22.3 Evidence-Based Dentistry & Biostatistics
Key Takeaways
- Evidence-based dentistry integrates best research evidence, clinical expertise, and patient values/preferences—not papers alone and not anecdote alone.
- Sensitivity detects disease when present; specificity excludes disease when absent; PPV and NPV depend on prevalence as well as test accuracy.
- Study designs rank from case reports through case–control and cohort studies to randomized controlled trials and systematic reviews/meta-analyses for interventional questions.
- Bias (selection, information, confounding, recall, publication) systematically distorts results; random error is addressed with sample size and statistics.
- P-values and confidence intervals require careful interpretation: statistical significance is not the same as clinical importance or proof of no bias.
22.3 Evidence-Based Dentistry & Biostatistics
Quick Answer: EBD = best evidence + clinical expertise + patient values. Frame questions with PICO, then match study design to the question. For tests, know sensitivity, specificity, PPV, NPV. For results, distinguish bias from chance, and never confuse a small p-value with automatic clinical importance.
This section is pure AFK reasoning skill: many stems describe a study or a diagnostic test and ask what it measures, what it cannot prove, or which design is strongest. Prevention claims from 22.1–22.2 are ideal material for appraisal practice.
What Evidence-Based Dentistry Is (and Is Not)
| Component | Role |
|---|---|
| Best available research evidence | Systematic, preferably patient-important outcomes |
| Clinical expertise | Diagnosis skill, technical ability, judgment of applicability |
| Patient values and circumstances | Preferences, culture, costs, access, risk tolerance |
Not EBD: “I always do it this way,” or “one viral post said,” or “the paper’s p < 0.05 so every patient must receive the product.”
Also not EBD: ignoring patient goals to follow a guideline blindly when comorbidities or values differ—guidelines inform, they do not replace consent.
The 5 A’s cycle (practical)
- Ask a focused clinical question.
- Acquire evidence (guidelines, systematic reviews, primary studies).
- Appraise validity, importance, applicability.
- Apply with the patient.
- Assess outcomes and improve.
PICO (and PICOT)
| Letter | Meaning | Example |
|---|---|---|
| P | Patient / Population / Problem | High-caries-risk children aged 6–8 |
| I | Intervention (or exposure) | School sealant program |
| C | Comparison | No sealant / usual care |
| O | Outcome | New DMFS over 24 months |
| T | Time (optional) | 24 months |
Diagnostic variant: P = patients suspected of disease; I = index test; C = reference standard; O = accuracy metrics.
Hierarchy of Evidence (Interventional Questions)
Higher designs better control bias when well conducted—quality still matters.
| Level (typical teaching ladder) | Design | Strength for therapy questions |
|---|---|---|
| Highest | Systematic review / meta-analysis of RCTs | Synthesizes multiple trials |
| High | Randomized controlled trial (RCT) | Randomization balances known/unknown confounders |
| Moderate | Cohort study (prospective/retrospective) | Good for prognosis, harms, when RCT unethical/impractical |
| Lower | Case–control study | Efficient for rare outcomes; recall/selection issues |
| Lower | Cross-sectional study | Prevalence and associations at one time—weak for causality |
| Lowest | Case series / case report | Hypothesis generating |
| Foundation | Expert opinion / pathophysiology | Useful but weakest alone |
Guidelines may be high impact clinically but quality varies—check whether recommendations are evidence-linked.
Matching design to question
| Question type | Preferred evidence |
|---|---|
| Therapy / prevention effectiveness | RCT → systematic review of RCTs |
| Harm / etiology | Cohort; case–control for rare harms |
| Diagnosis | Cross-sectional accuracy studies vs reference standard |
| Prognosis | Inception cohort |
| Prevalence | Cross-sectional survey |
AFK trap: a large case series of implant success without controls “proves” superiority of a brand—it does not.
Diagnostic Test Metrics
2×2 structure
| Disease present | Disease absent | |
|---|---|---|
| Test positive | True positive (TP) | False positive (FP) |
| Test negative | False negative (FN) | True negative (TN) |
Definitions and formulas
| Metric | Formula | Plain language | When you care most |
|---|---|---|---|
| Sensitivity | TP / (TP + FN) | Of those with disease, % testing positive | Rule out when high (SnNOut): Negative test → less likely disease |
| Specificity | TN / (TN + FP) | Of those without disease, % testing negative | Rule in when high (SpPIn): Positive test → more likely disease |
| PPV | TP / (TP + FP) | Of positive tests, % truly diseased | Depends on prevalence—higher prevalence → higher PPV |
| NPV | TN / (TN + FN) | Of negative tests, % truly non-diseased | Higher when disease is rare (often) |
| Accuracy | (TP+TN)/all | Overall correct | Can mislead with imbalanced prevalence |
| Likelihood ratios | Combine sensitivity/specificity | How much a result shifts odds | Advanced but elegant |
Prevalence dependence—critical AFK concept
Sensitivity and specificity are often treated as properties of the test (in a defined population and threshold). PPV and NPV change with prevalence:
| Scenario | Effect |
|---|---|
| Screening low-risk population with imperfect specificity | Many false positives → PPV drops |
| Testing high-risk clinic population | PPV rises for the same test |
| Very rare disease | Even specific tests can yield low PPV |
Example narrative: A pulp vitality test with good sensitivity still needs clinical context; a positive “caries detection device” in a sealed fissure may be FP noise.
ROC thinking (recognition)
Changing the positivity threshold trades sensitivity vs specificity. ROC curves plot that tradeoff; area under curve summarizes discrimination.
Core Biostatistics for AFK
Types of data
| Type | Examples | Typical stats |
|---|---|---|
| Continuous | Pocket depth mm, DMFS count | Means, t-tests, regression |
| Binary | Caries yes/no | Proportions, chi-square, logistic regression |
| Ordinal | Pain scale 0–10 categories | Nonparametric or ordinal models |
| Survival | Time to implant failure | Kaplan–Meier, Cox models |
Central tendency and spread
| Measure | Use |
|---|---|
| Mean | Symmetric continuous data |
| Median | Skewed data (income, some cost data) |
| SD | Spread around mean |
| IQR | Spread around median |
| 95% CI | Interval likely to contain the true parameter with 95% confidence under model assumptions |
Hypothesis testing language
| Concept | Meaning |
|---|---|
| Null hypothesis (H₀) | Typically “no difference / no association” |
| p-value | Probability of data as extreme as observed if H₀ were true—not the probability H₀ is true |
| α (alpha) | Pre-set type I error rate (often 0.05) |
| Type I error | False positive: reject true H₀ |
| Type II error | False negative: fail to reject false H₀ |
| Power | 1 − β; probability of detecting a true effect |
| Clinical vs statistical significance | Tiny irrelevant difference can be “significant” in huge samples; important effects can be nonsignificant if underpowered |
Confidence interval pearl: a mean difference of 0.1 mm CAL with 95% CI 0.05–0.15 may be statistically significant but clinically trivial; a large effect with wide CI crossing the null is uncertain.
Relative vs absolute measures
| Measure | Idea | Trap |
|---|---|---|
| Relative risk (RR) | Risk_intervention / Risk_control | “50% reduction” can hide small absolute change |
| Odds ratio (OR) | Odds_exposed / Odds_unexposed | ≈ RR when outcomes rare; otherwise not identical |
| Absolute risk reduction (ARR) | Risk_control − Risk_intervention | More transparent for patients |
| NNT | 1 / ARR | Patients needed to treat for one extra good outcome |
| NNH | For harms | Number for one extra harm |
Prevention example: If sealants reduce 2-year occlusal caries from 10% to 5%, RR = 0.5 (50% relative reduction), ARR = 5%, NNT = 20.
Bias and Validity
Internal vs external validity
| Validity | Question |
|---|---|
| Internal | Are the study’s conclusions correct for the participants studied? (bias/confounding minimized) |
| External (generalizability) | Do results apply to my patient / my setting? |
Major bias types
| Bias | What goes wrong | Dental example |
|---|---|---|
| Selection bias | Who enters or remains in the study is non-comparable | Clinic volunteers for a “new whitening gel” are healthier and more compliant |
| Information / measurement bias | Exposures or outcomes measured differently across groups | Examiner knows which group got sealant and scores caries more leniently |
| Recall bias | Cases remember exposures differently than controls | Oral cancer cases over-report alcohol vs controls |
| Interviewer bias | Questioning differs by case status | |
| Lead-time bias | Early detection looks like longer survival without changing course | Screening “improves survival” only by diagnosing earlier |
| Length-time bias | Screening finds slowly progressive disease preferentially | Overestimates benefit of screening |
| Publication bias | Positive studies publish more readily | Meta-analyses overstate effects |
| Confounding | Third factor associated with exposure and outcome distorts effect | Coffee “causes” oral cancer if smoking not controlled |
| Attrition bias | Differential loss to follow-up | Failures drop out of implant follow-up |
| Spectrum bias | Diagnostic study uses extreme sick vs healthy only | Overstates real-world accuracy |
Controlling confounding
| Method | Notes |
|---|---|
| Randomization | Balances confounders in RCTs |
| Restriction | Study only non-smokers |
| Matching | Pair on age/sex |
| Stratification / multivariable analysis | Statistical adjustment—residual confounding possible |
Appraising a Prevention Claim (Worked Pattern)
Claim: “Product X toothpaste cuts cavities 40%.”
- PICO: Who? vs what? measured how?
- Design: RCT of children? industry-funded?
- Outcomes: DMFS increments? patient-important?
- Absolute numbers: 40% of what baseline? NNT?
- Bias: blinding of examiners? attrition?
- Applicability: fluoride background, diet, Canadian setting?
- Harms/cost: fluorosis, price, adherence?
- Patient values: taste, cost, preferences.
Ethics of Evidence Use
- Do not overstate certainty to sell procedures.
- Disclose material alternatives (including monitoring/prevention).
- Research participation needs consent and REB oversight (jurisprudence crossover).
- Industry relationships require transparent appraisal.
Rapid review list
- EBD triad: evidence + expertise + patient values
- PICO structures searchable questions
- Therapy: prefer well-done RCTs and systematic reviews
- Sensitivity/specificity vs PPV/NPV (prevalence matters for PPV/NPV)
- SnNOut / SpPIn mnemonics
- p-value ≠ probability the null is true; ≠ clinical importance
- ARR and NNT communicate absolute benefit
- Selection, information, confounding, recall, publication, lead-time, length-time biases
- Internal validity first, then external applicability
Mastery of Chapters 22.1–22.3 means you can deliver fluoride and sealants correctly, speak the language of population oral health and DMFT, and critically appraise the evidence behind any prevention or treatment claim—the full prevention and EBD slice of the AFK blueprint.
Evidence-based dentistry is best defined as the integration of:
A new caries-detection device is evaluated: of 100 teeth truly with dentin caries, 90 test positive; of 100 truly sound teeth, 80 test negative. Sensitivity and specificity are:
Which study design generally provides the strongest evidence that a preventive intervention causes a reduction in new caries, assuming high methodological quality?
Positive predictive value (PPV) of a diagnostic test is most accurately described as: