Statistics · Advanced
Read PCA, forest and survival plots
Read model based displays while keeping their assumptions visible.
12 min read · Oct 26, 2026
Key points
- Interpret PCA scores and loadings.
- Read estimates and intervals in forest plots.
- Account for event timing and censoring.
Contents
Read model based displays while keeping their assumptions visible.
Some figures display a statistical model or a complex structure rather than measurements alone. This tutorial introduces principal component analysis, forest plots and survival curves. Each can be valuable in a research or review paper, but each answers a different question and depends on choices that must be reported.
Use this article to learn how to read the axes, identify the unit, interpret uncertainty and draft a careful explanation. It is an introduction to reporting and interpretation, not a complete training course in multivariate analysis, evidence synthesis or event time modelling. All supplied illustrations and numerical teaching examples are synthetic.
Know which question each display answers
| Display | Main question | What it does not establish by itself |
|---|---|---|
| PCA score plot | Which samples vary together along prominent directions? | A new biological type or a significant group difference |
| Forest plot | How do estimates and intervals compare on a common scale? | That every study estimates the same effect or has low bias |
| Survival curve | How does the estimated chance of remaining event free change with time? | That censoring is harmless or a group difference is causal |
The first step is always to identify how the plotted values were produced. A complicated figure can be accurately drawn from an inappropriate analysis. Looking professional does not make the underlying inference valid.
PCA replaces many variables with new axes
Principal component analysis, or PCA, constructs weighted combinations of measured variables. The first principal component captures the greatest variance available under the chosen preprocessing. Subsequent components capture additional variance while remaining orthogonal to earlier components. Scores are the coordinates of observations on these axes. Loadings describe how the original variables contribute to the components. Official software documentation explains this decomposition and its preprocessing assumptions. S35
Imagine a hypothetical materials project measuring strength, density, conductivity and cost for several designs. A score plot might place designs with similar combinations of these measurements near one another. It does not mean the horizontal axis is “quality”. A component that combines high strength with high cost may represent a tradeoff rather than an improvement.
To interpret an axis, inspect the loadings and return to the original variables. Naming a component “motivation” because it separates survey respondents is not enough. The interpretation must fit the measured items and the scientific purpose, and may require a different measurement model if the goal is to establish a latent construct.
Explain centring, scaling and explained variance
Centring subtracts each variable's mean. Scaling may additionally divide by a measure such as its SD. If one variable is measured in large numerical units and another in small units, an unscaled PCA can be dominated by the larger variance variable. Standardisation changes the question to variation relative to each variable's spread; it is not automatically better in every application.
State transformations, missing value treatment and feature selection. A PCA of all measured variables differs from a PCA of variables selected because they distinguish groups. If measurements were transformed or batch adjusted first, describe that process and its assumptions.
Put the proportion of variance explained on each displayed axis. For the supplied synthetic dataset, the first component accounts for approximately 39.24% and the second for 25.03%, for a combined 64.27%. About 35.73% is therefore outside the displayed plane. Two points close together in this view can differ on omitted components.
The sign of a component can be reversed without changing the underlying PCA solution, provided the associated scores and loadings are reversed together. Do not attach scientific importance to left versus right without examining the loading direction. Reproducibility requires documenting the calculation, not expecting every implementation to place a cluster on the same side of a page.
Do not turn a PCA picture into a hypothesis test
Colouring samples by treatment, institution or device batch can help reveal patterns. Visible separation may reflect the target phenomenon, a nuisance factor or a combination. Check batch, processing date, measurement quality and other plausible explanations before interpreting the pattern.
Ellipses around groups need explicit definitions. They may summarise dispersion, uncertainty about a group centre or another model based quantity. Overlap or nonoverlap of unspecified ellipses is not a universal significance test.
If PCA is part of a prediction pipeline, fit scaling and components using training data only, then apply those learned transformations to evaluation data. Fitting preprocessing on the full dataset can leak information about the evaluation distribution. A plot used solely for exploratory description has a different purpose, which should be stated rather than confused with predictive validation.
A useful hypothetical reporting sentence is: “The first two components summarised 64% of the standardised feature variation; the score plot suggested a batch related pattern that was investigated in the original measurements.” It states what the display contributes and what remains to be checked.
Read a forest plot from its scale outward
A forest plot places several effect estimates and uncertainty intervals on one horizontal scale. Each row may represent a study in a review, a subgroup estimate or a coefficient from one fitted model. These are different uses, so define the row unit before interpreting the symbols.
A mean difference has a null value of zero. A risk ratio, odds ratio or hazard ratio has a null value of one. Ratios are often shown on a logarithmic scale because reciprocal ratios then occupy symmetric positions around one. These measures are not interchangeable: an odds ratio is not generally a risk ratio, and a hazard ratio is not a ratio of cumulative event probabilities.
Read the effect direction and reference group. “Favours treatment” can be ambiguous unless the outcome and subtraction or ratio direction are defined. A reduction in an adverse event may be beneficial; a reduction in a desired completion rate may not be.
Consider a separate hypothetical estimate of a risk ratio of 0.80 with a 95% confidence interval from 0.60 to 1.07. The point estimate corresponds to a 20% lower relative risk, but the interval includes values representing both lower and slightly higher risk. It is not evidence of an exactly zero effect. The absolute difference also depends on the underlying risk in the comparison group.
Understand weights, pooling and heterogeneity
Marker area may represent statistical weight. State the weighting rule; a large marker does not automatically mean a study has better methods. A precise study can still be biased. A pooled diamond usually shows a summary estimate and its confidence interval, but a pooled result is justified only when the studies address a coherent comparison.
Heterogeneity refers to variation among study effects beyond what a chosen common effect model attributes to sampling error. A model with random effects includes a distribution of underlying effects, and its mean interval differs from a prediction interval for an effect in a comparable new setting. Estimates of heterogeneity can be uncertain, especially with few studies. Cochrane's handbook discusses these distinctions and cautions against pooling before considering the review question, design and risk of bias. S21
Two terms commonly accompany the plot. Tau squared, written τ², represents estimated variance among underlying study effects in the model. I squared, written I², describes the proportion of observed variability attributed to heterogeneity rather than sampling variation in a conventional formulation. I² is not the percentage of studies that are biased or the probability that the review is wrong.
Avoid counting how many studies cross a p value threshold. Different sample sizes can give different p values for similar effect estimates. Inspect effect direction, magnitude, uncertainty, design and risk of bias. If a review cannot justify a pooled estimate, a forest style display of individual effects can still support a structured synthesis without a diamond.
Survival curves require an event and a time origin
Survival analysis concerns time until a defined event. The event might be device failure, school withdrawal, completion of a task or a biological endpoint. “Survival” need not mean life or death. Define when timing begins, what counts as an event and how observation ends.
A Kaplan Meier estimate combines the conditional proportions remaining event free at successive event times. The risk set is the group still observed and event free immediately before a time. Right censoring means the event has not been observed by the last known observation time; the future event time is unknown. NIST describes this estimator in the context of reliability data. S34
Censoring is not a recorded absence of the event forever. A device still working when the study ends contributes information until then, but does not justify claiming that it will never fail. Similarly, a participant lost to observation has an unknown subsequent outcome.
Calculate the first steps of the supplied example
The synthetic event time dataset contains eight units. Events occur on days 2, 4, 5, 7 and 9. Censoring occurs on days 3, 6 and 8. There are no tied times in this example.
At day 2, all eight units are at risk and one experiences the event. The estimated probability of remaining event free becomes 7 divided by 8, or 0.875. At day 3, one unit is censored. The curve does not fall, but that unit leaves later risk sets.
Immediately before day 4, six units remain at risk. One event occurs, so multiply the earlier estimate by 5 divided by 6: 0.875 × 5/6 = 0.7292 after rounding. At day 5, five units are at risk and one event occurs; the estimate becomes 0.7292 × 4/5, approximately 0.5833 using unrounded intermediate values.
This is why dividing the total number of observed events by eight does not reproduce the curve at every time. The calculation uses the changing risk set. At the final event on day 9, only one unit remains at risk and the estimated curve reaches zero. That dramatic last step is based on very little information and should not be interpreted as a precise statement about a wider population.
Assess censoring and competing events
The usual interpretation of a Kaplan Meier curve requires appropriate assumptions about censoring. In simplified terms, censored units should not have systematically different subsequent event prospects from comparable units still observed, conditional on the relevant structure of the analysis. This cannot generally be established by looking at the curve alone.
If devices are removed because technicians notice signs of imminent failure, treating their removal as ordinary noninformative censoring can distort reliability estimates. If students with worsening circumstances are more likely to leave a follow up study, their unknown outcomes may differ from those who remain. Report reasons for loss and examine sensitivity where possible.
A competing event prevents the event of interest from occurring or being observed in the defined way. In an education study, transferring to another institution may need a different treatment from simply losing contact, depending on the estimand. For some questions, cumulative incidence methods are more appropriate than treating competing events as ordinary censoring. Define the scientific question before choosing the curve. The survival package vignette explains why cumulative incidence and a naive curve that censors competing events answer different questions. S409
Show censoring marks, meaningful time units, confidence intervals when appropriate and numbers at risk at useful times. Sparse tails should be easy to recognise. Do not extend a curve beyond supported follow up simply to fill the plotting area.
Explain comparisons without relying on a single p value
A log rank test and a Cox regression address particular questions about event time distributions or hazards. A Cox model commonly assumes proportional hazards for its usual interpretation. If curves cross or relative hazards vary over time, a single constant hazard ratio may be inadequate. Model diagnostics and alternative estimands may be necessary.
An event free probability at a specified time, a difference in restricted mean event free time or another justified measure can sometimes answer the substantive question more directly. Choose the analysis in consultation with the design and available information rather than attaching whichever p value a plotting function provides.
The same discipline applies to all three displays. PCA requires an account of preprocessing and omitted dimensions. Forest plots require effect definitions, study comparability and uncertainty. Survival curves require event definitions, risk sets and censoring assumptions. The legend should summarise these essentials while Methods provides the full analysis.
Practice and model answers
Task: A PCA plot shows two coloured groups with little overlap. Can the paper claim that two biological subtypes have been discovered?
Model answer: Not from the plot alone. Check feature selection, preprocessing, batches and other explanations, then seek appropriate independent validation. Describe the visible pattern without turning it into a confirmed category.
Task: A forest plot's pooled interval excludes the null but several study intervals do not. Is this automatically a contradiction?
Model answer: No. The pooled estimate and individual estimates have different uncertainty. Assess whether pooling was appropriate, how heterogeneous the effects are and whether bias affects the evidence. Do not use the number of individually significant studies as the deciding rule.
Task: A Kaplan Meier curve remains flat after several participants are lost. Does this show that those participants remained event free?
Model answer: No. The flat segment means no event was observed among the relevant risk sets during that interval. Censored participants' later event times are unknown. The inference depends on the censoring assumptions and remaining information.
Sources
Downloads
- Synthetic teaching data: synthetic_PCA_inputs.csv (3 KB)
- Synthetic teaching data: synthetic_survival.csv (1 KB)
- Calculation record: verified_calculations.json (1 KB)
- Python script that reproduces the examples: reproduce_examples.py (3 KB)
- Calculation record: figure_tutorial_calculations.json (1 KB)
- Calculation record: computation_environment.json (1 KB)
- About the teaching data: README.txt (8 KB)
Sign in to download these files.