Psychology dissertation regression analysis can test how an outcome varies with one or more predictors while estimating uncertainty and controlling specified covariates. The software may fit a model in seconds, but a defensible dissertation must justify the model, code variables correctly, inspect assumptions, distinguish prediction from causation, and report coefficients in language readers can understand.
This guide explains linear, logistic, and related regression models through psychology-specific examples. It covers planning, data preparation, diagnostics, interactions, model comparison, interpretation, reporting, and common mistakes without turning the dissertation into a software manual.
What a Psychology Dissertation Regression Analysis does
Regression represents a conditional relationship between an outcome and one or more predictors. In a linear model, each coefficient estimates the expected difference in a continuous outcome associated with a one-unit predictor difference, holding the other included predictors constant. In logistic regression, coefficients describe change in the log odds of a binary outcome and are often presented as odds ratios.
Regression can address several purposes:
- estimate an association, such as the relationship between sleep duration and academic stress;
- adjust an estimate for defensible covariates;
- compare the incremental contribution of a predictor block;
- test whether an association differs across levels of another variable;
- predict an outcome for new observations when prediction is genuinely the aim.
These purposes are not interchangeable. A model designed to explain a theoretical relationship may not be an accurate prediction model. An adjusted association from observational data is not automatically causal.
Start with the research question, not the software menu
Write the question and estimand before choosing commands. Identify the outcome, focal predictor, comparison or unit change, population, time point, and covariates. The psychology dissertation research-question guide can help align the claim with the design.
Consider this question: Among first-year university students, how is weekly sleep duration associated with end-of-term perceived stress after accounting for baseline stress and paid-work hours? The outcome is end-of-term stress, the focal predictor is weekly sleep duration, and the primary coefficient represents the expected stress-score difference per additional hour of sleep, conditional on the specified covariates.
A weak version would ask whether several available variables “predict mental health” without defining the outcome, timing, theory, or intended interpretation. Adding every column to a model does not create a stronger question.
Table of Contents
Choose a model that matches the outcome
The outcome’s scale and distribution are the starting points. Do not select ordinary linear regression merely because it is familiar. Likewise, do not convert a continuous score into high and low groups without a sound substantive reason, because categorisation discards information and makes results depend on an arbitrary threshold.
| Outcome | Common model | Psychology example |
|---|---|---|
| Continuous | Linear regression | Stress score |
| Binary | Binary logistic regression | Completed treatment: yes or no |
| Ordered categories | Ordinal logistic regression | Low, moderate, high engagement |
| Unordered categories | Multinomial logistic regression | Preferred coping strategy |
| Count | Poisson or negative-binomial regression | Number of help-seeking contacts |
| Repeated observations | Multilevel or mixed model | Daily mood nested within person |
The table is a starting map rather than a diagnosis. Count models differ in their distributional assumptions. Ordinal models require attention to proportional-odds assumptions. Repeated data require a model that represents dependency. Consult an appropriate statistical text and supervisor when the design exceeds the taught model.
Design the regression before collecting data
A good analysis plan lists the primary outcome, focal predictors, covariates, coding, transformations, interaction terms, missing-data strategy, diagnostic checks, influential-case rules, model-comparison plan, and reporting outputs. Separate confirmatory analyses from exploratory extensions.
Use the site’s hypothesis guide to connect directional predictions to variables and models. A hypothesis such as “social support will predict wellbeing” is incomplete unless the construct measures, direction, timing, and analysis are clear.
Plan an informative sample size
Regression sample size depends on the primary effect, number and structure of predictors, measurement quality, outcome distribution, anticipated missingness, and desired precision. A fixed rule such as ten or twenty participants per predictor is not a universal justification.
Plan around the smallest effect that would matter for the focal test, using an analysis appropriate to the intended model. Allow for exclusions and attrition without inflating the final claim. For logistic models, sparse outcomes can create unstable estimates even when the total sample appears large. The sampling guide explains why representativeness and information are not guaranteed by a large number alone.
Preregister decisions where appropriate
A preregistration can distinguish planned models, variables, exclusions, and interactions from later exploration. It does not prevent justified changes. Record deviations, explain why they occurred, and label the revised work accurately.
Prepare variables without manufacturing a result
Protect the raw data and create a reproducible working file. Check ranges, labels, duplicates, impossible values, missing codes, scale scoring, and reverse-scored items before fitting the primary model. Document every recode.
Code categorical predictors transparently
A categorical predictor is represented through indicator or contrast variables. State the reference category because each coefficient is interpreted relative to it. For a study-mode variable with in-person, hybrid, and online categories, choosing in-person as the reference yields separate hybrid-versus-in-person and online-versus-in-person estimates.
Do not enter arbitrary category numbers such as 1, 2, and 3 as if they were equally spaced unless that assumption is meaningful. Preserve categories, choose planned contrasts, or justify an ordinal representation.
Centre and scale variables for a reason
Mean-centring a continuous predictor changes the zero point, so the intercept and lower-order coefficients may become easier to interpret. Standardising can place predictors on standard-deviation units, but standardised coefficients do not solve confounding or make constructs equivalent.
Always report what transformation was used. Retain unstandardised coefficients when the original units are interpretable, such as hours of sleep or points on a validated scale.
Handle missing data explicitly
Describe how much data are missing, where missingness occurs, and why the chosen method is plausible. Complete-case analysis can reduce precision and introduce bias when complete cases differ systematically from incomplete cases. Single mean substitution usually understates uncertainty and distorts relationships.
More advanced approaches, including multiple imputation, need compatible models and transparent reporting. Do not apply them as a button-click repair. The broader data-analysis guide covers missing-data reasoning and data screening.
Specify covariates from theory and design
Covariates should enter because they address the research question, design, or plausible confounding structure. Selecting them only because their individual p-values are small creates unstable, data-dependent models. Automatically controlling every available variable can also introduce bias by adjusting for mediators, colliders, or consequences of the exposure.
Draw a simple causal or temporal diagram before modelling. Ask whether each candidate covariate precedes the focal predictor, affects the outcome, lies on the proposed pathway, or reflects sample selection. Even if the dissertation is not making a causal claim, this exercise clarifies what “holding constant” means.
Use hierarchical entry purposefully
Hierarchical regression enters blocks in a theory-driven order. A researcher might enter baseline distress first, demographic covariates second, and social support third to examine the additional variance associated with social support. The order must be justified before looking at results.
Hierarchical regression is not the same as multilevel modelling. The former concerns block entry; the latter models nested or clustered data.
Understand the linear regression assumptions
Assumptions concern the model and its errors, not simply whether each raw variable passes a normality test. UCLA’s statistical consulting resources emphasise regression diagnostics because violations can make estimates or inferences misleading.
| Issue | What to inspect | Possible response |
|---|---|---|
| Linearity | Residual and partial plots | Model a justified curve or transform |
| Unequal variance | Residual spread across fitted values | Use suitable robust inference or respecify |
| Dependent errors | Design, clustering, repeated records | Use a model representing dependency |
| Influential cases | Leverage, residuals, influence measures | Verify, explain, and run sensitivity checks |
| Collinearity | Predictor relations and variance inflation | Reconsider redundant predictors |
| Residual distribution | Q-Q plot and extreme residuals | Assess inference and model form |
Do not delete a valid observation merely because it is inconvenient. Check for data-entry errors, determine whether the case belongs to the target population, inspect its influence, and compare conclusions with and without it when justified. Report the decision and sensitivity analysis.
Do not use assumption tests mechanically
Large samples can make trivial deviations statistically detectable, while small samples give assumption tests little power. Combine design knowledge, graphical diagnostics, and model-specific reasoning. A significant normality test on a predictor does not by itself invalidate linear regression.
Check logistic and other models on their own terms
Logistic regression does not require normally distributed predictors or residuals. It does require appropriate outcome coding, independent observations when using a standard model, a plausible relationship between continuous predictors and the log odds, adequate information, and no perfect separation.
Assess calibration, discrimination only when prediction is an aim, influential observations, functional form, and sparse cells. An odds ratio can look dramatic when the outcome is rare or the estimate is imprecise, so report confidence intervals and absolute outcome frequencies.
For count, ordinal, multinomial, and multilevel models, use diagnostics and fit criteria appropriate to those models. Do not recycle an ordinary least-squares checklist.
Test interactions without confusing main effects
An interaction asks whether the association between one predictor and the outcome differs across another predictor. For example, does the association between academic workload and stress differ by perceived social support?
Include the component terms when fitting their interaction. After an interaction, a lower-order coefficient is conditional: it describes the predictor’s association when the other interacting variable equals zero or its reference level. Centring can make that point meaningful, but it does not change the evidence for the interaction.
Interpret the interaction with predicted values or simple slopes at defensible values. A graph with confidence bands is usually more informative than a coefficient alone. Avoid describing one subgroup as significant and another as not significant unless the difference between subgroup effects has itself been tested.
Distinguish mediation from ordinary regression
Mediation proposes a pathway through which an exposure relates to an outcome. Several regressions can estimate an indirect effect, but cross-sectional associations do not establish temporal or causal mediation. The mediator should be theoretically justified and measured at an appropriate time.
Report the indirect effect with uncertainty and discuss alternative explanations. Do not claim that a mechanism is proven because one path is statistically significant. If the design measures predictor, mediator, and outcome simultaneously, use cautious associational language.
Interpret coefficients, uncertainty, and model fit
For linear regression, report the unstandardised coefficient, standard error or confidence interval, test statistic, p-value where required, and model context. A coefficient of -1.8 for sleep hours means the model estimates 1.8 fewer stress-score points per additional hour of sleep, conditional on included predictors. It does not mean sleep “causes” stress to fall.
For logistic regression, an odds ratio above 1 indicates higher odds and below 1 indicates lower odds relative to the predictor unit or reference category. Odds are not probabilities. Translate selected estimates into predicted probabilities when that helps the reader and the values are within the observed data range.
Report effect size and uncertainty
APA’s quantitative reporting standards recommend effect sizes and confidence intervals or statistical significance information. The American Statistical Association cautions that a p-value does not measure effect size, practical importance, or the probability that a hypothesis is true.
A small p-value cannot repair a weak design, and a large p-value is not proof of no relationship. Discuss estimate magnitude, interval width, measurement quality, sample information, and compatibility with theory together.
Use model fit measures carefully
R-squared describes variance accounted for in the analysed sample under a linear model. Adjusted R-squared penalises model complexity, but neither demonstrates causation or out-of-sample accuracy. Logistic pseudo-R-squared measures do not have the same interpretation as ordinary R-squared.
When comparing nested models, report why the comparison answers the question and what changed. Do not select a final model solely because one fit statistic improves slightly.
Report a regression analysis clearly
The method section should identify the outcome, predictors, coding, reference groups, transformations, covariates, interaction terms, missing-data approach, diagnostics, software and version, and planned sensitivity analyses. The results should follow the hypotheses rather than the sequence of software clicks.
| Report element | Include | Avoid |
|---|---|---|
| Sample | Analysed N and exclusions | Only recruited N |
| Model | Outcome, predictors, coding, estimator | “A regression was run” |
| Coefficient | B or odds ratio with interval | P-value alone |
| Diagnostics | Relevant checks and responses | A list of unexplained tests |
| Fit | Appropriate fit measure and comparison | Fit as proof of truth |
| Interpretation | Units, reference, conditional meaning | Causal language from association |
Use a compact coefficient table rather than pasting raw software output. Give meaningful variable labels and reference categories. Align decimal places, define abbreviations, and introduce the table in the text without repeating every cell. The results-section guide provides a complete reporting workflow.
Worked psychology example
Suppose a dissertation examines whether sleep duration is associated with end-of-term stress among first-year students. Baseline stress and paid-work hours are prespecified covariates. Social support is an exploratory moderator.
- Score the stress and support measures according to their manuals.
- Verify ranges, missing values, duplicates, and time-point identifiers.
- Describe the analysed sample and variables.
- Plot stress against sleep and inspect functional form.
- Fit the prespecified main-effects model.
- Inspect residuals, influence, variance, and collinearity.
- Fit the labelled exploratory sleep-by-support interaction.
- Graph model-predicted stress across observed sleep values at defensible support levels.
- Report estimates, intervals, p-values if required, fit, diagnostics, and sensitivity checks.
If the adjusted sleep coefficient is negative, interpret it as a conditional association in score units. If the interval is wide, acknowledge limited precision. If the interaction is uncertain, do not claim that social support definitely buffers stress.
Common regression mistakes and repairs
Selecting predictors by p-value
Repair: choose focal variables and covariates from the question, theory, design, and a documented plan.
Calling every predictor a cause
Repair: use association language unless design and assumptions support a causal interpretation.
Ignoring reference categories
Repair: state coding and interpret each categorical coefficient relative to its reference.
Deleting outliers automatically
Repair: verify values, assess influence, justify exclusions independently of results, and report sensitivity checks.
Testing many models without disclosure
Repair: distinguish the primary model from exploratory alternatives and control the interpretation of multiplicity.
Reporting beta and p only
Repair: report unstandardised units, interval estimates, model context, fit, diagnostics, and substantive meaning.
Treating software defaults as methodology
Repair: document the estimator, coding, missing-data handling, diagnostic decisions, and reasons for each model.
A practical regression workflow
- Define the question, outcome, focal predictor, and estimand.
- Choose a model family that matches the outcome and dependency structure.
- Prespecify coding, covariates, interactions, missing data, and diagnostics.
- Plan sample information for the focal effect.
- Protect raw data and create reproducible cleaning and scoring steps.
- Describe variables and inspect relationships before modelling.
- Fit the planned model and examine model-specific diagnostics.
- Investigate influential cases without outcome-driven deletion.
- Interpret coefficients in units with intervals and reference groups.
- Label deviations and exploratory analyses honestly.
- Produce a concise table, useful plot, and reproducible syntax.
- Check every claim against the design and evidence.
Psychology dissertation regression analysis checklist
- The model answers a defined research question.
- The outcome scale matches the model family.
- Predictors, covariates, and interactions have theoretical roles.
- Categorical coding and reference groups are documented.
- Transformations and scoring are reproducible.
- Missing data and analysed sample size are reported.
- Diagnostics match the fitted model.
- Influential observations are investigated, not silently removed.
- Coefficients are interpreted conditionally and in clear units.
- Effect estimates and uncertainty accompany p-values.
- Confirmatory and exploratory models are separated.
- Claims do not exceed the design.
Frequently asked questions
How many predictors can I include?
There is no universal limit. Capacity depends on sample information, outcome distribution, effect size, collinearity, missingness, model complexity, and the precision required for the focal estimate.
Do predictors need to be normally distributed?
No. Ordinary linear regression does not require each predictor to be normal. Focus on model form, residual behaviour, independence, influence, and whether inference is appropriate.
Should I standardise every predictor?
No. Standardisation may help compare scales or interpret an interaction, but original units are often clearer. Explain any standardisation and retain unstandardised estimates when meaningful.
Can I use regression with Likert-scale data?
A multi-item scale score is sometimes modelled as approximately continuous when its construction and distribution support that decision. A single ordered item may require an ordinal model. Justify the scale treatment rather than relying on a slogan.
What if an assumption is violated?
First identify the cause and relevance. A different functional form, robust inference, a model for clustering, or a different model family may be appropriate. Do not transform or delete data mechanically.
Does a significant regression coefficient prove prediction?
No. Statistical association in the analysed sample is not evidence of accurate out-of-sample prediction. Predictive claims need validation procedures and suitable performance measures.
Conclusion
Regression is most useful when the research logic remains visible. Define the estimand, match the model to the outcome, prepare variables reproducibly, choose covariates deliberately, inspect model-specific assumptions, and interpret coefficients with uncertainty. Keep causal, explanatory, and predictive claims distinct.
If you seek dissertation support, use feedback to test your model logic, coding, diagnostics, tables, and interpretation while retaining authorship and responsibility for every analytic decision. Ethical support should improve understanding, not manufacture results or conceal uncertainty.
Authoritative references
- APA Style Journal Article Reporting Standards for Quantitative Research
- American Statistical Association Statement on Statistical Significance and P-Values
- UCLA Statistical Consulting: Regression Diagnostics in SPSS
- UCLA Statistical Consulting: Logistic Regression Diagnostics
- UCLA Statistical Consulting: Introduction to Regression in R
