario. The solution: teach decision-making, not definitions. Every sampling method should be taught by asking "WHEN would you use this โ and when would it fail?"
5-Step Teaching Sequence for Sampling
Anchor to the Problem
Start with: "We want to know the prevalence of hypertension in Odisha. How would you study 4.5 crore people?" Let students struggle โ their guesses reveal gaps you need to fill.
Teach the Framework First
Introduce the Population โ Frame โ Sample โ Statistic pathway before any method. Students who skip this confuse "who I want to study" with "who I actually studied."
Contrast Probability vs Non-Probability
Use one vivid example each: NFHS methodology (probability) vs OPD convenience sample (non-probability). Ask: "Which one can you use to claim Odisha's hypertension prevalence? Why?"
Teach Each Method with a FAIL case
For every method, show when it breaks. Systematic: the ANC register with every 7th patient being a high-risk referral. Cluster: studying one ward and claiming it represents Delhi.
Apply the Decision Matrix
Give 3โ4 research scenarios; students work in pairs to choose and defend their method. No single right answer for some โ the debate IS the learning.
Exam-Proof with Common Traps
End each class by showing the 3 most common exam mistakes: Quotaโ Stratified; Large n โ no bias; Cluster needs DEFF correction. Repetition of these traps is essential.
Exam-Focused: High-Yield Contrast Pairs
| Pair | Similarity (Why Students Confuse Them) | Key Difference | Exam Tip |
|---|---|---|---|
| Stratified vs Quota | Both divide population into subgroups before sampling | Stratified: RANDOM selection within strata โ probability method. Quota: CONVENIENCE selection within quotas โ non-probability | If random selection within groups โ Stratified. If researcher fills quotas by convenience โ Quota |
| Cluster vs Stratified | Both use groups/subpopulations as part of design | Cluster: Randomly SELECT clusters, study ALL units within. Stratified: Create strata, randomly select INDIVIDUALS within each | Cluster = select whole groups then study everything inside. Stratified = study a random sample from each group |
| Sampling Error vs Sampling Bias | Both affect accuracy of estimates from samples | Error: Random; decreases with n; unavoidable. Bias: Systematic; does NOT decrease with n; caused by poor method | A biased large sample is worse than a small unbiased one. Bias can only be fixed by changing the method, not by adding subjects |
| Multi-stage vs Cluster | Both involve selecting groups (clusters) at some point | Cluster: One-stage โ select clusters, study all. Multi-stage: Two or more stages of selection (e.g., districts โ villages โ households โ individuals) | NFHS uses multi-stage (PSUs โ households โ individuals). A simple village survey using whole villages is cluster |
| Systematic vs SRS | Both are probability methods; both give unbiased samples from lists | SRS: Truly random each time. Systematic: Periodic โ vulnerable to periodicity bias if list has cyclical pattern | If OPD register has every Monday = highest severity, systematic sampling at interval=7 will always select Monday cases โ biased |
| Snowball vs Purposive | Both are non-probability; both used in qualitative research | Snowball: Participants recruit others (chain referral); grows from initial seeds. Purposive: Researcher actively selects information-rich cases based on criteria | Snowball = participants drive recruitment. Purposive = researcher drives recruitment based on judgement |
Real Indian Health Research Scenarios โ Apply Your Method
| Research Question | Recommended Method | Justification | Sampling Frame | Practical Challenge |
|---|---|---|---|---|
| Prevalence of anaemia among adolescent girls in Odisha | Multi-stage + Stratified | Large state; heterogeneous urban/rural; hierarchical population structure | School/PHC registers at each stage; voter rolls for households | Non-school-going girls missed; consent from parents |
| Immunisation coverage in a district โ WHO survey | EPI 30ร7 Cluster (PPS) | WHO's validated method; no individual frame available; clusters selected by population size | Village list with population for PPS; no individual-level frame needed | Random walk method for household selection within cluster |
| Risk behaviour among truck drivers on national highway | Snowball / RDS | Hidden population; no sampling frame exists; trust-based recruitment needed | None available | Chain-referral bias; need multiple seeds at different highway stops |
| Comparing treatment outcomes: DOTS vs self-administered therapy in TB patients | Systematic from RNTCP register | Sequential patient list available; need unbiased allocation to study arms | District RNTCP patient register | Ensure register is complete; check for periodicity in registration patterns |
| Qualitative study on barriers to institutional delivery among tribal women | Purposive Sampling | Qualitative; need information-rich cases; tribal women with experience of home delivery | No frame; ASHA worker referrals for identification | Language barriers; trust building; purposive selection criteria must be explicit |
| Monitoring vaccine coverage in 30 PHC areas after campaign | LQAS (Lot Quality Assurance) | Need accept/reject decision for each PHC area; small sample per area; operational monitoring | PHC area population list; community health workers' records | Threshold and sample size based on LQAS tables; decision rule must be pre-specified |
| Blood pressure survey in a medical college OPD (pilot study) | Consecutive Sampling | Pilot only; all eligible patients in OPD over 2 weeks; simple and complete within available setting | OPD attendance register as guide; not a frame | Healthcare-seeking bias; generalisation limited to OPD-attending population only |
๐ Final Teaching Principle: Every sampling decision is a compromise between scientific rigour, feasibility, cost, and ethics. There is rarely one "correct" answer. Teach students to justify their choice โ explaining what they gain, what they lose, and what limitations they must acknowledge in their discussion section. That skill is what separates a good researcher from one who merely follows a textbook formula.