Sampling Methods in Field Research: What I Learned

Survey

Statistics

Field Research

From stratified to cluster sampling — practical insights from 7 years of field survey work across India.

Author

Pawan Jangra

Published

May 10, 2025

The Reality of Field Sampling

Textbooks make sampling sound clean. Reality is messier. In my years of field survey work — from rural Haryana to urban Maharashtra — I’ve encountered almost every challenge sampling theory describes, plus a few it doesn’t.

This post is my attempt to bridge the theory-practice gap.

Simple Random Sampling: Rarely Simple in Practice

Code

library(tidyverse)

# Simulate a village population
set.seed(42)
village_pop <- data.frame(
  household_id = 1:500,
  income = exp(rnorm(500, log(8000), 0.8)),
  num_members = rpois(500, 4.5) + 1,
  landholding = rexp(500, 1/2.5)  # acres
)

# SRS: draw 50 households
srs_sample <- village_pop |>
  slice_sample(n = 50)

# Compare estimates to true values
comparison <- bind_rows(
  village_pop |> summarise(group = "Population", 
                           mean_income = mean(income),
                           mean_size = mean(num_members)),
  srs_sample |> summarise(group = "SRS Sample (n=50)", 
                           mean_income = mean(income),
                           mean_size = mean(num_members))
)

knitr::kable(comparison, digits = 1,
             caption = "SRS vs Population parameters")

SRS vs Population parameters
group	mean_income	mean_size
Population	10575.1	5.4
SRS Sample (n=50)	9324.0	5.3

The estimates are close — but in field reality, SRS requires a complete sampling frame (list of all households), which rarely exists in rural areas.

Stratified Sampling: My Preferred Approach

When I joined SSER, we almost always used stratified sampling for district-level surveys. Here’s why it works:

Code

# Add strata based on landholding
village_pop <- village_pop |>
  mutate(
    stratum = case_when(
      landholding == 0         ~ "Landless",
      landholding < 1          ~ "Marginal (<1 ac)",
      landholding < 5          ~ "Small (1-5 ac)",
      TRUE                     ~ "Large (5+ ac)"
    ) |> factor(levels = c("Landless","Marginal (<1 ac)","Small (1-5 ac)","Large (5+ ac)"))
  )

# Proportional stratified sample
strat_sample <- village_pop |>
  group_by(stratum) |>
  slice_sample(prop = 0.1) |>  # 10% from each stratum
  ungroup()

# Visualise income by stratum
ggplot(village_pop, aes(x = stratum, y = income, fill = stratum)) +
  geom_boxplot(alpha = 0.7, outlier.shape = 21) +
  geom_jitter(data = strat_sample, aes(x = stratum, y = income),
              width = 0.2, alpha = 0.6, color = "black", size = 1.5) +
  scale_y_log10(labels = scales::label_number(prefix = "₹", big.mark = ",")) +
  scale_fill_manual(values = c("#2563eb","#7c3aed","#0d9488","#d97706")) +
  labs(title = "Income Distribution by Land Stratum",
       subtitle = "Black dots = sampled households (10% per stratum)",
       x = NULL, y = "Monthly Income (log scale)") +
  theme_minimal(base_size = 11) +
  theme(legend.position = "none",
        axis.text.x = element_text(angle = 15, hjust = 1))

Cluster Sampling: Practical but Tricky

Cluster sampling is often the only feasible approach for large geographic surveys. We used it extensively when surveying across multiple districts.

Code

# Simulate 20 clusters (villages), sample 5 clusters
set.seed(99)
n_clusters <- 20
clusters <- data.frame(
  cluster_id = 1:n_clusters,
  district = sample(c("Rohtak","Hisar","Karnal","Panipat"), 
                    n_clusters, replace = TRUE),
  pop_size = sample(200:800, n_clusters)
)

# Within-cluster correlation (ICC) effect
icc <- 0.15  # typical for income surveys
design_effect <- 1 + (mean(clusters$pop_size) - 1) * icc

cat("Design Effect (DEFF):", round(design_effect, 2))

Design Effect (DEFF): 84.51

Code

cat("\nEffective sample size reduction:", round((1 - 1/design_effect)*100, 1), "%")


Effective sample size reduction: 98.8 %

The key lesson: Always account for the design effect when reporting standard errors from cluster samples. Ignoring it gives you artificially small standard errors and overconfident conclusions.

Practical Tips from the Field

Listing comes first — Before sampling, build your sampling frame carefully. In rural areas, this often means a fresh household listing exercise.
Replacement rules matter — Define substitution procedures before going to field. What happens when a sampled household is locked or refuses?
Random starts in systematic sampling — Always use a random start for systematic samples, not just household #1.
Weight your data — After collecting, compute and apply sampling weights before any analysis. The survey package in R handles this beautifully.

Field sampling is as much art as science. The best designs balance statistical efficiency with logistical feasibility.

--- title: "Sampling Methods in Field Research: What I Learned" description: "From stratified to cluster sampling — practical insights from 7 years of field survey work across India." author: "Pawan Jangra" date: "2025-05-10" categories: [Survey, Statistics, Field Research] image: "https://images.unsplash.com/photo-1454165804606-c3d57bc86b40?w=600&q=80" toc: true --- ## The Reality of Field Sampling Textbooks make sampling sound clean. Reality is messier. In my years of field survey work — from rural Haryana to urban Maharashtra — I've encountered almost every challenge sampling theory describes, plus a few it doesn't. This post is my attempt to bridge the theory-practice gap. ## Simple Random Sampling: Rarely Simple in Practice ```{r} #| message: false library(tidyverse) # Simulate a village population set.seed(42) village_pop <- data.frame( household_id = 1:500, income = exp(rnorm(500, log(8000), 0.8)), num_members = rpois(500, 4.5) + 1, landholding = rexp(500, 1/2.5) # acres ) # SRS: draw 50 households srs_sample <- village_pop |> slice_sample(n = 50) # Compare estimates to true values comparison <- bind_rows( village_pop |> summarise(group = "Population", mean_income = mean(income), mean_size = mean(num_members)), srs_sample |> summarise(group = "SRS Sample (n=50)", mean_income = mean(income), mean_size = mean(num_members)) ) knitr::kable(comparison, digits = 1, caption = "SRS vs Population parameters") ``` The estimates are close — but in field reality, SRS requires a complete sampling frame (list of all households), which rarely exists in rural areas. ## Stratified Sampling: My Preferred Approach When I joined SSER, we almost always used stratified sampling for district-level surveys. Here's why it works: ```{r} #| fig-height: 5 # Add strata based on landholding village_pop <- village_pop |> mutate( stratum = case_when( landholding == 0 ~ "Landless", landholding < 1 ~ "Marginal (<1 ac)", landholding < 5 ~ "Small (1-5 ac)", TRUE ~ "Large (5+ ac)" ) |> factor(levels = c("Landless","Marginal (<1 ac)","Small (1-5 ac)","Large (5+ ac)")) ) # Proportional stratified sample strat_sample <- village_pop |> group_by(stratum) |> slice_sample(prop = 0.1) |> # 10% from each stratum ungroup() # Visualise income by stratum ggplot(village_pop, aes(x = stratum, y = income, fill = stratum)) + geom_boxplot(alpha = 0.7, outlier.shape = 21) + geom_jitter(data = strat_sample, aes(x = stratum, y = income), width = 0.2, alpha = 0.6, color = "black", size = 1.5) + scale_y_log10(labels = scales::label_number(prefix = "₹", big.mark = ",")) + scale_fill_manual(values = c("#2563eb","#7c3aed","#0d9488","#d97706")) + labs(title = "Income Distribution by Land Stratum", subtitle = "Black dots = sampled households (10% per stratum)", x = NULL, y = "Monthly Income (log scale)") + theme_minimal(base_size = 11) + theme(legend.position = "none", axis.text.x = element_text(angle = 15, hjust = 1)) ``` ## Cluster Sampling: Practical but Tricky Cluster sampling is often the only feasible approach for large geographic surveys. We used it extensively when surveying across multiple districts. ```{r} # Simulate 20 clusters (villages), sample 5 clusters set.seed(99) n_clusters <- 20 clusters <- data.frame( cluster_id = 1:n_clusters, district = sample(c("Rohtak","Hisar","Karnal","Panipat"), n_clusters, replace = TRUE), pop_size = sample(200:800, n_clusters) ) # Within-cluster correlation (ICC) effect icc <- 0.15 # typical for income surveys design_effect <- 1 + (mean(clusters$pop_size) - 1) * icc cat("Design Effect (DEFF):", round(design_effect, 2)) cat("\nEffective sample size reduction:", round((1 - 1/design_effect)*100, 1), "%") ``` **The key lesson:** Always account for the design effect when reporting standard errors from cluster samples. Ignoring it gives you artificially small standard errors and overconfident conclusions. ## Practical Tips from the Field 1. **Listing comes first** — Before sampling, build your sampling frame carefully. In rural areas, this often means a fresh household listing exercise. 2. **Replacement rules matter** — Define substitution procedures before going to field. What happens when a sampled household is locked or refuses? 3. **Random starts in systematic sampling** — Always use a random start for systematic samples, not just household #1. 4. **Weight your data** — After collecting, compute and apply sampling weights before any analysis. The `survey` package in R handles this beautifully. Field sampling is as much art as science. The best designs balance statistical efficiency with logistical feasibility.