Sampling is not just "n"
When an institution comes to us asking "how many people should we interview?", that is not our first question. We look at a set of things first.
Sample size is settled after that. Because n should be the output of the methodology, not its starting point.
- Which population are we representing?
- Which subgroups do we want to analyse separately?
- How is the geographic distribution shaped?
- How good is the list or sampling frame?
- Where is the risk of non-response highest?
- Can the field actually implement this design?
Probability Sampling
The idea behind probability sampling is simple: knowing how, and with what probability, a unit in the target population entered the sample.
ID’s applied experience covers stratified, cluster, multi-stage and systematic sampling designs. Stratification, clustering, allocation, selection probabilities, design weights and non-response adjustments are handled together rather than one after another.
A design question
A Provincial Map Is Not Always a Population Map
Imagine a youth study across Türkiye. Run a large number of interviews in İstanbul and very few in the smaller provinces and you get a distribution that looks sensible for national estimates.
But if the institution’s question is how regions differ from one another, the design has to change. Now it is not only the national total that matters but the precision of the regional estimates — which may mean allocating more sample to some regions than their population share would suggest.
Those unequal selection probabilities are then corrected with weights. It is one of the clearest illustrations of why sampling and analysis should never be separated.
Sample Size & Statistical Power
A larger sample is not always better. The aim in setting sample size is not to reach the highest number possible; it is to find the size that can answer the research question with adequate precision and statistical power.
Depending on the study, confidence level, margin of error, expected proportion, design effect, subgroup requirements, expected non-response and statistical power are weighed together.
Power analysis matters most in experimental studies and impact evaluations: too small a sample and the analysis may miss a real effect; too large and you pay in cost, field burden and time for precision nobody needed.
Sampling Weights
Two people’s answers can sit side by side in a dataset and still not carry the same representational weight. One may have had a 1 in 100 chance of selection, the other 1 in 1,000.
Real fieldwork is more complicated than that first weight. Final weights may therefore apply design weights, non-response adjustment, post-stratification, calibration and raking.
Our sampling work is as much about computing selection probabilities, building design weights, making non-response corrections and producing the final analysis weights as it is about drawing the sample.
Weight = 1 / Selection Probability
Design Effect
In real fieldwork, participants are rarely selected independently from across the whole country. People living in the same neighbourhood, village or cluster are more likely to resemble one another, and that similarity reduces the effective information in the sample.
Complex sampling designs therefore account for the design effect (DEFF). A simple random sample of 1,000 may look sufficient in theory, but with a design effect of 1.5 the same precision needs a larger sample.
The point is not to inflate the formula. It is to let the real structure of the field enter the statistical calculation.
Non-response Is Data Too
Looking only at who answered is not enough. Who did not answer is sometimes the most important question in the study.
If participation is systematically lower among young people, among higher income groups, in particular sectors or particular regions, the resulting data can be biased.
Non-response is therefore treated not as a count of missing interviews but as a potential representation problem, and corrected by strata, region, demographics or other auxiliary variables where needed.
Statistics: From Description to Explanation
Statistics does not stop at producing percentages. The first result of a study is usually simple. Our next questions are not.
Our analytical capacity therefore covers descriptive statistics, hypothesis testing, regression modelling, logistic regression, multivariate analysis, factor analysis, reliability and validity analysis, index construction and impact analysis.
- Where is it higher?
- Is the difference between groups meaningful?
- Could that difference be driven by another variable?
- Which factors explain the result?
- How far does one result predict another?
A data story
A Percentage Is Sometimes the Beginning, Not the End
Say an enterprise survey finds that 40% of firms report difficulty accessing finance. On its own that is a useful finding. For a decision it may not be enough.
The statistician asks what comes next: what is the rate among micro-enterprises? Does it change for women entrepreneurs? Is there a sector effect? Does the regional difference hold? Does the result survive controlling for firm age and headcount?
This is where a simple percentage becomes a multivariate analysis — and where data moves from description to explanation.
Measurement, Reliability & Validity
Statistics does not only count people or firms. Sometimes what needs measuring cannot be observed directly: satisfaction, trust, quality of life, perception, attitude, institutional capacity.
In those cases the scale itself has to be tested. Cronbach’s Alpha, ICC, Kappa, EFA, CFA, SEM, ROC analysis and validity tests are used as required.
The aim is not to have used an advanced method. It is to show how reliable the measurement is, and to choose the method the research question actually calls for.
Academic Depth + Field Reality
What distinguishes ID’s sampling and statistics capacity is that theoretical sampling knowledge and real field implementation sit inside the same team.
Our technical leadership combines more than 20 years of official statistics and sampling work at TurkStat with more than 15 years of field experience in private sector and international research — from industry, building, agriculture and tourism surveys through to today’s large-scale household and enterprise studies.
So when we design a sample we look not for the best model in theory, but for the strongest model that can actually be implemented in the field.
