Missing data is one of those problems almost every researcher meets, and almost nobody plans for. You collect your questionnaires, open your spreadsheet, and there it is: blank cells where answers should be. A few students skipped the income question. Three respondents left the last page blank. One interview ended early.
It is tempting to panic, or to quietly delete every incomplete row and hope nobody notices. Neither reaction helps. How you handle missing data affects your sample size, your results, and whether a reviewer trusts your findings. The good news is that you do not need advanced statistics to handle it well. You need a clear, honest process: find out how much is missing, work out why it might be missing, choose a method that fits that pattern, and report what you did.
This guide walks you through that process step by step, with a worked example you can adapt to your own thesis or paper.
What Is Missing Data, and Why Does It Matter?
Missing data simply means a value that should have been recorded was not. It can happen at the item level, when a respondent skips one question, or at the unit level, when a whole questionnaire is never returned. It can also happen during data entry, coding, or merging files.
It matters for three reasons. First, it reduces your usable sample, which reduces statistical power. Second, if the missing values are not random, deleting them can bias your results: the people who remain in your analysis may be systematically different from the people who dropped out. Third, reviewers and examiners now expect you to explain missing data. A short, transparent paragraph about it is a sign of careful research, not a weakness.
The Three Types of Missing Data
Before you choose a method, you need to think about why the data are missing. Researchers usually describe three patterns, based on ideas developed by Donald Rubin and widely used in research methods teaching.
1. Missing Completely at Random (MCAR)
The missing values have nothing to do with any variable in your study, observed or unobserved. For example, a page of a printed questionnaire was accidentally torn, or a server glitch lost a random set of responses. If data are truly MCAR, the remaining cases are still a representative sample, just a smaller one.
2. Missing at Random (MAR)
The missingness is related to other variables you did observe, but not to the missing value itself once those variables are taken into account. For example, older respondents might be more likely to skip an income question, but within each age group, skipping is unrelated to actual income. This is the most common assumption behind modern imputation methods.
3. Missing Not at Random (MNAR)
The missingness is related to the missing value itself, even after accounting for observed variables. For example, people with very high incomes may be less willing to report their income, or students with severe anxiety may skip anxiety items. MNAR is the hardest pattern to handle, because the information you need is, by definition, in the values you do not have.
You can test for MCAR patterns and look for clues about MAR, but you can never prove MNAR from the data alone. That is why reporting and sensitivity checks matter so much: you show the reader you thought about it, rather than pretending the problem did not exist.
Step 1: Count and Describe the Missing Data
Start with a simple audit before you delete or fill in anything.
- For each variable, calculate the number and percentage of missing values.
- For each case, count how many items are missing.
- Look at the overall picture: is missingness concentrated in one question, one section, or one group of respondents?
Most statistical software can do this in a few clicks: in SPSS, use Analyze, then Missing Value Analysis, or simply run Frequencies and check the missing counts. In R, packages for missing-data summaries give you the same picture.
Write these numbers down. They belong in your methods section later, and they guide every decision that follows.
Step 2: Look for a Pattern
Next, ask whether missingness looks random or systematic.
- Compare respondents with and without missing values on key variables such as age, gender, group, or pre-test score. Large differences suggest the data are not MCAR.
- Check whether missingness clusters: do later items have more blanks than earlier ones? That can signal fatigue. Do sensitive items have more blanks? That can signal reluctance.
- Plot or tabulate missingness by group. If one group has far more missing data, flag it.
If the pattern looks random and the amount is small, simple methods may be fine. If a clear pattern appears, you need a method that uses the observed information, such as multiple imputation, and you need to discuss the pattern in your write-up.
Step 3: Decide on a Strategy — Deletion or Imputation
Broadly, you have two families of choices: remove the incomplete cases or values, or fill them in with plausible estimates. Neither is always right. The choice depends on how much is missing, why it is missing, and what analysis you plan to run.
Deletion Methods
Listwise deletion, also called complete-case analysis, drops any case with a missing value on any variable in the analysis. It is simple, transparent, and the default in many programs. Its cost is lost data: if five variables each miss a different 5 percent of cases, listwise deletion can easily discard a quarter of your sample. It is most defensible when missingness is MCAR and small, typically under about 5 percent per variable.
Pairwise deletion keeps a case for every analysis where it has complete data, and drops it only where a value is missing. It preserves more data than listwise deletion, but different analyses then use different sample sizes, which can confuse readers and complicate comparisons.
Variable deletion, dropping an entire variable because it has too much missing data, is a last resort. Use it only when a variable is so incomplete that any estimate would be guesswork, and say so explicitly.
Imputation Methods
Imputation replaces a missing value with an estimate. Simple single imputation fills each blank once, while multiple imputation fills each blank several times and combines the results, which properly reflects the uncertainty of guessing.
Mean or median substitution replaces a missing value with the average of the observed values. It is easy to explain, but it artificially reduces variation and can distort relationships between variables. Use it, if at all, only when missingness is tiny and you say so.
Regression imputation predicts the missing value from other variables. It preserves relationships better than the mean, but a single predicted value still pretends you know more than you do.
Multiple imputation creates several complete datasets, each with slightly different plausible values for the blanks, runs your analysis on each one, and pools the results. Because it accounts for the uncertainty in the filled-in values, it is now the recommended approach in many methods textbooks when data are MAR and missingness is more than trivial. It takes more steps, but modern software has made it accessible, and it is well worth learning for a thesis.
A simple decision rule many supervisors suggest is this: if less than about 5 percent is missing and the pattern looks random, listwise deletion with a clear report is usually acceptable. If more is missing, or the pattern is systematic, move to multiple imputation and run a sensitivity check. Treat any fixed cut-off as a guide, not a law — your supervisor, journal, or examiner may expect a specific approach.
Step 4: Run a Sensitivity Check
A sensitivity check asks a plain question: would my conclusion change if I handled the missing data differently? Run your key analysis twice, for example once with listwise deletion and once with multiple imputation, and compare the main results. If the conclusion is the same, say so: that strengthens your paper. If it changes, report both results honestly and discuss why. Reviewers respect that far more than a single analysis that hides the uncertainty.
Step 5: Report What You Did, Clearly
Whatever you choose, report it in your methods section. A good missing-data paragraph covers four points: how much was missing, for which variables; what pattern you found; which method you used and why; and whether a sensitivity check changed anything. Add the final analytic sample size next to the original sample size, so the reader can see exactly how many cases were lost or retained.
Worked Example: A Student Survey With Blanks
Imagine an illustrative example, not real data: you survey 120 undergraduate students about study habits and exam stress. When you check the file, 12 students skipped the weekly study-hours question, 18 skipped the household-income question, and 6 left the final stress scale incomplete. Overall missingness is modest, but the income question is the worst, and students who skipped income also tended to be older students in part-time work.
Step 1 shows you the counts. Step 2 shows a pattern linked to an observed variable, age and work status, which points toward MAR rather than MCAR. Deleting all 36 affected cases listwise would leave you with far fewer than 120 students and would disproportionately remove older, working students. Mean substitution for income would flatten an important variable. A sensible choice here is multiple imputation using age, work status, and study hours to predict plausible income values, followed by a sensitivity check comparing the imputed analysis with a complete-case analysis. In your write-up, you report the original 120, the missing counts per variable, the pattern you found, and the fact that the main conclusion about study hours and stress held under both approaches. That paragraph takes only a few sentences, but it answers the examiner’s questions before they are asked.
Common Mistakes to Avoid
- Deleting cases without looking first. Deletion feels tidy, but blind deletion can quietly bias your sample and waste hard-won data.
- Treating a blank as zero. A missing answer is not a zero score, a “no,” or a neutral response. Coding it as zero invents data and distorts your results.
- Using mean substitution for large amounts of missing data. It shrinks variation, weakens correlations, and makes your estimates look more precise than they are.
- Ignoring the pattern. If one group has far more missing data, that is a finding in itself. Investigate it before you analyse anything else.
- Saying nothing at all. The most common mistake is silence. Even a brief, honest paragraph is better than leaving the reader to guess how many cases were dropped and why.
- Imputing without thinking about MNAR. If you suspect the missingness depends on the unobserved value itself, say so, consider a sensitivity analysis under different assumptions, and discuss it as a limitation. Do not let a sophisticated method hide a substantive problem.
Frequently Asked Questions
How much missing data is too much?
There is no single magic threshold, but as a rough guide, under about 5 percent per variable with a random-looking pattern is usually manageable with simple methods, while larger amounts call for multiple imputation and a sensitivity check. What matters more than the percentage is the pattern and whether your conclusions survive a different handling method.
Can I just delete incomplete cases?
Yes, if missingness is small and looks random, and you report how many cases you dropped and why. If missingness is large or systematic, deletion alone risks bias and lost power, and reviewers will likely ask why you did not consider imputation.
Is mean imputation ever acceptable?
For a very small amount of missing data, mean substitution is sometimes used for simplicity, especially in coursework. It is rarely the best choice for a thesis or publication, because it underestimates variation. If you use it, say so and keep the amount involved tiny.
What if only one item of a scale is missing?
Check the scale’s own scoring instructions first, because some scales allow prorating, which means averaging the answered items, when only a small number of items are missing. If the instructions do not allow it, do not invent a rule. Either treat the scale score as missing for that respondent or use an imputation approach and report it.
What software can I use?
SPSS includes missing-value analysis and multiple imputation routines. In R, widely used packages handle both diagnosis and multiple imputation. Even a spreadsheet audit of counts and patterns is a good start before you move to a statistical package.
Final Thought
Missing data is normal. Every dataset has gaps, and handling them carefully is part of doing good research, not a sign that your study failed. Count the gaps, look for a pattern, choose a method that fits that pattern, check whether your conclusion survives a different choice, and tell the reader what you did. Do those five things, and a blank cell stops being something to hide and becomes evidence that you handled your data with care.