What Is Multivariate Analysis in Medical Research?
The term "multivariate analysis" is used broadly in medical literature, but it carries a specific technical meaning that is often confused. Strictly speaking, multivariate analysis refers to statistical methods that model multiple outcome (dependent) variables simultaneously for example, MANOVA (Multivariate Analysis of Variance), which tests whether a treatment affects blood pressure, cholesterol, and BMI all at once.
However, in clinical research publications, "multivariate analysis" almost universally refers to what statisticians call multivariable regression a model with a single outcome variable and multiple predictor (independent) variables. You will see this usage throughout peer-reviewed journals, thesis guidelines, and ethics committee submissions. For the purposes of this guide, we use "multivariate" and "multivariable" interchangeably to match the language used in clinical practice, while being clear about the underlying method.
In essence, multivariate analysis in medical research means: examining the relationship between multiple predictors and a single outcome, all in one model, to determine the independent effect of each predictor while controlling for the others.
Univariate vs Multivariate Analysis The Critical Difference
Understanding the distinction between univariate and multivariate analysis is fundamental to reading and conducting clinical research correctly. The key is not just what they measure it is what they cannot account for.
Univariate Analysis
- One predictor, one outcome
- Produces crude (unadjusted) estimates
- Cannot control for confounders
- Used for initial screening of associations
- Examples: chi-square, independent t-test, simple logistic regression
- Reports: crude OR, crude RR, mean difference
Multivariate Analysis
- Multiple predictors, one outcome
- Produces adjusted estimates
- Controls statistically for included confounders
- Used for inferring independent associations
- Examples: multiple logistic/linear/Cox regression
- Reports: aOR, aHR, adjusted 2 coefficient
Why Crude Estimates Can Be Misleading
Consider a study examining whether smoking is a risk factor for diabetic nephropathy. The univariate analysis yields a crude OR of 1.95 (95% CI 1.382.76, p < 0.001) a seemingly robust association. However, smokers in the dataset also tend to have worse blood glucose control and higher rates of hypertension compared to non-smokers. When HbA1c and hypertension are entered into the multivariate model alongside smoking, the adjusted OR for smoking drops to 1.31 (95% CI 0.891.93, p = 0.162) no longer statistically significant.
The apparent association between smoking and nephropathy was largely explained by the fact that smokers had poorer metabolic control not by smoking itself. This is confounding, and it is the primary reason multivariate analysis exists.
Confounding Variables The Core Problem Multivariate Analysis Solves
A confounding variable (or confounder) is a variable that is associated with both the predictor (exposure) and the outcome and its presence distorts the observed relationship between them, making the association appear stronger, weaker, reversed, or even non-existent compared to the true underlying relationship.
The Classic Confounding Diagram
(e.g., HbA1c level)
(e.g., Smoking)
(e.g., Nephropathy)
The confounder is associated with both the exposure and the outcome independently. This creates a spurious (or distorted) association between exposure and outcome in crude analysis.
Three Criteria for a True Confounder
A variable is a confounder if it meets all three of the following criteria simultaneously:
- Associated with the exposure the confounder must be more common (or have a different distribution) among exposed vs unexposed subjects
- Associated with the outcome the confounder must independently predict the outcome, even in the absence of the exposure
- Not on the causal pathway the confounder must not be an intermediate step between the exposure and the outcome (a mediator is different from a confounder)
Classic Examples of Confounding in Medical Research
| Observed Association | True Confounder | What Was Actually Happening |
|---|---|---|
| Coffee drinking lung cancer | Cigarette smoking | Coffee drinkers smoked more; smoking caused the cancer, not coffee |
| Hormone therapy less heart disease | Socioeconomic status / healthy user bias | Women on HRT tended to be healthier overall; early observational studies were confounded |
| Stork sightings higher birth rates | Rural vs urban setting | Rural areas have more storks AND higher birth rates spurious ecological correlation |
| Gray hair heart disease risk | Age | Older people have more gray hair AND more heart disease; age drives both |
| Shoe size reading ability in children | Age | Older children have larger feet AND read better; age is the confounder |
Residual Confounding
Multivariate analysis controls only for the confounders you actually measure and include in the model. Variables that are unmeasured, measured imprecisely, or simply not available in your dataset cannot be adjusted for. This is called residual confounding it is an inherent limitation of all observational studies and should be acknowledged in your discussion section. Only a properly conducted randomized controlled trial, with randomization achieving balance on all characteristics (measured and unmeasured), can eliminate confounding entirely.
Regression-Based Multivariate Analysis Which Method for Which Outcome?
The choice of regression method is determined primarily by the nature of your outcome variable, not by your personal preference or what is most familiar. Using the wrong regression method is one of the most common methodological errors reviewed in thesis defenses and manuscript peer review.
| Outcome Variable Type | Clinical Example | Method | Output Statistic |
|---|---|---|---|
| Binary (yes / no) | Diabetes: present / absent | Multiple Logistic Regression | Adjusted Odds Ratio (aOR) |
| Continuous (normally distributed) | HbA1c (%), blood pressure (mmHg) | Multiple Linear Regression | 2 coefficient (change per unit) |
| Time-to-event (with censoring) | Overall survival (months to death) | Cox Proportional Hazards | Adjusted Hazard Ratio (aHR) |
| Count data (non-negative integers) | Number of ED visits per year | Poisson / Negative Binomial | Incidence Rate Ratio (IRR) |
| Ordinal (ranked categories) | Pain score 15, ECOG PS 04 | Ordinal Logistic Regression | Proportional Odds Ratio |
| Binary, rare outcome (< 10%) | Rare complication rate | Logistic Regression (OR RR) | aOR (approximates RR when rare) |
Multiple Logistic Regression The Most Common Method in Clinical Research
Multiple logistic regression is by far the most frequently used multivariate method in clinical and epidemiological research. It is appropriate whenever your outcome is binary meaning each patient either did or did not experience the event of interest: 30-day readmission (yes/no), postoperative complication (yes/no), disease diagnosis (positive/negative), treatment response (yes/no).
Crude OR vs Adjusted OR What Changes and Why
The key output of logistic regression is the odds ratio (OR). From univariate logistic regression, you obtain a crude OR the association between one predictor and the outcome without accounting for anything else. From multiple logistic regression, you obtain an adjusted OR (aOR) the independent association between each predictor and the outcome, with all other variables in the model held constant.
When you read a clinical paper reporting aOR = 2.14 (95% CI 1.483.09) for ECOG performance status 2 versus 0, it means: after adjusting for age, disease stage, treatment type, and all other variables in the model, patients with ECOG PS 2 had 2.14 times the odds of the outcome compared to patients with ECOG PS 0. The other variables' influence has been statistically removed from this estimate.
Clinical Example: Risk Factors for Diabetic Nephropathy
Study: Cross-sectional study of 520 patients with type 2 diabetes mellitus (T2DM) at a tertiary diabetes clinic. Outcome: diabetic nephropathy (defined as eGFR < 60 mL/min/1.73m2 or urine ACR 30 mg/g). Predictors: HbA1c, hypertension, smoking, diabetes duration, age, and BMI.
| Predictor | Crude OR (95% CI) | Crude p | Adjusted OR (95% CI) | Adjusted p |
|---|---|---|---|---|
| HbA1c (per 1%) | 1.82 (1.542.15) | < 0.001 | 1.74 (1.462.08) | < 0.001 |
| Hypertension (yes vs no) | 2.34 (1.673.27) | < 0.001 | 1.91 (1.342.72) | < 0.001 |
| Smoking (yes vs no) | 1.95 (1.382.76) | < 0.001 | 1.31 (0.891.93) | 0.162 |
| Age (per 10 years) | 1.42 (1.181.71) | < 0.001 | 1.28 (1.051.56) | 0.013 |
| Diabetes duration (per year) | 1.11 (1.071.15) | < 0.001 | 1.09 (1.051.13) | < 0.001 |
| BMI (per kg/m2) | 1.08 (1.031.14) | 0.003 | 1.03 (0.981.09) | 0.231 |
How to Interpret the Odds Ratio
- aOR = 1.0: No association between predictor and outcome after adjustment
- aOR > 1.0: Higher exposure is associated with increased odds of the outcome
- aOR < 1.0: Higher exposure is associated with reduced odds of the outcome (protective)
- If 95% CI crosses 1.0: The association is not statistically significant at alpha = 0.05
- aOR vs RR: When outcome prevalence exceeds 10%, the OR overestimates the relative risk consider Poisson regression or modified Poisson regression to estimate the risk ratio directly
Multiple Linear Regression When Your Outcome Is Continuous
When your outcome variable is continuous and approximately normally distributed such as HbA1c level, systolic blood pressure, serum creatinine, or length of hospital stay multiple linear regression is the appropriate multivariate method. Instead of an odds ratio, it produces a regression coefficient (2) for each predictor, representing the expected change in the outcome per one-unit increase in the predictor, with all other variables held constant.
Clinical Example: Predictors of HbA1c Level in T2DM
Study: Cohort of 380 outpatients with type 2 diabetes. Outcome: HbA1c (%) at 12-month follow-up. Predictors: diabetes duration, physical activity score, medication adherence score, age, and BMI.
| Predictor | 2 Coefficient | 95% CI | p-value | Interpretation |
|---|---|---|---|---|
| Medication adherence (per 10 pts) | 0.31 | 0.38 to 0.24 | < 0.001 | Each 10-point increase in adherence HbA1c 0.31% lower |
| Physical activity (per 10 pts) | 0.18 | 0.24 to 0.12 | < 0.001 | Each 10-point increase in activity HbA1c 0.18% lower |
| Diabetes duration (per year) | +0.12 | +0.09 to +0.15 | < 0.001 | Each additional year of diabetes HbA1c 0.12% higher |
| Age (per 10 years) | +0.08 | +0.01 to +0.15 | 0.024 | Each 10 additional years of age HbA1c 0.08% higher |
| BMI (per kg/m2) | +0.04 | 0.01 to +0.09 | 0.097 | Not independently associated after adjustment |
Key Metrics in Multiple Linear Regression
- R2 (coefficient of determination): The proportion of variance in the outcome explained by all predictors combined. R2 = 0.41 means the model explains 41% of the variability in HbA1c.
- Adjusted R2: Penalized for the number of predictors; always report adjusted R2 instead of R2 when comparing models with different numbers of variables.
- Standardized 2 (beta): Allows comparison of effect sizes across predictors measured in different units a standardized 2 of 0.42 for adherence vs 0.28 for activity indicates adherence has a stronger independent effect on HbA1c.
Cox Regression When Your Outcome Is Time-to-Event
Cox proportional hazards regression extends the multivariate framework to time-to-event outcomes with censoring the same data type analyzed by the Kaplan-Meier method and log-rank test. While KM gives you an unadjusted survival curve and the log-rank test gives an unadjusted comparison between groups, Cox regression gives you an adjusted hazard ratio (aHR) for each predictor while simultaneously controlling for all other variables in the model.
Study: Retrospective cohort of 290 patients with stage III colorectal cancer undergoing curative resection. Outcome: Overall survival (time from surgery to death). Predictors entered: age, ECOG performance status, T stage, N stage, pre-operative CEA level, and vascular invasion on pathology.
| Predictor | Adjusted HR | 95% CI | p-value |
|---|---|---|---|
| ECOG PS 2 (vs PS 0) | 2.14 | 1.483.09 | < 0.001 |
| N2 disease (vs N1) | 1.73 | 1.222.46 | 0.002 |
| CEA > 5 ng/mL (vs 5) | 1.58 | 1.112.25 | 0.011 |
| Vascular invasion (yes vs no) | 1.44 | 0.992.10 | 0.058 |
| T4 vs T3 stage | 1.31 | 0.921.87 | 0.132 |
| Age (per 10 years) | 1.19 | 0.971.46 | 0.094 |
For a full explanation of Kaplan-Meier curves, censored data, and the log-rank test underlying Cox regression, see our article on survival analysis in medical research.
How to Build a Multivariable Regression Model Step by Step
Model building is where multivariate analysis goes right or wrong. The approach must be pre-specified in your study protocol or statistical analysis plan before looking at the data. Here is the recommended process for clinical research.
Define your primary outcome and research question
Before selecting any variables, clearly state what you are trying to answer: "What are the independent risk factors for X?" or "Is predictor Y independently associated with outcome Z after adjusting for potential confounders?" The outcome type (binary, continuous, time-to-event) determines your regression method.
Identify candidate variables a priori from the literature
List all variables you plan to enter into the model based on biological plausibility and existing evidence before analyzing your data. Include known confounders even if they turn out to be non-significant in your univariate screening. This approach is called a priori (or theory-driven) variable selection and is the scientifically defensible standard.
Check events per variable (EPV) constraints
For logistic regression: you need 10 outcome events per predictor. For Cox regression: 10 events per predictor. For linear regression: 1020 observations per predictor. If your study is underpowered for the number of variables you want to include, prioritize the most clinically important ones and acknowledge the limitation.
Run univariate screening (optional, not decisive)
Running univariate analysis on all candidate variables gives you crude effect sizes to compare with adjusted results later which is valuable for demonstrating confounding. Some researchers use p < 0.10 or p < 0.20 as a screening threshold to reduce the variable list, but this approach risks excluding important confounders. Always override screening criteria when a variable is a known confounder from the literature.
Check model assumptions before interpreting results
For logistic regression: check linearity in the logit for continuous predictors, absence of multicollinearity (VIF < 5), and adequate sample size. For linear regression: check normality of residuals, homoscedasticity (constant variance), and absence of influential outliers (Cook's D). For Cox regression: verify the proportional hazards assumption using Schoenfeld residuals or log-log plots.
Report both crude and adjusted estimates in your results table
The standard reporting format for multivariate analysis shows the crude (univariate) OR and the adjusted OR side by side for each predictor, with 95% CI and p-values for both. This allows readers to identify which associations changed substantially after adjustment direct evidence of confounding. See the diabetic nephropathy example above for the correct table format.
Common Mistakes Researchers Make in Multivariate Analysis
Mistake 1: Calling univariate results "multivariate" in the paper
One of the most common errors in submitted manuscripts reporting a single-variable chi-square or t-test result as "multivariate analysis." A univariate test with one predictor is not multivariate by any definition. This mislabeling confuses readers, misrepresents what was done analytically, and is immediately flagged by statistical reviewers.
Mistake 2: Using stepwise regression as the primary analytical strategy
Forward, backward, or bidirectional stepwise regression selects variables based purely on statistical thresholds, ignoring clinical knowledge. The resulting model is optimized for the current sample and tends to overfit it will not replicate in a new dataset. More critically, it may drop important confounders simply because they failed to reach the entry threshold, producing biased adjusted estimates for the retained predictors.
Mistake 3: Including too many variables relative to events (overfitting)
Including 15 predictors in a logistic regression model when only 40 patients experienced the outcome (EPV = 2.7) is a severe overfitting problem. The model memorizes the sample data instead of detecting real patterns it will have inflated apparent precision, spuriously significant p-values, and will fail completely when applied to a new dataset. The 95% confidence intervals will be artifactually narrow.
Mistake 4: Ignoring multicollinearity
Including two highly correlated predictors simultaneously such as serum creatinine and estimated GFR (which is calculated from creatinine), or systolic and diastolic blood pressure, or BMI and body weight creates multicollinearity. The model cannot separate their individual effects: standard errors explode, CIs widen dramatically, regression coefficients become unstable and may even reverse sign, and the model is essentially meaningless for those predictors despite appearing to run correctly.
Mistake 5: Interpreting OR as Risk Ratio when outcome prevalence is high
The odds ratio approximates the risk ratio (relative risk) only when the outcome is rare typically defined as prevalence below 10%. When 3040% of subjects experience the outcome (very common in hospital-based cross-sectional studies), the OR substantially overestimates the risk ratio, sometimes by 50100%. Reporting "patients with X had 3.4 times the risk of Y" when the OR is 3.4 and outcome prevalence is 35% is a significant overstatement.
Mistake 6: Not checking the proportional hazards assumption in Cox regression
Cox regression assumes that the hazard ratio between groups remains constant over the entire follow-up period. This is the proportional hazards (PH) assumption. Researchers frequently fit a Cox model and report the hazard ratio without ever verifying this assumption. If the PH assumption is violated which is common with immunotherapy trials, rapidly progressing diseases, or long follow-up the reported hazard ratio is a meaningless average of a time-varying effect.
Frequently Asked Questions
Use multiple linear regression when your outcome is continuous and approximately normally distributed blood pressure in mmHg, HbA1c percentage, hospital length of stay in days, laboratory values. The output is an unstandardized regression coefficient (2) representing the change in the outcome per one-unit change in the predictor, adjusted for all other variables.
If your continuous outcome is highly skewed (e.g., length of stay, cost data), consider log-transforming it before entering it into linear regression, or use a generalized linear model with an appropriate link function.
- Logistic regression: 10 outcome events per predictor variable in the model
- Cox regression: 10 events (not patients events) per predictor
- Linear regression: 1020 observations per predictor
An adjusted odds ratio (aOR) comes from multiple logistic regression it shows the association between a predictor and the outcome after statistically controlling for all other variables in the model, giving the independent effect of that predictor.
When the crude OR and aOR differ substantially (change-in-estimate criterion: typically > 10% change in the OR on the log scale), confounding is present. Only the aOR should be used to characterize a variable as an independent risk factor. The crude OR is still valuable to report because showing the change from crude to adjusted demonstrates the presence and direction of confounding.
When predictors are highly correlated, the model cannot reliably estimate the independent effect of each one. Standard errors inflate, confidence intervals widen dramatically, and regression coefficients become numerically unstable sometimes reversing sign between similar models.
How to detect: Variance inflation factor (VIF). VIF > 5 is concerning; VIF > 10 indicates severe multicollinearity. Also inspect the correlation matrix of all continuous predictors before building the model.
Solution: Remove one of the collinear pair, combine them into a composite variable (e.g., a composite metabolic score), or choose one based on clinical priority. Do not include eGFR and creatinine in the same model.
- It relies entirely on p-value thresholds to select variables, ignoring clinical knowledge
- It produces models overfit to the current sample that do not replicate in new datasets
- It may exclude important confounders that happen to be non-significant in this particular dataset
- It capitalizes on chance associations in the data
A variable with p = 0.15 in univariate analysis may still be an important confounder if it is biologically linked to both the exposure and the outcome, it must be included in the model regardless of its univariate p-value. Excluding it produces a misspecified model with biased adjusted estimates for all remaining predictors.
Use a liberal threshold for univariate screening if you use one at all (p < 0.20 rather than p < 0.05), but always override the threshold for variables that are known confounders from the existing literature.
Run Multivariate Regression Analysis Free
StatClinic includes logistic regression, linear regression, Cox proportional hazards, and multiple regression tools all free, no login, with adjusted ORs, hazard ratios, and APA-formatted output tables generated automatically.
Open Regression Tools