Psychology dissertation missing data decisions can change sample descriptions, effect estimates, uncertainty, and the conclusions you are able to defend. Missing values are not merely empty cells. They record a break between the study you planned and the information you observed, and that break may be related to participants, measures, procedures, or outcomes.
This guide explains how to prevent, diagnose, analyse, and report incomplete quantitative psychology data. It covers item nonresponse, attrition, skipped-by-design questions, complete-case analysis, maximum likelihood, multiple imputation, sensitivity analysis, SPSS and R workflows, and worked dissertation examples. The aim is a transparent decision process, not a universal software recipe.
What counts as a Psychology Dissertation missing data?
A value is missing when the intended observation is unavailable for analysis. That may happen because a participant skipped an item, withdrew before follow-up, encountered a technical failure, was not asked a routed question, or supplied a response that could not be scored. These situations look similar in a spreadsheet but do not have the same meaning.
Start by separating true absence from valid values. Zero may be a legitimate score. “Prefer not to answer” is a meaningful response category that should not be coded as zero. A scale item omitted through survey routing differs from an item displayed but left unanswered. Create explicit missing-value codes in the data dictionary, then convert them correctly for the analysis software.
Do not overwrite the raw file. Preserve the original export, store a working copy, and document every recode. Our psychology dissertation data collection guide explains how to build these controls before analysis.
| Situation | Meaning | Initial action |
|---|---|---|
| Skipped item | The question was shown but no answer was recorded | Check burden, sensitivity, position, and participant characteristics |
| Skipped by design | Routing made the question inapplicable | Keep a distinct structural-missing code and respect the survey logic |
| Participant attrition | A later wave or endpoint was not completed | Compare retained and lost participants using available baseline information |
| Technical loss | A device, platform, upload, or recording failed | Document the incident and whether failure relates to condition or setting |
| Invalid response | A recorded value cannot be used under a prespecified rule | Apply the rule consistently and report the resulting missingness |
| Scale-level incompleteness | One or more items needed for a score are absent | Follow the instrument’s scoring guidance before considering imputation |
Table of Contents
Prevent missingness before choosing a method
The strongest missing-data strategy begins during design. The National Academies recommends careful design and conduct to limit the amount and impact of missing data, followed by analysis that makes assumptions explicit. Statistical methods cannot recover information that the study never collected without relying on assumptions.
Reduce avoidable burden. Keep measures aligned with the research questions, test the survey on common devices, check routing, use accessible wording, and make the participant journey realistic. For longitudinal work, collect reliable contact details with consent, schedule reminders proportionately, and explain why follow-up matters. Avoid coercive persistence.
Do not make every online item compulsory automatically. Forced answers can create distress, careless responses, or false data, especially for sensitive mental-health, trauma, sexuality, or demographic questions. A carefully placed “prefer not to answer” option may protect autonomy while revealing a type of nonresponse that should be described.
A psychology dissertation pilot study can reveal confusing items, broken validation, excessive completion time, device problems, and unclear withdrawal processes. Use the pilot to improve feasibility rather than to promise that no data will be missing.
Map the amount and pattern of missing data
Before selecting an analysis, produce a variable-level table showing observed and missing counts and percentages. Then examine case-level patterns. A dataset with 5% missing entries scattered across many people differs from one in which 5% of participants have no outcome data. The same overall percentage can conceal very different analytical problems.
Plot or tabulate missingness by study wave, condition, recruitment source, site, device, measure, and relevant observed participant characteristics. In a longitudinal anxiety study, for example, ask whether follow-up completion differs by baseline anxiety, treatment condition, age, or earlier engagement. These comparisons do not prove the missingness mechanism, but they help you formulate plausible assumptions.
Check whether the apparent absence is created by coding. Impossible dates, strings in numeric columns, special codes such as 999, and reverse-scored items can be misread by software. The SPSS psychology dissertation workflow explains system-missing and user-defined values, data dictionaries, and reproducible syntax.
Describe participant flow
Keep recruitment, consent, eligibility, commencement, completion, exclusion, and analysis as separate counts. If 300 people open a survey, 240 consent, 220 meet eligibility rules, 205 reach the final page, and 196 enter the primary model, report that flow. Do not label all 104 differences “dropout.”
For repeated measures, show attendance or response at every wave. A table may list the number with observed outcome data at baseline, post-test, and follow-up by study group. Also state whether your analysis uses participants with partial repeated data.
Understand MCAR, MAR, and MNAR carefully
Missing-data mechanisms describe how the probability of missingness relates to observed and unobserved information. They are assumptions about a data-generating process, not labels that software can establish with certainty.
Missing completely at random
Data are missing completely at random, or MCAR, when the probability of absence does not depend on observed or missing values relevant to the analysis. A random equipment failure may be a plausible example. Under MCAR, incomplete cases are effectively a random subset, although precision is still lost.
Missing at random
Data are missing at random, or MAR, when differences in missingness can be explained by observed information included in the analysis or imputation model. Suppose follow-up anxiety is more often missing among participants with high baseline anxiety, and baseline anxiety is recorded. Conditional on that information and other relevant variables, MAR may be a defensible working assumption.
Missing not at random
Data are missing not at random, or MNAR, when missingness still depends on the unseen value after conditioning on observed data. Participants experiencing the greatest deterioration might be least likely to complete a follow-up, even among people with the same recorded baseline characteristics. Observed data alone cannot generally distinguish MAR from MNAR, so substantive knowledge and sensitivity analysis matter.
| Mechanism | Plain-language question | Implication |
|---|---|---|
| MCAR | Is absence unrelated to relevant observed and unobserved values? | Deletion may remain unbiased for some targets but loses information |
| MAR | Can observed variables explain relevant differences in absence? | Likelihood or imputation methods may be defensible if models are well specified |
| MNAR | Could absence depend on the unseen value even after using observed information? | Primary assumptions need scenario-based sensitivity analysis |
A formal MCAR test is not a complete decision rule. Failure to reject does not prove MCAR, especially in a modest dissertation sample. Equally, a visible association between missingness and one observed variable does not reveal the full mechanism. Explain the study process that could create the pattern.
Choose an analysis that matches the question
The correct response depends on the estimand, design, variables missing, proportion and pattern of absence, plausible mechanism, sample size, and software. Do not choose a method from the percentage alone. Research shows that the proportion missing is not sufficient to decide between complete-case analysis and multiple imputation.
Available-case summaries
Descriptive statistics are often calculated from the observations available for each variable. Report the denominator so readers can see when means, correlations, or percentages rely on different subsets. Pairwise deletion may produce a correlation matrix assembled from inconsistent samples and may not be positive definite, so use it cautiously.
Complete-case analysis
Complete-case analysis includes only cases observed on every variable required for that model. It is simple and transparent but can reduce precision sharply. It may also bias estimates when complete cases differ systematically in ways relevant to the target relationship.
Complete-case analysis is not automatically wrong, and multiple imputation is not automatically superior. The validity of complete cases depends on the analysis and missingness structure. If you use it, report how many cases remain, compare relevant observed characteristics, justify the assumptions, and consider a sensitivity analysis.
Scale scoring rules
Many psychological measures provide rules for scoring a scale when a limited number of items are missing. Follow the official manual or validation source. Person-mean substitution permitted by a scoring manual is not the same as replacing missing variables with the full-sample mean. Document the minimum items required, prorating rule, and number of affected participants.
Maximum-likelihood approaches
Full-information maximum likelihood estimates model parameters using all available observations under specified assumptions, often MAR. It is common in structural equation and longitudinal models. Mixed-effects models may use incomplete repeated outcomes without deleting every participant, provided the model and assumptions are appropriate.
Maximum likelihood does not fill cells in a completed spreadsheet. It estimates the model directly from available information. This can be efficient, but it is not assumption-free and may require specialist software or supervision.
Multiple imputation
Multiple imputation replaces each missing value with multiple plausible draws from an imputation model, analyses each completed dataset, and combines estimates and standard errors. The multiplicity represents uncertainty about unseen values. It is not acceptable to impute once and treat predictions as observed facts.
The imputation model should include variables in the final analysis, variables related to missingness, and useful predictors of incomplete values. Match the imputation method to variable type. Binary, ordinal, nominal, count, and continuous variables should not all be treated as normally distributed without justification.
Preserve important features such as interactions, non-linear terms, clustering, and repeated structure. An imputation model that conflicts with the analysis model can introduce bias. With a small or highly sparse dataset, complex imputation models may be unstable. Seek statistical advice instead of accepting default settings blindly.
Methods to avoid as automatic fixes
Replacing every missing value with the sample mean usually understates variability and distorts correlations. Last observation carried forward assumes that an earlier score remains valid at later occasions, which is often implausible in psychological change studies. Creating a “missing” category for a continuous predictor changes the variable’s meaning. These techniques need a rare, explicit justification, not convenience.
Build a defensible multiple-imputation workflow
- Define the primary analysis and target parameter before imputation.
- Audit codes, ranges, missing counts, patterns, and participant flow.
- State why MAR is a plausible primary assumption and what remains uncertain.
- Select analysis variables, auxiliary predictors, and variables related to missingness.
- Choose imputation models suited to each variable’s scale and constraints.
- Generate multiple datasets with a recorded random seed and enough imputations for stable results.
- Check convergence and compare observed with imputed distributions.
- Fit the identical substantive model in each dataset.
- Pool estimates and standard errors using the software’s established rules.
- Compare with prespecified sensitivity analyses and report the full implementation.
The Flexible Imputation of Missing Data online text explains mechanisms, model building, diagnostics, multilevel data, and sensitivity analysis. The mice package documentation describes chained-equation implementation in R. SPSS also provides multiple-imputation procedures, but menu availability does not establish that a selected model is appropriate.
Check the imputation, not only the final p value
Inspect trace or convergence information where applicable. Compare distributions and ranges of observed and imputed values. Look for impossible ages, scale scores outside permitted bounds, implausible category combinations, or excessive heaping. Differences between observed and imputed distributions can be expected under MAR, so the purpose is not to force identical shapes. Ask whether differences make sense under the model and study process.
Record the number of imputations, iterations, seed, software version, methods by variable, predictor matrix or equivalent specification, constraints, included auxiliary variables, diagnostics, and pooling procedure. Syntax or code belongs in an appendix or repository if permitted.
Plan sensitivity analyses for uncertain assumptions
A primary MAR analysis cannot demonstrate that MNAR is harmless. Sensitivity analysis asks how conclusions change under credible alternative assumptions. Options may include complete-case comparison, different imputation specifications, delta adjustment, pattern-mixture scenarios, best-case and worst-case bounds, or models that reflect plausible dropout processes.
Choose scenarios using psychological and practical knowledge. In a therapy study, you might explore whether missing follow-up scores are plausibly worse than MAR imputations by a meaningful amount. Do not search many scenarios and report only the one preserving significance. Show the range of conclusions and identify the point at which interpretation changes.
Preregister the primary method and major sensitivity analyses when possible. Our psychology dissertation preregistration guide explains how to document contingencies without pretending all problems are predictable.
Worked psychology dissertation examples
Example 1: incomplete questionnaire items
A cross-sectional study examines whether sleep quality predicts academic stress after adjustment for age and workload. Some participants omit one or two stress-scale items. The student first checks the scale manual. It permits prorating when a stated minimum number of items is completed. The student applies that rule consistently, records how many scores were prorated, and leaves scores missing when the minimum is not met.
For covariate missingness, the student maps patterns and compares observed characteristics. The primary regression uses a justified multiple-imputation model including the outcome, predictors, auxiliary indicators of survey engagement, and all model terms. A complete-case model is reported as a sensitivity comparison. The results section gives denominators, implementation details, pooled estimates, and any material differences.
Example 2: longitudinal attrition
An experimental dissertation measures mood at baseline, post-intervention, and four-week follow-up. Follow-up is more often absent among people with higher baseline distress. Deleting everyone without all three waves would discard their earlier observations and may create a selectively retained sample.
The student uses a repeated-measures model capable of using partial outcome histories under a documented MAR assumption, including baseline variables related to attrition. They compare groups at each flow stage and perform a sensitivity scenario in which missing follow-up mood is shifted in an unfavourable direction. Interpretation focuses on effect estimates and intervals across scenarios, not simply whether each p value remains below a threshold.
Common mistakes and repairs
| Mistake | Problem | Repair |
|---|---|---|
| Reporting only an overall missing percentage | It hides variables, waves, cases, and patterns | Report variable-level and participant-level missingness with flow |
| Calling data MAR because a test was non-significant | MAR versus MNAR cannot be established from observed data alone | Use process knowledge, observed predictors, and sensitivity analysis |
| Deleting incomplete cases by default | Information and representativeness may be lost | Justify complete cases against the estimand and mechanism |
| Using mean substitution | Uncertainty and relationships are distorted | Use an appropriate scoring rule, likelihood method, or multiple imputation |
| Imputing the outcome without the analysis predictors | The imputation and analysis models are incompatible | Include the substantive-model variables and justified auxiliary information |
| Reporting “multiple imputation was used” | The analysis cannot be reproduced or evaluated | Report software, models, variables, imputations, diagnostics, and pooling |
| Hiding sensitivity results that differ | Readers cannot judge assumption dependence | Report all prespecified scenarios and explain material changes |
How to report missing data
The APA quantitative reporting standards ask researchers to report the amount of missing data and how it was handled analytically. In the method section, state preventive procedures, coding rules, planned assumptions, primary method, and sensitivity analyses. In the results, report actual amounts and patterns, achieved analysis samples, diagnostics, implementation, and deviations.
A clear report answers these questions:
- Which variables, items, and waves contain missing values?
- How many participants and observations are affected?
- What reasons for absence are known?
- Which observed characteristics predict missingness?
- What mechanism is assumed, and why is it plausible?
- How were scale scores handled?
- Which cases entered each analysis?
- What software, settings, models, and pooling rules were used?
- Which sensitivity analyses were performed?
- Did conclusions change across reasonable assumptions?
Use the psychology dissertation results-section guide to integrate participant flow, descriptive tables, estimates, and uncertainty. The broader psychology dissertation data-analysis guide can help align the missing-data method with the primary model.
Frequently asked questions
How much missing data is acceptable?
There is no universal safe percentage. Impact depends on which variables are missing, the pattern and mechanism, the analysis, and the information available to address missingness. Even a small amount can matter if it concentrates in the primary outcome or one study group.
Can I remove participants with missing values?
Sometimes, but not by default. State the complete-case assumptions, quantify the loss, and consider whether retained participants differ in relevant observed ways. Compare an appropriate alternative when the primary conclusion may depend on deletion.
Is multiple imputation always the best method?
No. Its suitability depends on the estimand, missingness, model, sample, and available predictors. Maximum likelihood, a valid scale-scoring rule, complete cases, or another design-specific method may be preferable. Complex cases need statistical supervision.
Should I impute individual scale items or the total score?
Follow the measure’s official scoring guidance and your analysis plan. Item-level imputation can preserve information but requires an appropriate model and enough data. Score-level approaches may be simpler but can discard item information. Report the chosen level and rationale.
Can a test prove my data are MCAR?
No single test proves the mechanism. Such tests have limited power and evaluate restricted conditions. Combine observed pattern checks with knowledge of recruitment, survey routing, attrition, technical failures, and participant behaviour.
What if sensitivity analyses disagree?
Report the disagreement. It shows that the conclusion depends on unverified assumptions. Describe which scenarios are most plausible, reduce the strength of claims, and recommend better follow-up or data collection rather than selecting the preferred result.
Conclusion
Psychology dissertation missing data should be treated as a design, measurement, and inference problem. Preserve the raw evidence, distinguish structural absence from nonresponse, map patterns, state the mechanism you assume, choose a method aligned with the primary model, and test credible alternatives. Transparent uncertainty is stronger than an automatic deletion or imputation rule.
If you need independent support, our psychology dissertation service can help you audit missingness, document an analysis plan, and explain your own results clearly. Ethical support should preserve your data and reasoning. It should never invent observations, conceal attrition, select a method to manufacture significance, or replace your supervisor’s requirements.
Authoritative references
- APA Style: Quantitative research reporting standards
- Jakobsen and colleagues: Practical multiple-imputation guidance
- National Academies: Prevention and treatment of missing data
- Van Buuren: Flexible Imputation of Missing Data
- Hughes and colleagues: Accounting for missing data in statistical analyses
- UCLA Statistical Methods and Data Analytics: Missing data in SPSS
- CRAN: mice multiple-imputation package
