Practice guideline, Colleague, Google, Pubmed, Dynamed
Practice guideline: vested interest
Colleague: unsystematic observation
Google: unfiltered information
Pubmed: difficult to locate most valid and up-to-date studies
Dynamed: most valid and up-to-date evidence
Evidence based medicine
- Conscientious, explicit and judicious use of ***current best practice in making decisions about care of individuals
- Integration of best **research evidence with our **clinical expertise and our ***patient’s unique values and circumstances
Why:
1. Best available evidence and tailor interventions to patient
—> Improve effectiveness of patient care
2. Evidence may be weak, contradictory, incomplete / vested interests
—> need to acquire most relevant evidence and appraise its quality
5A approach to EBP
Assess clinical scenario
Ask PICO question
- patient/population group
- intervention
- comparison group
- outcome
Acquire the best evidence for the question
Appraise the validity of evidence (external / internal)
Apply with patienyts’ unique values and circumstances
Best research evidence
- Relevant (patient-oriented outcomes)
- Valid (critical appraisal)
- Up-to-date
Attributes of causality
- Part of scientific theory (can be tested / form hypothesis)
- Reversible (no cause no outcome)
- Consistent (same effect across settings)
—> associations are not always causal —> therefore can only make inferences
Variables
- Qualitative (categorical)
- Binary (dichotomous)
- Nominal (unordered: several distinct categories cannot be ordered e.g. ethnicity)
- Ordinal (ordered: several distinct categories that can be ordered e.g. smoking) - Quantitative (numerical)
- Discrete (integers)
- Continuous (within a range)
Both qualitative and quantitative data can be presented in frequency distribution
- Summarise the data
—> identify outliers, reveal possible errors - Visualised using graphical display
—> Qualitative: pie chart, bar chart
—> Quantitative: Histogram (frequency polygon), Box-and-whisker plot
Central tendency
Mode:
- not particular good indicator of central tendency
- only means to measure central tendency when data is NOMINAL values
Median:
- literal measure of central tendency
- LESS SENSITIVE to outliers
Mean:
- arithmetic mean
- SENSITIVE to outliers (extreme values)
Dispersion
- Range (difference between largest and smallest)
- VERY SENSITIVE to outliers - Quartile
- Inter-quartile range (difference between first and third quartile)
- summarise variation of data
- can be estimated from cumulative frequency curve - Percentile
- give information on spread
- can be estimated from cumulative frequency curve - Standard deviation
- how dispersed the data are
- average difference between mean and dat value
- square root of variance
Sample variance vs Population variance
Sample variance:
- sum of squared difference / (number of values - 1)
Population variance:
- sum of squared difference / number of values
Sampling
Different samples —> Sampling variation
Sample many times —> Sampling distribution
More samples we draw —> mean of sampling distribution closer to population mean
Effect measures
Judging whether exposure causes outcome - strength of possible association
- Risk ratio / Relative risk: how much more likely exposed will become cases
- RATIO of probabilities
* **- incidence of outcome in exposed / incidence of outcome in unexposed
- a/(a+b) / c/(c+d)
- RR < 1 —> Protective factor, lower risk
- RR = 1 —> no evidence of association between exposure and outcome
- RR > 1 —> Risk factor, higher risk
- further from 1, stronger the association - Odds ratio: odds of outcome in exposed to odds of outcome in unexposed
- Odds is not probability
- but RATIO of probability of getting to probability of not getting
- odds = 兩個ratio相除
* **- odds of outcome in exposed / odds of outcome in unexposed
- a/(a+b) / b/(a+b) / c/(c+d) / d/(c+d)
- OR < 1 —> exposed group less likely to have the outcome, lower odds of outcome
- OR = 1 —> No association between exposure and outcome
- OR > 1 —> exposed group more likely to have the outcome, higher odds of outcome
* **- OR is always further away from 1 than corresponding RR, except for rare outcome (incidence is very low, RR / OR approximately equal)
In Case-control studies: only OR can be calculated because no incidence (prevalence is inflated by selecting cases)
New cases involved time: use RR
In general: RR are better since take into account incidence rate + represent likelihood of outcome
However, in case-control study, only OR can be calculated
Prevalence vs Incidence
Prevalence: for Cross-sectional studies
Incidence: for Cohort studies
Prevalence: no. of EXISTING cases at a designated time
- depends on: incidence, duration (chronic / acute)
- proportion, not rate
—> Point prevalence: at a time point
—> Period prevalence: during specified time period
Incidence: frequency of OCCURRENCE of NEW cases in a given time period
- measure of risk
- not directly measurable unless population is followed over time
- frequency count / proportion / rate
—> Cumulative incidence (incidence proportion): new cases/population at start period
—> Incidence rate (incidence density): new cases/total person time at risk
Reliability vs Validity
Reliability: produce same results if repeated
Validity: extent to which measures true value of variable
Hierarchy of evidence
Evidence summaries > Systematic reviews / Meta-analyses > individual studies
- Randomised Controlled Trial (control intrinsic and confounding factors)
- Cohort studies
- Case-control studies / cross-sectional / ecological
- Expert opinion
- Pathophysiologic reasoning
RCT not always ethical / feasible
—> then use observational studies
—> use guides to assess causation:
E.g. Bradford Hill’s criteria, Koch’s postulate
Internal validity: Can we trust the results?
Null hypothesis: (Default)
- a variable has NO association with another variable / 2 population distributions do not differ from each another
Alternative hypothesis:
- a variable has an association with another variable / 2 populations differ from each other
P-value:
- probability that a result that is extreme from observed result is obtained, assuming that the null hypothesis is true
- probability that a result is affected by random error, assuming null hypothesis is true
- smaller p-value —> stronger confidence that we are correctly rejecting null hypothesis
- p < 0.05: observed results is unlikely due to chance, statistically significant
Power:
- probability of correctly rejecting the null hypothesis, given that null hypothesis is wrong
- (1-β)
- ability of a study to demonstrate an association
- Good: as high as possible (80%)
Confidence interval
More informative than p-value:
- gives range of plausible values for true value
(E.g. 95% confident that true value lies within specific range)
- tell precision of estimate width decreases with increasing sample size
- suggest whether an association exist (if the confidence interval crosses the null e.g. risk ratio = 1)
Random errors
Random error (chance): (lack of reliability —> cannot produce same results if repeated)
- Type 1 (α error / p-value): rejecting null hypothesis when it is true (false positive)
- probability of committing Type 1 error: p-value - Type 2 (β error / 1-power): accepting null hypothesis when it is false (false negative)
- probability of avoiding Type 2 error: power
When try to reduce Type 2 error (increase power) —> Type 1 error will increase consequently, to keep both errors to acceptable level:
- increase effect size —> large effect size —> small p-value
- increase sample size —> smaller sampling variation —> smaller p-value
- standard procedures
- regularly calibrate
Systematic errors
consistently wrong results due to problems with study design (lack of validity: cannot measure true value)
- must be identified early since cannot controlled for in the analysis
- reduce systematic error
—> appropriate instrument
—> reduce biological variations
- Information bias
—> Observer bias (by investigators)
—> Recall bias (by subjects, particular problem if public is aware of exposure-disease association)
- reduce bias:
—> Blind assessment of outcome status (investigator blinded)
—> Prospective assesssment of outcome (no need to recall)
—> Use objective measures - Selection bias:
- systematic difference between subjects and target population
- subjects not representative of target population —> affects EXTERNAL validity
- subjects groups are not comparable —> affects INTERNAL validity
- reduce bias:
—> random sample
—> maximise response rate
—> minimise attrition rate
—> select case and control from same underlying population
***Confounding
Type of bias
—> Distortion of observed association due to other factors that are common causes of both exposure and outcome (confounders)
—> leading to estimated association is not the true causal effect
Positive confounders: overestimate associations
Negative confounders: underestimate associations
Confounders:
- must cause the exposure
- must cause the outcome
- must not be on causal pathway between exposure and outcome (confounders cannot be result of exposure), otherwise called mediator
Reduce confounding:
- Randomisation (randomly distribute confounders, only possible when you assign people to groups i.e. interventional studies)
- Restriction (not allow certain confounders to be present)
- Matching (equal representation of subjects with certain confounders between study group —> may introduce selection bias)
- Stratifying (into sub-samples according to specified confounder during analysis)
- Multivariable analysis (e.g. regression analysis: adjust for a number of confounders simultaneously)
- Quantitative analysis (multiple linear regression)
- Time to occurrence (cox regression)
Reporting bias
Dissemination of research findings is influenced by nature of results
—> distort results of systematic reviews
—> biased / incomplete picture
- Publication bias (positive results more likely to be published)
- Time lag bias (positive results published sooner)
- Language bias
- Citation bias
- Selective outcome reporting bias (secondary outcome more favourable —> report as major outcome want to look for in the beginning)
Reduce:
- Register clinical trials in public registries before carrying out
—> Ensure transparency and subject to public scrutiny
—> Pre-specified protocol
—> Pre-specified outcomes
- Fully report all methods and results including negative results
- Share individual patient data so other researchers can replicate
Epidemiology study design
Observational (without intervention)
- Group
—> descriptive (Disease mapping)
—> analytical (Ecological studies)
- Individual
—> descriptive (Cross-sectional + Cohort)
—> analytical (Cross-sectional + Case-control + Cohort)
Interventional (RCT and other experiments)
Observational group descriptive
Disease mapping (Routine data, Surveillance data):
- Comparison of health indicators across populations
- Monitoring of time trends
- Evaluate and planning public health programmes
Observational group analytical
Ecological studies
- analyse associations of exposure with outcome
- provide preliminary evidence for association of interest
- examine possible exposure-disease relationship on population level
- ECOLOGICAL FALLACY (以全概偏)
—> inferences about individual risks can be erroneous
