QA and QC are not the same thing
The distinction matters to us, because finding an error is not enough. The aim is to build the system that stops the error recurring.
- QA — Quality Assurance
- Focuses on preventing the error from occurring.
- QC — Quality Control
- Focuses on detecting and correcting the error that did occur.
Quality by Design
We start quality at the design stage. The QA/QC process comes into play before fieldwork begins: the sampling plan, the questionnaire, skip and filter structures, mandatory fields, logic rules, coding structure, interview protocol and enumerator training are all checked.
In electronic questionnaires, errors that can be prevented are prevented before any data is collected. If a participant under 18 is not shown a question meant only for adults, that is not a programming detail. It is a quality measure.
Pilot Testing
We find the large errors on a small field. The pilot is the first real test of the QA/QC system. During it we do not only look at whether the survey works.
Necessary revisions follow. An error noticed in 30 interviews is manageable; the same error noticed after 10,000 is a data quality problem.
- Is the interview duration realistic?
- Are the questions understood?
- Do participants struggle with particular questions?
- Do the filters work correctly?
- Can enumerators apply the questions consistently?
Interviewer Quality
Data quality starts with the enumerator. If the same question is applied differently by two enumerators, two different datasets can result.
Training therefore does not consist of reading out the questionnaire. Teams are trained on the purpose of the study, the target group, the standard for asking a question, neutrality, probing rules, sensitive questions, ethics, device use and the field protocol.
But training alone is not enough. Performance has to be monitored once fieldwork starts.
Real-Time Quality Monitoring
We see quality while the field is still running. Digital collection systems show us not only the answers but the field process itself.
Daily checks therefore go beyond the count of completed surveys: interview durations, completion rates, refusal rates, enumerator-level distributions, question-level answer patterns, GPS information, time stamps and paradata can all be examined.
Unusual records surface early as a result.
A realistic example
Is the Fastest Enumerator Really the Best One?
One enumerator completes 28 interviews a day. The others average 12 to 15. At first glance this person looks highly productive.
Then quality control shows a different picture: interview durations are very short, many questions carry the same answer pattern, GPS points cluster closely together, and some long modules barely vary.
High productivity here is not an indicator of quality. It is a risk signal. Catching the problem behind that kind of apparent success is exactly what QA/QC is for.
Back-Check
We verify that the interview actually happened. Back-checking is one of the core instruments of the field quality system: a defined share of completed interviews is re-checked.
The participant is contacted again, and whether the interview took place, basic demographic information, the interview duration and certain critical answers can be verified.
Our corporate QA/QC system applies verification rates of 20% and above depending on project structure. But what matters to us is not the rate alone. It is that the checking is risk-based.
Risk-Based QC
Not every record carries the same risk. Random back-checking matters, but random checking alone is not enough.
Some records carry higher risk: very short interviews, unusual answer patterns, enumerators with very high completion rates, long runs of the same option, interviews concentrated around the same GPS point, suspicious time stamps.
Duplicate Detection
In large-scale studies duplicate records arise for different reasons. The same person may have been interviewed twice, the same enterprise visited by two different enumerators, or the same record uploaded twice for a technical reason.
Duplicate checks can weigh ID fields, telephone, address, GPS, time, demographic combinations and answer patterns together. The aim is not to delete identical rows but to detect genuine repetitions.
Missing Data
Missing data is sometimes a technical problem and sometimes a methodological one. A missing value is not only an empty cell; it asks a question: why is it missing?
Did the participant not answer? Did the enumerator skip the question? Was there a programming error? Is the question systematically blank in a particular group?
If missing values concentrate in a particular group, the problem is not only one of data cleaning. It may be a bias problem. We therefore analyse the structure of the missingness as well.
Outliers
Not every outlier is an error. A business may have reported 500 employees. If most firms have 5 to 10, the record looks like an outlier — but it is not necessarily wrong.
We therefore do not delete outliers automatically. The logical context, the other variables, the enterprise profile and the field record are checked first.
Good data cleaning is not deleting the unusual value. It is being able to explain it.
Response Pattern Analysis
The rhythm of the answers carries information too. Some quality problems do not appear in a single variable but reveal themselves in answer patterns.
The same option marked throughout a long scale, the middle option chosen for every question, unusually low variance, particular answer combinations repeating excessively — all can be investigated.
This kind of analysis is especially valuable for detecting enumerator effects in large datasets.
Enumerator-Level Analytics
We monitor quality by team as well as by record. Each enumerator can be treated as a separate performance profile: average interview duration, refusal rate, missing data rate, back-check success, suspicious record rate and completion speed can all be compared.
This shows whether a quality problem comes from individual records or from systematic enumerator behaviour.
Field Feedback Loop
A control has value only if it returns to the field. Quality control should not only produce a report. When a problem is detected, the process feeds back into fieldwork.
If a question is consistently applied incorrectly, enumerators are retrained. If the wrong target group is being reached in a region, the field plan is revised. If a group of records is suspect, re-verification is carried out.
- Detect
- Verify
- Correct
- Prevent
Data Cleaning is not Data Manipulation
Data cleaning is not making the data look better. It is sometimes misunderstood: the aim is not to improve the appearance of the results.
The aim is to correct genuine errors, flag uncertainties, make the data structure transparent and produce a reliable analysis-ready dataset. Every correction should be as traceable as possible.
Audit Trail
You have to know what changed. In corporate data quality systems, keeping a record of the changes made matters.
A cleaning log, correction log, version control and issue tracker can be used as required, so it is possible to follow which record changed, why, by whom and on what date.
This matters particularly for data integrity in international projects.
Data Quality Dimensions
Quality data is not one single thing. We assess data quality across several dimensions, and quality is only meaningful when they are considered together.
- Accuracy
- How close is the data to reality?
- Completeness
- How complete is the required information?
- Consistency
- Are the variables compatible with one another?
- Validity
- Do the values comply with the defined rules?
- Timeliness
- Is the data available at the moment it is needed?
- Uniqueness
- Does the same observation appear more than once?
A Large Survey Is a Quality System
In a field study the size of the World Bank–KOSGEB survey — 11,062 enterprises across all 81 provinces — quality cannot be managed by looking at the final dataset.
In operations of this kind, field monitoring, data verification and back-check processes run together inside a structured QA/QC system. Our corporate verification standard of at least 20% is explicitly defined in the project documentation.
In the same way, digital routing and validation rules, interview durations, completion rate, paradata, duplicate, missing data and logical consistency checks are core parts of the quality system.
A large field operation is not only a data collection system. It is a quality management system.
