Biostatistics and Research Methods for Ophthalmologists
Last updated: October 2026
Key points in 60 seconds
Sensitivity = TP/(TP+FN); a sensitive test, if negative, rules out (SnNout).
Specificity = TN/(TN+FP); a specific test, if positive, rules in (SpPin).
Predictive values depend on prevalence: screening in low-prevalence populations gives many false positives.
Know study designs, common biases, trial terms and agreement statistics for exams and journal club.
The statistics, study design and research ethics that appear in every ophthalmology exam and every journal club, explained with eye examples. Includes a calculator for diagnostic test accuracy.
Diagnostic tests: the 2 × 2 table
Disease present
Disease absent
Test positive
True positive (TP)
False positive (FP)
Test negative
False negative (FN)
True negative (TN)
Sensitivity = TP ÷ (TP + FN): of those with disease, how many test positive. A sensitive test, if negative, rules disease out (SnNout).
Specificity = TN ÷ (TN + FP): of those without disease, how many test negative. A specific test, if positive, rules disease in (SpPin).
Positive predictive value = TP ÷ (TP + FP) and negative predictive value = TN ÷ (TN + FN). These depend on prevalence: screening for glaucoma in the general population (low prevalence) gives many false positives even with a good test.
Likelihood ratios: LR+ = sensitivity ÷ (1 − specificity); LR− = (1 − sensitivity) ÷ specificity. Unlike predictive values, they do not change with prevalence.
The default numbers show a test with 90% sensitivity and 90% specificity used where prevalence is 2%: the positive predictive value is only about 16%.
Measures of effect
Measure
Formula
Example (event in 20% of controls, 15% of treated)
Absolute risk reduction (ARR)
Control risk − treated risk
20% − 15% = 5%
Relative risk (RR)
Treated risk ÷ control risk
0.15 ÷ 0.20 = 0.75
Relative risk reduction (RRR)
1 − RR
25%
Number needed to treat (NNT)
1 ÷ ARR
1 ÷ 0.05 = 20
Odds ratio (OR)
(a × d) ÷ (b × c) from a 2 × 2 table
Used in case–control studies; close to RR when the outcome is rare
Confidence interval: the range of values compatible with the data. If a 95% CI for an RR or OR includes 1 (or for a difference includes 0), the result is not statistically significant at p < 0.05.
p-value: the probability of a result at least this extreme if there were truly no difference. It does not measure the size or importance of an effect.
Type I error (α): finding a difference that is not real (false positive). Type II error (β): missing a real difference. Power = 1 − β, usually set at 80–90%; it rises with sample size and effect size.
Study designs and the evidence pyramid
Design
What it does
Eye example
Main weakness
Systematic review and meta-analysis
Combines all good trials
Cochrane review of anti-VEGF for DME
Only as good as the included trials; publication bias (funnel plot asymmetry)
Randomised controlled trial
Random allocation removes confounding
LiGHT (SLT vs drops), Protocol T
Cost, strict selection limits generalisability
Cohort study
Follows exposed and unexposed people forward
Smokers vs non-smokers for cataract
Confounding, loss to follow-up
Case–control study
Compares cases with controls, looking back at exposures
Uveal melanoma and sunlight
Recall and selection bias
Cross-sectional study
Measures prevalence at one time point
Prevalence of glaucoma in a district
Cannot show cause and effect
Case series and case reports
Describe new findings
First reports of a drug side effect
No comparison group
Common biases
Selection bias: the study population does not represent the target population (for example, tertiary hospital patients used to estimate community prevalence).
Lead-time bias: earlier diagnosis by screening makes survival look longer even if outcome is unchanged. Length-time bias: screening picks up slower, milder disease.
Recall bias (case–control studies), observer bias (reduced by masking), attrition bias (dropouts), and publication bias.
Confounding: a third factor linked to both exposure and outcome (for example, age in studies of diabetes and cataract). Randomisation, matching and multivariable analysis reduce it.
Trials: key terms
Intention-to-treat: analyse people in the groups they were randomised to, whatever happened; preserves randomisation. Per-protocol: only those who completed treatment as planned.
Masking (blinding): single (patient), double (patient and assessor). Non-inferiority trials test whether a new treatment is not worse than the standard by more than a set margin (for example faricimab vs aflibercept).
Survival analysis: Kaplan–Meier curves for time-to-event data such as graft survival or trabeculectomy failure, compared with the log-rank test; hazard ratios from Cox regression.
Agreement and reliability
Bland–Altman plot: agreement between two measurement methods (for example two tonometers): plots the difference against the mean, with the mean bias and 95% limits of agreement. Correlation is not agreement.
Kappa statistic: agreement between observers beyond chance for categories (for example grading of diabetic retinopathy): above 0.8 is very good, 0.6–0.8 good.
Intraclass correlation: reliability of continuous measurements.
Research ethics and consent
The Declaration of Helsinki sets the principles for research involving people; studies need ethics committee approval and informed consent. In India, research follows the ICMR National Ethical Guidelines and trials are registered with CTRI.
Consent for treatment: patients must be told about material risks, benefits and reasonable alternatives (UK: Montgomery 2015). Record the discussion, not just a signature.
Confidentiality can be broken only in limited situations, such as serious risk to others (for example an unsafe driver who refuses to stop).
Patel D. Biostatistics and Research Methods for Ophthalmologists. RPC Rounds (Sadbhaav Eye Clinic). Updated October 2026. Available at: https://www.sadbhaavclinic.com/biostatistics-for-ophthalmologists.html. Accessed [date].
Written and medically reviewed by Dr. Dhaval Patel, MD (Ophthalmology, AIIMS New Delhi) · Editorial policy