Psychology dissertation logistic regression analysis is appropriate when the outcome has two categories and the research question concerns adjusted associations or predicted probabilities. The method can examine outcomes such as treatment completion, diagnostic status, relapse, help-seeking, or correct versus incorrect responses. A defensible analysis requires more than selecting “Binary Logistic” in software: you must define the event, plan predictors, inspect sparse data, check functional form, interpret odds ratios correctly, and match evaluation to the study’s purpose.
This guide explains binary logistic regression for globally based psychology students. It covers model design, outcome coding, sample size, assumptions, separation, interactions, predicted probabilities, SPSS and R workflows, model performance, APA-style reporting, and common errors. It also distinguishes explanatory modelling from individual prediction so the statistical output answers the question you actually asked.
What does logistic regression estimate?
Binary logistic regression models the probability that an event occurs. Because probabilities must remain between zero and one, the model transforms probability into log odds, also called the logit. Predictors are then related linearly to that transformed outcome.
A coefficient represents the expected change in log odds for a one-unit predictor increase, holding other model variables constant. Exponentiating the coefficient gives an odds ratio. An odds ratio above one indicates higher odds of the coded event, below one indicates lower odds, and equal to one indicates no change in odds.
Odds are not probabilities
If the probability of completing therapy is 0.75, the odds are 0.75 divided by 0.25, or 3 to 1. A doubling of odds does not mean a doubling of probability. The probability change associated with an odds ratio depends on the starting probability and the values of other predictors.
For accessible interpretation, report odds ratios with confidence intervals and add predicted probabilities or marginal effects for realistic cases. Never translate an odds ratio directly into a percentage-point probability change without calculating that change from the fitted model.
Table of Contents
Logistic regression as a generalised linear model
Logistic regression is a generalised linear model with a binomial outcome distribution and usually a logit link. This separates it from ordinary linear regression, which assumes a continuous outcome and errors that behave differently. The broader psychology dissertation regression guide maps models to outcome types, while this article focuses on the binary case.
When is psychology dissertation logistic regression appropriate?
Use binary logistic regression when each independent observational unit has an outcome coded into two meaningful categories. Examples include attended versus did not attend, improved versus did not improve under a prespecified definition, and met versus did not meet diagnostic criteria.
| Outcome structure | More suitable starting model | Psychology example |
|---|---|---|
| Two categories | Binary logistic regression | Completed an online intervention: yes or no |
| Three or more unordered categories | Multinomial logistic regression | Preferred support route: peer, tutor, or clinician |
| Ordered categories | Ordinal logistic regression | Low, moderate, or high engagement |
| Repeated binary observations | Mixed-effects or GEE logistic model | Daily help-seeking recorded within participants |
| Time until an event with censoring | Survival model | Weeks until treatment dropout |
Logistic regression versus chi-square
A chi-square test can examine association between categorical variables in a contingency table. Logistic regression can include continuous and categorical predictors simultaneously, estimate adjusted associations, test interactions, and generate predicted probabilities. The psychology dissertation chi-square guide is more suitable when the main question is a simple table-based association and adjusted modelling is unnecessary.
Do not dichotomise a continuous outcome for convenience
Turning a continuous wellbeing score into “low” and “high” discards information and makes results depend on an often arbitrary threshold. It can also reduce power and conceal differences within categories. Use a binary outcome only when the categories are substantively meaningful, clinically or theoretically justified, and defined before analysis.
Define the question and estimand first
State the population, binary event, time point, focal predictor, comparison, covariates, and intended interpretation. For example: “Among first-year university students, what is the adjusted association between perceived belonging at enrolment and use of counselling services during the first semester?”
Decide whether the aim is explanation, causal estimation, or prediction. An explanatory model estimates associations that test a theory. A causal model requires design and identification assumptions beyond regression. A prediction model estimates probabilities for new individuals and needs performance evaluation and validation. Mixing these aims produces confusing predictor selection and reporting.
Specify the event category
Record exactly which value represents the event. If treatment completion is coded 1 and non-completion 0, positive coefficients correspond to higher odds of completion. Reversing the coding reverses coefficient signs and inverts odds ratios. Check the software output rather than assuming its event convention.
Plan predictors before viewing outcomes
Select focal predictors and covariates using the research question, design, prior evidence, and a plausible causal or temporal structure. Avoid entering every available variable, testing many combinations, or retaining predictors solely because their univariable p values are small.
For observational research, distinguish confounders from mediators, colliders, and consequences of exposure. The confounding variables guide explains why indiscriminate adjustment can introduce rather than remove bias.
Code categorical predictors transparently
A predictor such as study mode with in-person, hybrid, and online categories needs indicator or contrast coding. State the reference category because each reported odds ratio compares one level with that reference. Do not enter category labels 1, 2, and 3 as a single continuous predictor unless an equal-step linear trend is substantively defensible.
Keep continuous predictors continuous
A one-unit odds ratio may be hard to interpret for a measure with a large range. Rescale the predictor to a meaningful unit, such as five points on a stress scale or ten years of age. Mean-centring can make the intercept and interaction terms easier to understand, but it does not change model fit.
Do not divide a continuous predictor into “low,” “medium,” and “high” merely to avoid checking nonlinearity. Retaining the original scale preserves information. If the logit relationship is curved, use a justified transformation or flexible function such as a spline and present predicted probabilities.
Plan sample size around information, not a slogan
Logistic regression information depends on the total sample, outcome prevalence, number of predictor parameters, predictor distributions, effect sizes, and degree of separation. A dataset with 500 participants but only 15 events can be less informative for event modelling than a smaller balanced dataset.
A fixed rule of ten events per predictor is not a universal sample-size justification. Riley and colleagues’ prediction-model sample-size research shows that requirements vary with anticipated model fit, number of parameters, overfitting control, and precision of the overall risk. Their method applies specifically to prediction-model development, but the wider lesson is useful: count parameters and plan around the analysis purpose.
For a hypothesis-driven dissertation, base planning on the focal coefficient or contrast, expected event proportion, predictor distribution, alpha level, and desired power or confidence-interval precision. Account for attrition and missingness without replacing planning with a post hoc power calculation. The power analysis guide provides a transparent framework.
Prepare and screen the data
Protect the raw file and create a reproducible working dataset. Check participant identifiers, duplicates, allowed ranges, missing-value codes, scale scoring, outcome coding, category frequencies, and timing. Produce a table crossing the binary outcome with each categorical predictor.
Look for sparse cells and separation
Sparse cells can make estimates unstable. Complete separation occurs when a predictor or combination of predictors perfectly distinguishes events from non-events. Standard maximum likelihood estimates may then grow without a finite solution, producing huge coefficients, standard errors, or convergence warnings.
UCLA’s SPSS logistic regression example recommends checking empty and small cells and warns that rare outcomes can be difficult even in an apparently large dataset. Heinze and Schemper’s separation study describes penalised likelihood as a way to obtain finite estimates when separation occurs.
Do not solve separation by silently deleting predictors, combining substantively different categories, or removing observations. Verify coding first. Then reduce complexity using theory, collect more informative data if feasible, or use a suitable penalised method with supervisor or statistical guidance. Report the problem and remedy.
Handle missing data explicitly
List missingness for the outcome and every predictor. Complete-case analysis can reduce the number of events and change the analysed sample. Single mean substitution distorts distributions and understates uncertainty. If multiple imputation is appropriate, make its model compatible with the logistic analysis and include variables that help explain missingness. See the missing data guide for planning and reporting.
Check assumptions and model adequacy
Logistic regression does not require normally distributed predictors or equal residual variance. Its important requirements concern independent observations, correct outcome structure, appropriate functional form, adequate information, absence of destructive collinearity, and a model suited to the sampling design.
| Issue | Useful diagnostic | Possible response |
|---|---|---|
| Dependent observations | Review clustering and repeated records | Use mixed-effects, GEE, or survey-aware methods |
| Nonlinearity in the logit | Plot predicted relationships; compare flexible functions | Use a justified transformation or spline |
| Sparse data or separation | Outcome-by-category tables, warnings, extreme estimates | Simplify defensibly or use suitable penalisation |
| Multicollinearity | Correlations, coding review, variance inflation diagnostics | Remove redundancy based on construct meaning |
| Influential observations | Deviance residuals, leverage, change in estimates | Verify cases and report sensitivity analyses |
| Poor prediction calibration | Calibration plot, intercept, and slope | Revise or validate the model; avoid threshold-only claims |
Independence is a design question
Students nested within classes, clients nested within therapists, or repeated binary observations within a participant are not independent. Ordinary standard errors can be too small when clustering is ignored. Use the multilevel modelling guide when the hierarchy is central.
Check linearity in the logit
The model assumes each continuous predictor has the specified relationship with the log odds, not with the raw probability. A curved probability plot is expected under a linear logit. Check whether the fitted shape is plausible across the observed predictor range.
Use subject-matter knowledge and graphical diagnostics. If a spline or polynomial is fitted, retain all parameters representing that predictor when testing it. Do not select a curve merely because it yields the smallest p value.
Assess influential cases without automatic deletion
An unusual predictor pattern or outcome can have substantial influence. Confirm that the record is accurate, examine leverage and changes in coefficients, and compare conclusions with and without the case as a sensitivity analysis. Exclusion needs a rule connected to data quality or the target population, not discomfort with the result.
Build the model in a reproducible sequence
Start with the prespecified analysis plan. Fit descriptive and unadjusted models only when they answer stated purposes, then fit the primary adjusted model. Use likelihood-ratio tests or clearly justified Wald tests for multi-parameter terms. Preserve syntax, software version, coding decisions, and deviations from the plan.
Test interactions as conditional effects
An interaction asks whether the association between one predictor and the outcome varies across another predictor. For example, the relationship between academic belonging and help-seeking might differ by study mode. Include the component terms whenever their interaction is present.
Odds-ratio interactions are difficult to communicate in isolation. Report predicted probabilities or contrasts at meaningful moderator values with confidence intervals. The moderation analysis guide provides an interpretation workflow.
Run psychology dissertation logistic regression in SPSS
In SPSS, choose Analyze, Regression, Binary Logistic. Place the binary outcome in Dependent and predictors in Covariates. Use Categorical to declare categorical predictors, choose reference categories deliberately, and request confidence intervals for Exp(B). Save predicted probabilities and diagnostic measures when appropriate.
Before interpreting coefficients, verify the event coding and categorical-variable table. Review cases included, convergence, model-comparison output, coefficients, standard errors, odds ratios, and confidence intervals. UCLA’s annotated SPSS output explains these components, including log odds, odds ratios, and reference coding.
A significant omnibus test does not guarantee useful prediction or correct specification. Pseudo R-squared values are not equivalent to ordinary linear-regression R-squared. Treat classification tables cautiously because results depend on the chosen probability threshold and event prevalence.
Run psychology dissertation logistic regression in R
A standard R model is glm(event ~ predictor1 + predictor2, family = binomial(link = "logit"), data = dat). Check factor reference levels before fitting. Use exp(coef(model)) only with appropriate confidence intervals, and obtain predicted probabilities with predict(model, type = "response").
The official R glm documentation defines the generalised linear modelling interface and binomial response options. Additional packages can estimate marginal effects, flexible functions, penalised models, calibration, or validation, but every package output still depends on the model and data supplied.
Interpret coefficients and predicted probabilities
Suppose the odds ratio for a five-point increase in belonging is 1.30, with 95% CI [1.08, 1.57]. Holding included covariates constant, each five-point increase is associated with 30% higher odds of the coded event. This is an association in odds, not proof of a causal effect and not necessarily a 30% increase in probability.
For categorical predictors, state both compared groups. “Online students had 1.45 times the odds of using counselling compared with in-person students” is clearer than “study mode predicted counselling.” Include the interval and explain whether the estimate is precise enough to support a useful conclusion.
Use predicted probabilities for concrete communication
Present model-based probabilities for realistic predictor profiles or average marginal effects. State what values were used for other predictors and whether results are averaged across the sample. Avoid fictional “typical students” whose covariate combination may not exist.
Evaluate a prediction model differently from an explanatory model
If the aim is individual prediction, coefficient significance is not the main criterion. Assess calibration, which compares predicted probabilities with observed event frequencies, and discrimination, which measures how well predictions separate events from non-events. These qualities are distinct: a model may rank people well but produce poorly calibrated probabilities.
Use internal validation such as bootstrapping or cross-validation to estimate optimism. A small random train-test split can waste information and yield unstable results. External validation in new data is stronger evidence of transportability, but may exceed a dissertation’s scope.
The TRIPOD+AI reporting guidance provides a 27-item framework for studies developing, validating, or updating prediction models, including regression-based models. Use it only when prediction is truly the aim, not as decoration for a theory-testing regression.
Worked psychology dissertation example
A researcher studies whether baseline therapeutic alliance predicts completion of a six-session online anxiety programme. Completion by week eight is coded 1 and non-completion 0. The focal predictor is the alliance score after the first session. Prespecified covariates are baseline anxiety severity, age, and study centre.
The researcher verifies outcome coding, inspects event counts within centre, checks missingness, and plots the alliance relationship on the probability and logit scales. A flexible term shows no strong evidence of curvature. There is adequate overlap, no separation warning, and sensitivity checks do not show one case driving the estimate.
The primary result reports the adjusted odds ratio per five alliance points with a confidence interval. Predicted completion probabilities are shown at three meaningful alliance values while other variables are averaged over observed participants. Because the study is observational, the discussion describes an adjusted association and does not claim that increasing alliance would necessarily cause completion.
Common logistic regression mistakes
| Mistake | Why it fails | Better practice |
|---|---|---|
| Calling an odds ratio a risk ratio | Odds and probability are different, especially for common outcomes | Use accurate language and add predicted probabilities |
| Ignoring the event code | The direction of every coefficient may be reversed | State and verify the coded event |
| Using ten events per variable as proof of adequacy | Information needs vary by prevalence, parameters, and purpose | Use an analysis-specific sample-size justification |
| Keeping only significant predictors | The model becomes unstable and outcome-driven | Prespecify terms from theory and design |
| Ignoring separation or convergence warnings | Maximum likelihood estimates may be effectively infinite | Diagnose sparse structure and use a defensible remedy |
| Reporting accuracy alone | A dominant outcome category can make trivial predictions look good | Match performance metrics to purpose and include calibration |
How to report logistic regression
In Methods, report the binary outcome and event code, predictor definitions, units, reference categories, covariate rationale, sample-size basis, missing-data method, functional-form checks, interactions, estimation method, diagnostics, software, and any validation strategy.
In Results, give total participants, events and non-events, missingness, descriptive distributions, coefficients or odds ratios with 95% confidence intervals, tests for multi-parameter terms, and relevant diagnostics. For prediction, add calibration, discrimination, optimism-adjusted performance, and the full model specification.
In Discussion, interpret magnitude and uncertainty in context. Explain limits arising from design, selection, measurement, missingness, sparse data, overfitting, and transportability. Do not convert an adjusted association into causation or present sample-specific performance as guaranteed future accuracy.
Frequently asked questions
How many participants do I need for logistic regression?
There is no universal number. Planning depends on event frequency, predictor parameters, anticipated effects or model fit, desired precision, missingness, and whether the goal is hypothesis testing or prediction.
Do predictors need to be normally distributed?
No. Logistic regression does not require normal predictors. Continuous predictors do need an appropriate functional relationship with the log odds, and the dataset needs sufficient information across predictor patterns.
What does Exp(B) mean in SPSS?
Exp(B) is the exponentiated coefficient, usually interpreted as an odds ratio. Its meaning depends on the predictor unit, event coding, reference category, other variables in the model, and any interactions.
Can I use logistic regression with a rare outcome?
Possibly, but few events increase instability, separation risk, overfitting, and imprecision. Simplify the model based on theory, plan sample size around events and parameters, and consider suitable penalised estimation with expert guidance.
Is the Hosmer-Lemeshow test enough?
No. A single goodness-of-fit p value is sensitive to grouping and sample size. Combine graphical calibration, residual and influence checks, functional-form assessment, and purpose-specific performance measures.
Should I report odds ratios or predicted probabilities?
Usually both. Odds ratios describe multiplicative changes in odds. Predicted probabilities make implications concrete. Each should include uncertainty and clearly stated comparison values.
Can logistic regression prove that a predictor causes the outcome?
No. Causal interpretation requires an appropriate design, temporal ordering, measured confounding control, correct model assumptions, and a defensible causal estimand. Regression adjustment alone is insufficient.
Conclusion
A strong psychology dissertation logistic regression analysis begins with a meaningful binary event and a clearly defined purpose. It preserves informative predictors, plans sample size around the outcome and parameters, checks sparse data and functional form, reports odds ratios without confusing them with probability, and evaluates prediction only when prediction is the aim.
If you need support, use ethical tutoring that helps you understand your own data, verify coding, interpret diagnostics, and explain the model. You should retain responsibility for the analysis, decisions, writing, and conclusions. Good assistance develops your statistical judgement rather than replacing it.
