A blank cell in a data table can represent four entirely different things: the question was never shown due to skip logic; the respondent saw it but refused or could not answer; the questionnaire was abandoned before reaching that question; or the selected person never participated at all. If missing data handling starts with “what should I fill in for all blanks,” the generation process has already been erased.
The correct order is to first reconstruct who was supposed to provide which information, where in the process the missing value arose, and whether the missingness correlates with known characteristics, and only then choose between valid denominators, complete-case analysis, weighting adjustments, imputation, or sensitivity analysis. Imputation is not about recovering true answers but about generating usable values for estimation under explicit assumptions; it cannot turn the unobserved into the observed, nor automatically eliminate bias.
Using a Denominator Tree to Reconstruct Blanks in Post-Sales Satisfaction
Suppose a survey has 1,000 eligible respondents, and 800 of them complete the questionnaire. Post-sales satisfaction is only shown to those who actually contacted customer service in the past 90 days:
1,000 eligible respondents
├─ 200 did not complete or start
└─ 800 completers
├─ 500 did not contact CS in last 90 days → Question not shown
└─ 300 contacted CS → Should answer satisfaction
├─ 250 provided a rating from 1 to 5
├─ 30 chose “Don’t know / can’t recall”
└─ 20 saw the question but did not answerIf the research question is “how satisfied are actual customer service users with their most recent interaction,” the most direct descriptive denominator is the 250 respondents who provided a valid scale rating, while disclosing that 50 out of the 300 eligible respondents did not provide a rating. The 500 who did not use customer service are not missing a satisfaction score, nor should they be scored as 0; for this question, they are structurally inapplicable.
If a report states “31.25% satisfaction among the 800 completers,” it places those who did not contact customer service in the wrong denominator. If it merely says “80% satisfaction among valid respondents,” it hides the fact that one-sixth of eligible respondents provided no rating. A denominator tree separates the two issues: what did valid respondents express, and might those without a rating differ from them?
Code Each Unobserved State Separately
Structural Non-display results from legitimate skip logic or research design, such as asking about the purchase experience only of purchasers. It indicates the question is not applicable to that person and typically should not be imputed as an experience rating.
Item Non-response occurs when a respondent completed other parts of the questionnaire but did not provide a valid value for a specific item, including refusal, “don’t know,” device malfunction, or answers invalidated by editing rules. Refusals and “don’t know” should preferably be recorded separately, as the former may relate to sensitivity and the latter may reveal boundaries of knowledge or memory.
Survey Break-off often has a continuous pattern: the remainder of the questionnaire is blank after a certain page. It affects later items and may correlate with questionnaire length, device, question sensitivity, or motivation. Deleting all partial responses discards valid information given before the break and may make the retained sample look more like a highly engaged subgroup.
Unit Non-response means that an eligible sampled or invited unit did not produce a usable questionnaire. Often there is no full data row available for item-by-item imputation, so handling typically relies on sampling frame information, response propensities, weighting adjustments, or follow-up studies. The U.S. Census Bureau’s SIPP methodology note explicitly distinguishes unit non-response from item non-response: the former is mainly addressed through weighting adjustments, while the latter may be handled using approaches such as editing or statistical imputation.
Post-processing Missingness occurs when an original value is excluded from the analysis due to out-of-range constraints, privacy protection, or quality rules. The original value and the reason for exclusion should be preserved and not conflated with a respondent’s active non-response.
The Missing Data Profile Should Cover Item, Location, Group, and Time
For each key variable, compute at least three numbers: the number who were supposed to answer, the number with valid responses, and the number of missing responses by reason. The U.S. Census Bureau’s approach to item response rates also uses “units required to answer the item” as the denominator, rather than indiscriminately including all sample cases.
Then map the missing data distribution to:
- Position in the questionnaire and the preceding page, to identify break-offs starting at a sensitive or high-burden module;
- Device, browser, language, channel, and survey version, to identify technical or translation problems;
- Known demographic and behavioral variables, to assess whether missingness is concentrated in a particular type of participant;
- Collection date and batch, to check for a specific configuration, recruitment, or reminder change;
- Observable proxies for key outcomes, such as whether those who fail to give a satisfaction rating are more likely to complain or churn earlier.
These comparisons cannot prove what values the missing respondents would have truly given, but they can rule out the careless assumption that “missingness is completely unstructured” and guide which auxiliary variables should enter the model.
Translate MCAR, MAR, and MNAR into Checkable Questions
Statistical textbooks often categorize missing data mechanisms into Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR). In practice, do not just assign abbreviations to the data. Instead, ask what assumptions the chosen handling method requires to be valid.
MCAR implies that missingness depends neither on observed information nor on the unobserved answer itself. For example, a brief random system failure unrelated to respondent characteristics might approximate this condition. If it holds, analyzing only complete cases mainly sacrifices precision, not necessarily introducing systematic bias; however, real surveys rarely prove it convincingly with a single table.
MAR implies that, after controlling for observed variables, missingness no longer depends on the unobserved answer. For instance, younger respondents and mobile users might be more likely to break off, and both age and device are fully recorded; the analysis handles missingness after adequately controlling for these. This remains an assumption that cannot be fully verified from the data alone; it simply lets observable information bear the explanatory burden.
MNAR implies that missingness is related to the unobserved answer itself, even after controlling for observed information. Higher-income individuals being more reluctant to report income, or highly dissatisfied respondents exiting without rating, could fall into this category. Standard complete-case analysis or conventional imputation cannot automatically fix this; it typically requires external information, follow-ups, selection models, or scenario-based sensitivity analysis to express uncertainty.
The Census Bureau’s current research note on missing data emphasizes that imputation and weighting adjustments leverage non-missing survey information or other data sources and require both statistical models and subject-matter expertise. This also clarifies that “just picking an algorithm” does not mean the missing data mechanism has been resolved.
Five Common Handling Approaches Address Different Problems
Valid or Complete Case Analysis uses only records that have values for all required variables; it is simple and easily reproducible. When missingness is minimal, does not affect decisions, or the mechanism approximates unstructured, it can serve as the primary or baseline analysis. However, when missingness concentrates in specific groups, it can alter the sample and cause the sample size to differ across models.
Available-Case Analysis Per Indicator lets each descriptive figure use the existing valid values for that particular question, preserving more data. The cost is that denominators differ among figures, and correlational or multivariate analyses cannot be pieced together casually. Each outcome must clearly report the actual number of cases.
Treating “Don’t Know” as a Meaningful State is appropriate for questions where knowledge, awareness, or eligibility itself is part of the conclusion. For example, regarding awareness of a policy, “don’t know” may be a genuine result. But for continuous income, coding refusals as a single numeric category does not complete the analysis.
Weighting Adjustments primarily target unit non-response or unbalanced sample response propensities. By leveraging the sampling frame and known characteristics, they aim to make respondents representative of the full target population. However, they can only correct for known dimensions included in the model; differences between respondents and non-respondents on unobserved factors can still exist.
Imputation uses logical rules, similar records, or statistical models to generate one or more plausible values for the missing entries. Single mean imputation compresses variance and distorts correlations, often unsuitable as a general-purpose solution. Multiple imputation generates several complete datasets and combines estimates to reflect some of the imputation uncertainty, but the results still depend on the models, variables, and assumptions about the mechanism. All imputed values should be flagged and not mixed with originally reported values.
Use Extreme Scenarios to Decide if Missingness is Actionable
Returning to the post-sales satisfaction example. Among the 250 valid responses, 200 are satisfied (80%), and 50 should have answered but did not provide a rating. If all 50 are assumed satisfied, the overall satisfaction rate is 250 out of 300, approximately 83.3%. If all 50 are assumed dissatisfied, the rate is 200 out of 300, approximately 66.7%. These extremes are not estimates of the true values for the missing respondents. They answer a decision question: under any plausible boundary, does the conclusion remain unchanged?
If the team’s action threshold for satisfaction is 75%, the missing data assumptions can flip the decision, making it inadequate to simply report 80%. The next steps should compare the channel, complaints, and usage characteristics of the missing group, consider follow-up studies or more refined models, and present scenario results alongside the main conclusion. If all plausible scenarios support the same action, there is no need to force the most complex imputation just for technical sophistication.
For model-based analyses, compare complete cases, assigned imputations, different auxiliary variables, and non-random shift scenarios. For longitudinal studies, test a plausible range where dropouts are worse off or better off than stayers. The point of sensitivity analysis is not to generate many numbers but to pinpoint which unverified assumption changes the decision.
First Reduce Avoidable Missingness, Then Discuss Imputation
Structural non-display should result from clear skip logic, and data should retain a flag for “eligible to answer.” Place key metrics where respondents still have energy, use reference periods that are easy to recall for questions, explain the purpose of sensitive questions, offer a legitimate refusal option, and avoid making them mandatory without strong justification. Multilingual and mobile versions need pre-testing, and before official data collection, verify that export coding distinguishes between ‘not shown,’ ‘don’t know,’ ‘refused,’ and ‘break-off.’
Also record the flow of the sample through invitations and questionnaires, rather than only retaining completers. Without the number of eligible invitations, contact outcomes, and basic sampling frame variables, it is difficult to diagnose unit non-response. If the business only tracks current users and has no information whatsoever on non-respondents, one should acknowledge the limits of what weighting can achieve.
Within survey tools, use questionnaire logic to record the population to which a question applies, distinguish complete from partial and quality-excluded cases after collection, and show the true denominator for each metric in the analysis. AI-assisted analysis should not independently interpret blanks as 0, negative, or neutral responses, nor silently impute values without a model explanation. These handling decisions should be recorded back in the survey data codebook and the pre-collection analysis plan, so that the same metric uses a consistent denominator across charts and reports.
The final report should at minimum disclose: the target population and response base, the reasons and proportions of missingness for key variables, the main handling methods and assumptions, the variables used in imputation or weighting, the proportion of imputed values and their flags, and whether conclusions change compared with alternative handling strategies. No single method will make missingness disappear; a mature analysis shows the reader where data is missing, what the results depend on, and how far the evidence could still be from the truth.