SPSS Survival Analysis for Nursing PhDs: What Went Wrong, and How I Fixed It

SPSS survival analysis is not something most PhD candidates learn until their committee forces the issue. I found this out again a few months back, when a Nursing PhD candidate at the University of Texas at Arlington sent me her rejected analysis chapter. This case is illustrative of a pattern I see often across nursing dissertations, and identifying details have been changed to protect confidentiality. Her research question was good. Her statistical method was wrong. This article walks through exactly what went wrong and how we fixed it, using the real variables and the real SPSS syntax we ended up running.

If a search for Cox regression in SPSS help or UT Arlington dissertation SPSS help brought you here, you are almost certainly dealing with the exact same mismatch she was.

The Problem: A Rejected Dissertation Chapter

Her study looked at time to hospital readmission among heart failure patients, using electronic health record data from a large hospital system. She had a binary outcome: readmitted within 30 days, yes or no. She ran a logistic regression on it in SPSS and submitted her chapter.

Her committee rejected it. One member said she was throwing away information by ignoring when patients were readmitted. Another pointed out that patients never readmitted during the study window were being treated the same as everyone else, which was not correct. She had never heard the term survival analysis before that meeting.

This is a more common mistake than people admit. Health researchers are trained to think in outcomes, not in time. This is exactly the kind of time-to-event analysis SPSS problem her data represented, and SPSS will happily run a logistic regression on it without ever warning you that you have picked the wrong tool.

Why Logistic Regression Was the Wrong Choice

I want to be blunt about this, because I see it in almost every third dataset that lands on my desk from health, nursing, and clinical research students. Logistic regression is not wrong because it is a bad technique. It is wrong here because it answers a different question than the one being asked.

Lost Information

Her outcome was time to readmission, measured in days from discharge. Some patients were readmitted at day 5, some at day 22, some never during the 90 day follow-up. Collapsing all of that into a single yes or no variable throws away the timing entirely, and the timing was the whole point of her research question.

Censoring Ignored

Patients who were never readmitted, or who dropped out of the study before the 90 day mark, were treated identically to those who were readmitted on day 89. That is not a small error. It biases every coefficient in the model because the model assumes complete information for every case.

Wrong Model for the Question

Logistic regression assumes a fixed binary outcome. It has no concept of “not yet, but might still happen.” A time-to-event process with censoring needs a model built for exactly that, which is where Cox regression comes in.

What Is SPSS Survival Analysis?

Survival analysis is a set of statistical methods used to analyse the time until an event occurs, while properly accounting for cases where the event has not happened yet by the end of the study (censored cases). Cox proportional hazards regression, first described by Sir David Cox in his 1972 paper in the Journal of the Royal Statistical Society, is the most widely used survival analysis method because it lets you include multiple predictors without assuming a specific shape for the underlying survival curve.

It is worth knowing there is a simpler alternative here too. The Kaplan-Meier estimator plots survival probability over time for one or two groups and is excellent for a quick visual comparison. But it cannot handle multiple predictors at once the way Cox regression can, which is exactly why we needed Cox regression here. Her model had seven predictors running simultaneously, not one grouping variable.

If you are also deciding between a survival model and a simpler time-based design, this guide on cross-sectional vs longitudinal design is worth reading before you commit to a method.

Restructuring the Data for Cox Regression

Before running anything, we had to rebuild her dataset around two variables that did not exist in her original file.

1. Time — days from discharge to readmission, or to the end of follow-up if not readmitted
2. Status — coded 1 if readmitted, 0 if censored (no readmission within 90 days, or lost to follow-up)

Once these two variables existed, the rest of her dataset (age, sex, ejection fraction, comorbidity count, discharge disposition, medication adherence, follow-up care) slotted in as predictors without any further changes.

Running COXREG in SPSS: Step by Step

Here is exactly what we did, in order.

1. Opened Analyze > Survival > Cox Regression.
2. Set Time as the time variable.
3. Set Status as the event indicator and defined the event as 1.
4. Added the predictors: age, sex, ejection fraction, comorbidity count, discharge disposition, medication adherence, follow-up care.
5. In Options, requested CI for exp(B) at 95%, plus Survival function and Hazard function plots.
6. In Plots, selected Log minus log for the categorical predictors, to check the proportional hazards assumption.
7. Ran the model.

The SPSS Syntax

For anyone who prefers syntax over menu clicks, this is what the final COXREG command looked like. This matches the syntax structure documented in IBM’s own SPSS Statistics COXREG reference, applied to her specific variables.

COXREG time
  /STATUS=status(1)
  /METHOD=ENTER age sex ef comorbidity discharge_meds followup
  /PRINT=CI(95) CORR
  /PLOT=SURVIVAL HAZARD LML
  /CRITERIA=PIN(.05) POUT(.10) ITERATE(20).

If you are still deciding whether logistic regression fits your own data, run it through the logistic regression calculator first and compare the fit against what a Cox model would give you on the same variables.

Checking the Proportional Hazards Assumption

This is the step most students skip entirely, mainly because nobody tells them it exists. SPSS does not check the proportional hazards assumption for you automatically. You have to request it and interpret it yourself.

Log-Minus-Log Plots for Categorical Predictors

For categorical variables like sex and discharge disposition, we checked the log-minus-log survival plots. Parallel lines across groups support the proportional hazards assumption. Her plots were reasonably parallel, so the assumption held for these variables.

Time-Dependent Covariates for Continuous Predictors

For continuous predictors like age and comorbidity count, log-minus-log plots do not work well. Instead, we created time-dependent covariates by interacting each predictor with a function of time, then tested whether that interaction was statistically significant. None were significant, which confirmed the proportional hazards assumption held across the full model.

Want a faster first check before you get this deep? Run your predictors through the regression assumption checker to catch obvious violations before you touch COXREG.

Interpreting the Hazard Ratios

We exponentiated the coefficients to get hazard ratios, which is the standard way to interpret Cox regression output. A hazard ratio of 1.45 for comorbidity count meant each additional comorbidity increased the hazard of readmission by 45 percent, holding all other variables constant. That is a very different, and much more clinically useful, statement than anything logistic regression could have produced from the same data.

For write-up purposes, report each hazard ratio with its 95% confidence interval in APA format, for example: HR = 1.45, 95% CI [1.12, 1.88], p = .003. Committees expect this exact format, not a bare coefficient.

Cox Regression vs Logistic Regression: Which Fit Better

We ran both models side by side on her data for comparison. The Cox model had better fit statistics and used every data point available, including the censored ones. It also let her generate survival curves for different patient profiles, which mattered a great deal to a nursing dissertation with clinical application in mind.

My opinion here, and I will say it plainly: any nursing, public health, or clinical dissertation using a time-to-event outcome with plain logistic regression should be sent back before it ever reaches a committee meeting, not caught by a committee member during the defence. That is a supervision failure, not a student failure. Most students are simply never taught to recognise when their outcome is a survival outcome wearing a binary variable’s clothes.

The Result: How Her Conclusions Changed

Two predictors that were statistically insignificant in her logistic regression became significant in the Cox model. One predictor that had looked significant in logistic regression turned out not to be, once censoring was properly accounted for. Her conclusions changed, and they became more defensible.

She rewrote her analysis chapter around hazard ratios instead of odds ratios, explained her choice of Cox regression, and resubmitted. Her committee approved it on the next pass. What stayed with me was something she said afterward: she thought “readmitted, yes or no” was all she needed, and had not realised the timing was the actual point of her whole research question.

That is the risk with SPSS specifically. It will run a logistic regression on a time-to-event outcome without a single warning. The output looks perfectly normal. It just is not answering your research question.

How I Help Nursing and Health Science PhD Candidates With SPSS

As a nursing PhD statistical consultant working across SPSS, R, and Stata, here is what I actually help with:

• Identifying whether your outcome needs logistic, Cox, or another survival model
• Restructuring your dataset for time-to-event analysis
• Running and interpreting Cox proportional hazards regression in SPSS
• Checking the proportional hazards assumption properly, for both categorical and continuous predictors
• Building and interpreting survival and hazard plots
• Comparing model fit so you can justify your choice to a committee
• Making sure you understand your own output well enough to defend it

A straightforward proportional hazards check usually turns around within a few days. A full model rebuild, like the one described above, typically takes longer depending on dataset size and how many predictors are involved.

I’ll make sure every number in your analysis is correct and airtight, worked through together so you can stand behind it in your defence, not handed to you as a black box. You can see the full scope of this on my statistical consulting services page, or book SPSS data analysis tutoring directly if you want to work through your own model with me.

If you are not sure whether your project needs this kind of outside check at all, this piece on how to know if you need a dissertation expert is a fair starting point.

Frequently Asked Questions

Siddharth Gupta, statistical consultant with over 12 years of research and analysis guidance experience. LinkedIn