How to Run Multinomial Logistic Regression in SPSS for Nursing Research

Multinomial logistic regression in SPSS for nursing
research is appropriate when your outcome has three or more
categories that have no natural order. A nursing researcher might use it
to examine whether age, frailty and comorbidity predict discharge to
home, supported home care or inpatient rehabilitation. This guide
explains how to identify the correct model, prepare nursing data, run
the procedure in SPSS, interpret the principal output and report the
findings without confusing odds, probability and causation.

Statistical integrity note: The worked nursing
example is hypothetical and is included to demonstrate the analytical
process. Do not copy its illustrative results into a dissertation. Your
reported figures must come from your own authorised dataset and SPSS
output.

Key Takeaways

• Use multinomial logistic regression when the dependent variable has
  at least three mutually exclusive, unordered categories.
• Do not call the procedure multiple linear regression: the two models
  answer different questions and require different outcome variables.
• Place categorical predictors in Factor(s) and
  continuous predictors in Covariate(s).
• Select and report the outcome reference category deliberately
  because every comparison is interpreted against it.
• Review case processing, model fitting, goodness of fit, pseudo-R²,
  likelihood-ratio tests, classification and parameter estimates
  together.
• Report adjusted odds ratios with 95% confidence intervals, not
  p-values alone.
• Interpret adjusted associations cautiously; regression does not
  establish that a predictor caused the nursing outcome.

Watch the SPSS Procedure

The short tutorial below demonstrates the SPSS interface and the
original programme-choice example. However, an outcome with vocational,
general and academic categories requires multinomial logistic
regression, not multiple linear regression. Therefore, use the
corrected terminology and interpretation provided in this article.

What Is Multinomial
Logistic Regression?

Multinomial logistic regression estimates how a set of predictors
relates to membership of an outcome category when the outcome contains
more than two categories. IBM describes the procedure as useful for
classifying subjects from a set of predictor values (IBM, 2026a).

Nursing Outcomes That Fit
the Model

The outcome categories must be mutually exclusive: each participant
or observation belongs to one category for the analysed outcome. Typical
nursing and healthcare examples include:

• discharge destination: home, supported home care or inpatient
  rehabilitation;
• preferred pain-management approach: pharmacological, physical or
  combined;
• principal reason for a missed appointment: transport, caring
  responsibility, work commitment or another reason;
• chosen nursing career pathway: clinical practice, education,
  management or research;
• primary help-seeking route: general practice, emergency department,
  community service or no formal service.

However, the categories should not represent a clear progression. If
the outcome is ordered—for example, mild, moderate and severe symptom
categories—ordinal logistic regression may be more appropriate. By
contrast, if there are only two outcome categories, such as readmitted
versus not readmitted, consider binary logistic regression. Finally, a
continuous outcome, such as length of stay measured in days, normally
requires a different regression model.

Multinomial
Logistic Regression Versus Multiple Regression

Unfortunately, the similarity of the names causes avoidable errors in
nursing dissertations.

Feature	Multiple linear regression	Multinomial logistic regression
Outcome	Continuous	Three or more unordered categories
Example	Pain score	Discharge destination
Main estimate	Regression coefficient	Log-odds coefficient and odds ratio
SPSS route	Analyze > Regression > Linear	Analyze > Regression > Multinomial Logistic
Typical model summary	R² and adjusted R²	Likelihood-ratio tests and pseudo-R²

Importantly, the number of predictors does not determine which model
is correct. Instead, the outcome’s measurement level is the first
decisive consideration. For a broader comparison of linear and logistic
approaches, read the existing nursing
regression analysis in SPSS guide.

Worked Nursing Research
Question

Assume that a nursing researcher wants to investigate factors
associated with discharge destination among older adult inpatients. The
hypothetical dependent variable is:

• 1 = home without formal support;
• 2 = home with community support;
• 3 = inpatient rehabilitation.

The predictors are:

• age in years: continuous covariate;
• comorbidity score: continuous covariate;
• frailty category: categorical factor;
• living arrangement before admission: categorical factor.

The research question is:

To what extent are age, comorbidity, frailty and pre-admission living
arrangement associated with discharge destination among older adult
inpatients?

An appropriate null hypothesis is that the included predictors do not
improve prediction of discharge destination compared with an
intercept-only model. Predictor-specific hypotheses can then examine
whether each variable contributes to the model after adjustment for the
others.

Therefore, this example uses association language deliberately.
Unless the design and identification strategy justify a causal
interpretation, the model should not be described as proving that
frailty or living arrangement causes a particular discharge
destination.

Check Whether
the Model Fits the Nursing Question

SPSS will produce output even when the chosen model is unsuitable.
Therefore, complete the following checks before opening the regression
dialogue box. If you are still choosing between regression and another
procedure, consult the statistical
tests for nursing dissertations guide.

1. Confirm That the
Outcome Is Nominal

In this example, the discharge categories represent different
destinations rather than low-to-high levels of the same construct.
Therefore, they can be treated as nominal. Coding rehabilitation as 3
does not mean it is mathematically “higher” than supported home
care.

2. Confirm
That Categories Are Mutually Exclusive

Similarly, each case must contribute one outcome category to the
analysis. If a patient receives community support and later enters
rehabilitation, define the outcome and observation period clearly. Do
not enter the same person in two categories without an appropriate
repeated-measures structure.

3. Review Independence

Moreover, ordinary multinomial logistic regression assumes that
observations are independent. Multiple admissions by the same patient,
patients nested within wards, or measurements collected repeatedly over
time may violate this assumption. Consequently, a multilevel, clustered
or repeated-measures approach may be required.

4. Check Category Sizes
and Sparse Data

Next, run frequencies for the dependent variable and categorical
predictors. Very small outcome groups or empty combinations of
predictors and outcomes can produce unstable estimates. They may also
create extremely wide confidence intervals or convergence problems.
Nevertheless, do not combine clinically different categories solely to
obtain a significant result. Any defensible combination should follow
the research question, clinical meaning and analysis plan.

Evaluate
Advanced Multinomial Model Assumptions

5. Examine Multicollinearity

In addition, highly overlapping predictors make their separate
adjusted contributions difficult to estimate. Review conceptual overlap
and suitable collinearity diagnostics before fitting the model. For
example, a frailty measure and a functional-dependence score may
represent closely related information.

6.
Assess Linearity in the Logit for Continuous Predictors

Furthermore, the model assumes an appropriate linear relationship
between each continuous predictor and the logit for each outcome
comparison unless a different functional form is specified. A continuous
variable is not suitable merely because SPSS labels it “scale.”
Therefore, consider whether transformations, nonlinear terms or
clinically justified categorisation are needed. Avoid data-driven
changes made only to obtain significance.

7. Consider
Independence of Irrelevant Alternatives

Finally, the standard multinomial logit model relies on an
independence-of-irrelevant-alternatives assumption. In practical terms,
the relative odds between two outcome categories should not depend
inappropriately on the presence of another alternative. Closely
substitutable categories can make this assumption questionable.
Therefore, address it through substantive reasoning and, where
necessary, specialist diagnostic or sensitivity analysis rather than
citing it without evaluation.

Prepare the Nursing Dataset
in SPSS

First, keep an untouched copy of the original data. In the working
file, each row should normally represent one patient or other
independent observation. Each column should represent one variable. For
a complete screening workflow, use the guide on cleaning
nursing research data in SPSS.

In Variable View:

1. Give variables short, meaningful names, such as
   discharge, age, comorbidity,
   frailty and living.
2. Add full variable labels.
3. Define value labels for every categorical code.
4. Specify missing values consistently.
5. Set the outcome and categorical predictors as nominal.
6. Set genuinely continuous predictors as scale.

Next, use Analyze > Descriptive Statistics >
Frequencies to review categorical variables. Use suitable
descriptive procedures for continuous variables. These checks can reveal
impossible values, unexpected missingness and influential observations.
The goal is not to delete inconvenient cases. Instead, identify data
problems and document every justified correction or exclusion.

How to Run
Multinomial Logistic Regression in SPSS

Once the question, variables and assumptions are defensible, follow
these steps.

Step 1: Open the Procedure

Select:

Analyze > Regression > Multinomial
Logistic

IBM identifies this as the menu route for the multinomial logistic
regression procedure (IBM, 2026b).

Step 2: Assign the Outcome

Move discharge into the Dependent box.
Confirm that its coding matches the codebook and that all categories
appearing in the dataset are legitimate.

Step 3: Assign Factors and
Covariates

Move categorical predictors such as frailty and
living into Factor(s). Move continuous
predictors such as age and comorbidity into
Covariate(s).

This distinction affects how SPSS represents the predictors. A factor
receives category comparisons; a covariate is normally represented
through a coefficient for a one-unit increase, subject to the model
specification.

Step 4: Choose the
Outcome Reference Category

The reference outcome is the category against which the other
outcomes are compared. SPSS can use the first or last category,
depending on the chosen setting and coding. IBM also provides guidance
on changing the default reference category (IBM, 2026c).

For example, “home without formal support” may be selected as the
reference because the researcher wants to compare supported home care
and rehabilitation with routine discharge home. The choice should follow
the research question and produce clinically interpretable
comparisons—not whichever category gives the most attractive
p-value.

Consequently, record the reference category in the methods and
identify it again when presenting results.

Configure the
Multinomial Model and Output

Step 5: Specify the Model

Initially, include each predictor once in a main-effects model. Add
interactions only when the question or analysis plan justifies them and
the dataset contains enough information. Specifically, an interaction
asks whether the association between one predictor and the outcome
differs according to another predictor. Therefore, do not add
interactions automatically.

Step 6: Request Useful
Statistics

Open Statistics and request the outputs needed for
evaluation, which may include:

• case-processing information;
• model-fitting information;
• goodness-of-fit statistics;
• pseudo-R² measures;
• likelihood-ratio tests;
• parameter estimates and confidence intervals;
• classification results;
• observed and predicted probabilities where relevant.

However, do not select every option merely to make the appendix
longer. Request outputs that answer the research question or check the
model.

Step 7: Save
Predicted Values Only When Needed

Where relevant, predicted category membership or probabilities can
support diagnostics, classification assessment or later analysis.
Nevertheless, save them only when they have a defined purpose. IBM
explains how NOMREG parameter estimates relate to predicted category
probabilities (IBM, 2026d).

Step 8: Run and Preserve
the Analysis

Finally, select Paste when possible so SPSS
generates syntax before running the command. Saved syntax creates a
reproducible record of the outcome, reference category, predictors and
selected statistics. Then run the syntax or select OK
to generate the output.

How to Interpret the SPSS
Output

Interpretation should move from data inclusion and overall model
performance to predictor-specific findings. In other words, do not begin
by searching the output for p-values below .05.

Case Processing Summary

Check how many cases were included and excluded and how observations
are distributed across the outcome categories. A large reduction from
the original sample may indicate missing data across one or more model
variables. Report the analysed sample and explain missing-data
handling.

Model Fitting Information

This table compares the intercept-only model with the final model
containing the predictors. A statistically significant likelihood-ratio
chi-square indicates that the predictor model improves fit relative to
the intercept-only model. It does not show that every predictor is
important, that predictions are accurate or that the relationships are
causal.

Goodness-of-Fit Statistics

Pearson and deviance statistics are commonly displayed.
Non-significant results are often treated as consistent with adequate
fit, but sparse cells can weaken these tests. They should therefore be
considered alongside the data structure, convergence, residual
information and other model evidence.

Pseudo-R² Measures

SPSS may report Cox and Snell, Nagelkerke and McFadden statistics.
These are called pseudo-R² measures because they are not equivalent to
the R² from linear regression. Avoid writing that the model “explained
exactly 42% of the variance” solely because Nagelkerke pseudo-R² equals
.42. Name the statistic and interpret it cautiously as a model-summary
measure.

Interpret Predictor-Level
Findings

Likelihood-Ratio Tests

The likelihood-ratio tests assess whether removing a predictor
worsens model fit. A significant result suggests that the predictor
contributes information across the outcome comparisons after adjustment
for other included variables. This table helps answer whether a
multi-category factor contributes overall, whereas the parameter table
examines its specific contrasts.

Classification Table

The classification table compares observed and predicted categories.
Overall percentage accuracy can be misleading when one outcome is
common. A model that predicts the largest group for nearly everyone may
appear accurate while performing poorly for smaller, clinically
important categories. Compare category-specific performance and, if
classification is central, consider validation and more suitable
performance measures.

Parameter Estimates and
Odds Ratios

More specifically, SPSS estimates a separate comparison for each
non-reference outcome against the reference outcome. If “home without
formal support” is the reference, the model may produce:

• supported home care versus home without formal support;
• inpatient rehabilitation versus home without formal support.

The coefficient B is expressed in log-odds units.
Meanwhile, Exp(B) is the adjusted odds ratio. An odds
ratio above 1 indicates higher odds of the stated outcome relative to
the reference outcome as the predictor increases or relative to the
predictor reference group. Conversely, an odds ratio below 1 indicates
lower relative odds.

Always read the category labels, predictor coding and reference
groups before interpreting direction. Also report the 95% confidence
interval. A very wide interval signals substantial uncertainty even if
the p-value is below .05.

Worked Interpretation
Example

Suppose the hypothetical output shows an adjusted odds ratio of 2.10
for severe frailty when comparing inpatient rehabilitation with
discharge home, with a 95% confidence interval from 1.30 to 3.39 and p =
.002.

A defensible interpretation would be:

After adjustment for age, comorbidity and pre-admission living
arrangement, patients classified as severely frail had 2.10 times the
odds of discharge to inpatient rehabilitation rather than home without
formal support compared with the frailty reference group (95% CI
1.30–3.39, p = .002).

Do not write that severe frailty made rehabilitation “110% more
likely” without distinguishing odds from probability. Do not state that
frailty caused the destination unless the study design and analysis
justify causation.

If another comparison is non-significant, avoid declaring that no
relationship exists. For example:

The estimate did not provide clear evidence of an adjusted
association between severe frailty and supported home care rather than
discharge home (OR 1.24, 95% CI 0.78–1.98, p = .36).

This wording recognises the estimate and uncertainty rather than
treating a non-significant p-value as proof of equality.

How
to Report Multinomial Logistic Regression in a Nursing Dissertation

The methods section should state:

• why multinomial logistic regression matched the outcome;
• the outcome categories and reference category;
• which variables were factors and covariates;
• how variables and missing values were coded;
• the rationale for predictor inclusion;
• relevant assumption and diagnostic checks;
• the significance threshold and software version.

The results section should normally include:

• analysed sample and category frequencies;
• overall likelihood-ratio model test;
• a named pseudo-R² measure;
• goodness-of-fit and convergence information where relevant;
• predictor-level likelihood-ratio tests;
• adjusted odds ratios, 95% confidence intervals and exact
  p-values;
• classification information only when it meaningfully addresses the
  aim.

In addition, STROBE provides a reporting framework for observational
cohort, case-control and cross-sectional studies. It can help
researchers report the design, variables, statistical methods, results
and limitations transparently (von Elm et al., 2007). Nevertheless, it
is a reporting guide rather than evidence that a study is
methodologically valid. After reporting the statistics, use the separate
guide to discuss
SPSS results in a nursing dissertation without repeating the results
chapter.

Example Results Paragraph
Template

A multinomial logistic regression examined whether [predictors] were
associated with [outcome], using [category] as the reference outcome.
The final model fitted better than the intercept-only model, χ²([df]) =
[value], p = [value]. [Named pseudo-R²] was [value]. The
likelihood-ratio tests indicated that [predictor] contributed to the
model, χ²([df]) = [value], p = [value]. After adjustment for [other
predictors], [group or one-unit increase] was associated with
[higher/lower] odds of [comparison outcome] rather than [reference
outcome] (adjusted OR [value], 95% CI [lower–upper], p = [value]).

Replace every bracketed item with verified output. Do not report
fictitious numbers or use the template without checking the direction
and reference categories.

Common Errors
Nursing Students Should Avoid

Calling It Multiple
Linear Regression

A three-category nominal outcome requires a categorical-outcome
model. Using the broad phrase “multiple regression” can obscure which
procedure was performed and how the coefficients should be
interpreted.

Placing
Categorical Predictors in Covariates

Numeric codes such as 1, 2 and 3 do not automatically make a variable
continuous. A ward type or frailty group normally belongs in
Factor(s) unless the planned model justifies a
different representation.

Ignoring the Reference
Category

An odds ratio is meaningless without stating both the outcome
comparison and the predictor’s reference group or unit.

Treating Pseudo-R² as Linear
R²

Pseudo-R² statistics do not support the same explained-variance
interpretation as ordinary least-squares R². Name the measure and avoid
overstating it.

Reporting Only
Significant Comparisons

Selective reporting hides uncertainty and distorts the analytical
story. Present the planned model and clinically relevant comparisons,
including estimates that are not statistically significant.

Confusing Odds With
Risk or Probability

Odds ratios compare odds, not raw probability and not automatically
risk ratios. The distinction becomes especially important when outcomes
are common.

Overfitting the Model

Adding many predictors, categories and interactions to a modest
dissertation dataset can create unstable estimates. Predictor selection
should follow the research question, clinical reasoning and an
appropriate analysis plan—not repeated attempts to find
significance.

Claiming
Causation From an Observational Model

Adjustment accounts only for variables measured and represented in
the model. Residual confounding, measurement error, selection processes
and reverse direction may remain.

When Specialist SPSS
Support May Help

Multinomial models become difficult when categories are sparse,
observations are clustered, predictors overlap, the model does not
converge or the output contains unstable estimates. Project-specific
review can help align the research question, codebook, authorised
dataset, assumptions, model and results reporting.

Our nursing
dissertation data analysis and SPSS support can review an existing
analysis or help you develop a transparent analysis plan. You can share
your research question, anonymised codebook, authorised dataset, SPSS
output and supervisor feedback for a project-specific review. However,
support should never involve inventing data, changing observations to
obtain significance or selecting a model only because it produces a
preferred result.

Need help checking your multinomial model? Request a
review before writing the final results chapter so that the outcome
coding, reference categories, assumptions and reported odds ratios
remain consistent.

Frequently Asked Questions

Can
multinomial logistic regression include both categorical and continuous
predictors?

Yes. In SPSS, categorical predictors are entered as factors and
continuous predictors as covariates. Their coding and functional form
must still be justified.

What
is the difference between binary and multinomial logistic
regression?

Binary logistic regression has a two-category outcome. Multinomial
logistic regression is used for three or more unordered outcome
categories.

Can
I use multinomial logistic regression for mild, moderate and severe
symptoms?

Possibly, but those categories have a natural order. Ordinal logistic
regression may use that ordering more efficiently if its assumptions are
defensible. Do not automatically discard the order.

Which outcome
category should be the reference?

Choose a clinically meaningful category that supports the research
question, such as usual care or the most common destination. State the
reference explicitly and do not change it merely to obtain favourable
results.

Does a
significant model mean every predictor is significant?

No. The overall model can improve on an intercept-only model even
when some individual predictors or category contrasts do not provide
clear evidence of association.

Is
Nagelkerke R² the percentage of variance explained?

Not in the same sense as R² in linear regression. Report it as
Nagelkerke pseudo-R² and interpret it cautiously.

Should
raw SPSS tables be pasted into the dissertation?

Usually, selected results should be reformatted into concise academic
tables. Retain the full output in an appendix only when required by the
university or necessary for verification.

Conclusion

Multinomial logistic regression in SPSS for nursing research is
suitable when a study predicts an unordered outcome with three or more
mutually exclusive categories. A defensible analysis begins by
confirming that the model matches the nursing question, preparing and
checking the dataset, assigning factors and covariates correctly, and
choosing a meaningful reference category. Interpretation should combine
overall model evidence with adjusted odds ratios, confidence intervals
and clearly identified comparisons. Accurate reporting also requires
restraint: pseudo-R² is not ordinary R², odds are not probability, and
adjusted observational associations are not automatically causal.

References

IBM. (2026a) Multinomial logistic regression. IBM SPSS
Statistics Documentation. Available at: https://www.ibm.com/docs/en/spss-statistics/32.0.0?topic=regression-multinomial-logistic
(Accessed: 29 August 2026).

IBM. (2026b) Multinomial logistic regression statistics. IBM
SPSS Statistics Documentation. Available at: https://www.ibm.com/docs/en/spss-statistics/31.0.0?topic=regression-multinomial-logistic-statistics
(Accessed: 29 August 2026).

IBM. (2026c) Reference category in multinomial logistic
regression. IBM Support. Available at: https://www.ibm.com/support/pages/reference-category-multinomial-logistic-regression
(Accessed: 29 August 2026).

IBM. (2026d) Compute predicted probabilities from multinomial
logistic regression for new cases or outside SPSS. IBM Support.
Available at: https://www.ibm.com/support/pages/compute-predicted-probabilities-multinomial-logistic-new-cases-or-outside-spss
(Accessed: 29 August 2026).

von Elm, E., Altman, D.G., Egger, M., Pocock, S.J., Gøtzsche, P.C.
and Vandenbroucke, J.P. (2007) ‘The Strengthening the Reporting of
Observational Studies in Epidemiology (STROBE) statement: guidelines for
reporting observational studies’, BMJ, 335(7624), pp. 806–808.
Available at: https://doi.org/10.1136/bmj.39335.541782.AD.

---