Confounding variables in a nursing dissertation need a reasoned selection process before you run an adjusted model. Begin by defining the exposure, outcome, population and timing. Then consider which prior factors could create a misleading comparison between the groups you want to study.
This guide explains how to justify adjustment decisions in quantitative nursing research. Its focus is the reasoning behind the model, rather than SPSS menu instructions. The examples are fictional teaching scenarios, and the numerical results are invented. They must not be presented as findings from a real study.
What are confounding variables in a nursing dissertation?
Consider an observational study examining the relationship between a nursing exposure and a later outcome. If the exposed and unexposed groups differ in another relevant way, the observed relationship may partly reflect that difference. A potential confounder is therefore considered in relation to a specific question and a proposed causal structure, rather than being an inherent property of a variable.
For example, imagine comparing fatigue scores between nurses who predominantly work nights and those who predominantly work days. Prior health, caring responsibilities and baseline fatigue might influence work-pattern selection and subsequent fatigue. Whether each factor should enter the analysis depends on its timing, measurement and plausible relationship with the exposure and outcome.
Do not automatically describe every demographic variable as a confounder. A baseline characteristic might be relevant to precision, a mechanism, a selection process or none of these. The methods chapter needs to explain the role you propose rather than use “controlled for demographics” as a substitute for reasoning.
VanderWeele (2019) explains why subject knowledge and causal assumptions matter in confounder selection. Statistical associations alone cannot reliably establish the temporal and causal roles needed to distinguish adjustment choices.
Start with the question and the effect of interest
Write the exposure and outcome in concrete terms. “Staffing and safety” is too broad to guide adjustment. A more usable question would specify the staffing measure, relevant time window, outcome definition and setting. The same variable can occupy different roles when these details change.
Next decide what you are trying to estimate. A study of an exposure’s overall relationship with an outcome is different from a study asking about a particular pathway after a later intermediate variable is considered. Do not choose the interpretation after looking at whichever model produces the most attractive result.
Also distinguish causal explanation from prediction. A prediction model may be evaluated for how accurately it forecasts an outcome. An explanatory analysis needs a defensible account of what its coefficients represent. A variable that improves prediction does not automatically belong in a causal adjustment set.
Write a short planning statement before opening the software: “We will examine X measured at time A in relation to Y measured at time B, considering the prior factors listed below.” This statement makes later decisions easier to evaluate and exposes unclear timing early.
Separate confounders, mediators, colliders and modifiers
| Proposed role | Meaning in the planned analysis | Question before adjustment |
|---|---|---|
| Confounder | A factor involved in a non-causal explanation for the exposure–outcome comparison | What prior relationships justify considering it? |
| Mediator | A variable on a proposed pathway from exposure to outcome | Would adjustment remove part of the overall relationship of interest? |
| Collider | A common effect of two variables on a causal path | Could conditioning create an association that was previously blocked? |
| Effect modifier | A factor across whose levels an effect differs on a specified scale | Should differences be examined and reported rather than averaged away? |
| Precision variable | A variable included to improve estimation under an appropriate model | Is the purpose clear and compatible with the study design? |
The table is a planning overview, not a rule that determines the correct role from a dataset. Draw the proposed sequence of events and discuss it with your supervisor. If a variable was measured after exposure began, that does not prove mediation, but it should prompt a careful review of whether it represents a consequence of the exposure.
Schisterman, Cole and Platt (2009) distinguish overadjustment bias from unnecessary adjustment. Their analysis supports a central warning: adding variables can change the question or introduce problems, so a longer model is not automatically a better model.
Build a covariate decision log
Create one row for each candidate covariate. Record its definition, measurement time, proposed role, supporting reasoning and intended treatment in the model. Include the reason for excluding a plausible factor, especially if readers might otherwise expect it.
For the fictional night-work study, a log might note that baseline fatigue was measured before the work-pattern period and could influence subsequent allocation or self-selection. By contrast, sleep measured during the work-pattern period could partly reflect a consequence of night working. The two measures should not be treated as interchangeable merely because both concern tiredness.
Do not claim that a causal diagram proves the assumptions used to draw it. Its value is making the assumptions explicit. If credible alternative structures would require different adjustment sets, describe that uncertainty and consider an appropriate sensitivity analysis with qualified statistical guidance.
Keep the log with your analysis plan. Later changes should record the reason and whether the decision was made before or after examining results. Transparent reporting is more defensible than presenting every exploratory choice as if it had been planned from the beginning.
Use study design to reduce avoidable confounding
Confounding is easier to address when important information has been considered before data collection. Once a dissertation dataset lacks a key baseline measure, software cannot recreate it reliably through wishful adjustment. Acknowledge the limitation instead of inserting an unrelated variable and treating it as equivalent.
Potential design approaches include restriction, matching and random allocation where appropriate to the research question and feasible design. Each changes what can be learned and which participants are represented. For example, restricting a study to a narrow population may simplify one comparison while limiting its applicability elsewhere.
Matching also creates analysis and reporting requirements. Record what was matched, how matching was performed, whether suitable matches were available and how the matched structure will be handled. Do not describe ordinary convenience sampling as matching simply because the final groups have similar average ages.
For a student analysing an existing dataset, the most useful design contribution may be an honest reconstruction of how participants entered the study. Consider whether exposure, outcome or relevant background factors influenced participation or inclusion in the analytical sample.
Choose an analysis method that fits the outcome
Stratification can make a simple relationship easier to examine within meaningful groups. Regression can estimate an exposure–outcome relationship while including a planned set of covariates. Other approaches require additional assumptions and expertise. Select the method for the question and data, rather than because it sounds more advanced.
The outcome’s scale matters. Continuous, binary, count and time-to-event outcomes usually require different modelling choices. A confounder selection argument does not remove the need to check assumptions, coding, clustering or the relationship between the model and the sampling design.
If you need software guidance after settling these decisions, use our separate guide to nursing regression analysis in SPSS. Keep the rationale for including variables in the methods chapter and the software execution details in a clearly documented analysis record.
Check how continuous covariates are represented. Converting a continuous measure into broad categories can discard information and leave differences within categories unaddressed. Conversely, assuming a simple linear relationship without checking its plausibility can also misrepresent the data.
A worked adjusted-results example
All numbers in this example are fabricated for teaching. No study or dataset is being reported. Imagine a cohort of 210 nurses with complete information for the illustrative variables. The exposure is predominant night work, and the continuous outcome is a later fatigue score.
| Teaching model | Illustrative estimated difference | Illustrative 95% confidence interval | Analytical sample |
|---|---|---|---|
| Unadjusted comparison | 2.4 points | 1.2 to 3.6 | 210 |
| Adjusted for the planned baseline covariates | 1.5 points | 0.2 to 2.8 | 210 |
A proportionate teaching interpretation would say that the adjusted estimated difference was smaller than the unadjusted difference in this fictional sample. It would identify the actual covariates and explain their selection. It would not state that adjustment “removed all confounding” or established that night work caused the difference.
Using the same analytical sample makes this particular comparison easier to interpret. If the adjusted model excludes participants because covariates are missing, a coefficient change may partly reflect a changed sample. Consider a comparable unadjusted model using the adjusted model’s sample, alongside an appropriate missing-data strategy.
The substantive meaning of 1.5 points also depends on the fatigue instrument and context. Statistical precision does not establish practical importance. Avoid calling a numerical difference clinically important unless a relevant, supported interpretation is available.
Why a changed odds ratio is not automatic proof of confounding
Students sometimes label a covariate a confounder whenever an odds ratio changes after adjustment. That shortcut is unreliable. Conditional and marginal odds ratios can differ because of noncollapsibility, even when the additional factor is not a confounder of the exposure–outcome relationship (Whitcomb and Naimi, 2021).
Therefore, separate the causal justification for adjustment from a descriptive comparison of model estimates. A change-in-estimate check may be part of an analysis strategy, but it cannot supply missing knowledge about time order and causal structure.
When reporting logistic regression, identify the outcome coding, exposure reference group, adjustment set and confidence interval. Do not switch between odds, risks and probabilities as though they were identical. If your intended audience needs an absolute interpretation, discuss a suitable approach with your supervisor or statistical adviser.
Address missingness and limited sample size
Before fitting the planned model, inspect how much data are missing for each variable and how the analytical sample changes. Describe the procedure actually used. A complete-case analysis, an imputation approach and a missing-indicator method are not interchangeable labels for the same process.
A small dataset may not support every parameter you would ideally estimate. Categories, interactions and nonlinear terms all add complexity. Do not use a universal participant-per-variable slogan as the only justification; the outcome distribution, parameter count and intended estimation task also matter.
If the dataset cannot answer the intended question convincingly, narrow the claim. A carefully qualified exploratory analysis can still be valuable. Overstating certainty to compensate for limited data makes the dissertation harder to defend.
Also consider whether participants are clustered within wards, hospitals or teams. Including a ward characteristic does not by itself resolve statistical dependence among observations. The sampling structure and the model’s uncertainty estimates need appropriate consideration.
Reporting confounding variables in a nursing dissertation
A useful methods paragraph identifies how candidate covariates were selected, when they were measured, how they were coded and which model was planned. It also explains departures from the plan. Keep the account specific enough for another reader to understand what was done.
For example, you might write a template such as: “Covariates were specified before modelling using the study question, measurement timing and a documented causal rationale.” Replace that sentence with the actual process used. Do not copy it if variables were selected entirely through an exploratory automated procedure.
STROBE’s explanation and elaboration addresses transparent reporting of observational studies, including estimates, adjustment and relevant analytical decisions (Vandenbroucke et al., 2007). Reporting guidance can help identify omissions, but it cannot convert a poorly justified adjustment set into a sound one.
Put a compact covariate table in the methods chapter where space allows. If fuller details are placed in an appendix, signpost them clearly and keep the essential rationale in the main text. Our critical dissertation writing guide can help you explain limitations without abandoning the argument.
Report remaining uncertainty in the discussion
Adjustment depends on measured variables, their quality and the assumptions behind the model. Unmeasured confounding may remain. A poorly measured covariate may also fail to represent the underlying factor well enough to address the intended concern.
Discuss specific weaknesses. Instead of writing only “confounding may exist”, identify an important unmeasured factor, explain how it might relate to the comparison and describe the resulting limit on interpretation. Do not invent its direction or magnitude unless you have evidence or a clearly stated sensitivity scenario.
Consider whether reverse causation is plausible. In a cross-sectional survey, exposure and outcome may be measured together, making temporal interpretation difficult. Statistical adjustment does not establish which came first.
Finally, keep recommendations proportionate. An observational association may justify further investigation or careful consideration, but it does not automatically establish that changing the exposure will produce the estimated outcome change.
Practical adjustment checklist
- Define the population, exposure, outcome and measurement times.
- Distinguish explanation from prediction.
- Record a rationale for each candidate covariate.
- Review mediators, colliders and possible selection processes.
- Check coding, missingness, sample size and clustering.
- Report unadjusted and adjusted results transparently where appropriate.
- State residual uncertainty without claiming that adjustment proves causation.
Prepare for a supervisor’s questions about adjustment
Practise explaining the model without naming the software. State the comparison of interest, identify the proposed prior influences and explain what changes when they are considered. If you can explain only which buttons were selected, return to the covariate decision log.
A useful rehearsal question is: “What would make this variable inappropriate to adjust for?” Consider whether it could be a consequence of exposure, a selection variable or a poorly measured proxy. The answer does not need to eliminate every uncertainty, but it should show that inclusion was a deliberate decision.
Also prepare to explain excluded variables. A dissertation can acknowledge that an important measure was unavailable without pretending the limitation has disappeared. Keep the distinction between what the model estimates under its assumptions and what the study can establish in practice. That distinction will help you write a discussion that is confident about the analysis while appropriately cautious about causal interpretation.
Frequently asked questions
Must age and sex always be included?
No. Their role depends on the question, causal structure and available information. Explain why each is relevant in your analysis rather than treating familiar demographic variables as mandatory confounders.
Can SPSS identify the correct confounders?
Software can estimate specified models and calculate statistical relationships. It cannot independently supply the clinical knowledge, time ordering and causal assumptions needed to determine a defensible adjustment set.
Should I include every statistically significant variable?
Statistical significance is not a sufficient selection rule for causal adjustment. A variable’s proposed role and timing matter even when its individual coefficient is imprecise. Conversely, a statistically associated variable can be inappropriate to adjust for.
Does an adjusted association show causation?
Not by itself. Causal interpretation requires additional assumptions about confounding, selection, measurement and the model. Use language that reflects what the study design and evidence can support.
What if the dataset lacks an important confounder?
Document the gap, consider whether the question remains answerable, and discuss appropriate sensitivity work if feasible. Do not claim that another loosely related variable fully replaces the missing information.
Conclusion and further support
Controlling confounding begins with the question and the proposed relationships between variables. A defensible dissertation explains why covariates were selected, fits an appropriate analysis and reports what adjustment cannot resolve. More variables and smaller p-values are not substitutes for that reasoning.
For help reviewing the rationale behind your model, explore nursing dissertation data analysis support. You can also explore our nursing dissertation support, read About Us and How It Works, or review service feedback. If you need guidance, request a review of your analysis plan. You should retain responsibility for data permissions, analytical decisions and the work submitted to your university.
Featured photograph: Nappy / Unsplash. Illustrative stock image; it does not depict the studies or participants discussed.
References
Schisterman, E.F., Cole, S.R. and Platt, R.W. (2009) ‘Overadjustment bias and unnecessary adjustment in epidemiologic studies’, Epidemiology, 20(4), pp. 488–495. Available at: https://doi.org/10.1097/EDE.0b013e3181a819a1 (Accessed: 17 October 2026).
Vandenbroucke, J.P. et al. (2007) ‘Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): Explanation and elaboration’, PLOS Medicine, 4(10), e297. Available at: https://doi.org/10.1371/journal.pmed.0040297 (Accessed: 17 October 2026).
VanderWeele, T.J. (2019) ‘Principles of confounder selection’, European Journal of Epidemiology, 34(3), pp. 211–219. Available at: https://doi.org/10.1007/s10654-019-00494-6 (Accessed: 17 October 2026).
Whitcomb, B.W. and Naimi, A.I. (2021) ‘Defining, quantifying, and interpreting “noncollapsibility” in epidemiologic studies of measures of “effect”’, American Journal of Epidemiology, 190(5), pp. 697–700. Available at: https://doi.org/10.1093/aje/kwaa267 (Accessed: 17 October 2026).
