AFP Critical Appraisal & Stats Flashcards
How will you summarise what the paper is about?
Population = People with AF Intervention = Apixaban 5mg BD Control = Warfarin INR 2-3 Outcomes = Stroke or systemic embolism Key findings
What should you think about before summarising the paper?
QR
What is the research q and what is its relevance to clinical practise
What acronym do you use to assess internal validity?
Recruitment Allocation Maintenance Baseline Outcome (and was it blinded) Stats
What is internal validity?
The extent to which the observed results represent the truth in the population we are studying and, thus, are not due to methodological errors
How can we reduce selection bias?
Use consecutive recruitment rather than non-consecutive
Consider the recruitment location (e.g. GP vs hospital)
What types of recruitment are there?
Consecutive vs non-consecutive
Single centre vs multicentre
Types of allocation?
Randomised vs non randomised
Allocation blinded vs open label
Letters/Packets vs automated voice recognition system (electronic)
Block randomisation vs whole group randomisation
Cluster randomisation vs whole group randomisation
What is block randomisation?
Wait for X number of study participants and then say 50% go to each arm
Throughout the whole study there are equal numbers in both arms
Helps to mitigate temporal trends
What is cluster randomisation? Disadvantages?
Used in multicentre studies
Used when you can’t deliver different interventions in one place (e.g. poster in a GP practice)
Needs higher numbers of people
More difficult to achieve balance across groups
Benefit of randomisation?
Reduces risk of confounders
Why does blinding help?
Reduces risk of placebo effect
Reduces bias in data analysis
Benefit of Electronic voice recognition system?
Minimizes risk of tampering
What is study maintenance and what is usually aimed for?
Maintenance is the drop out rate from a study
Usually aim for <20% but ideally <10%
What bias is created if there is a poor maintenance?
Attrition bias - causes study to be under powered
Types of blinded outcome?
Single: patient only
Double: patient and physicians interacting
Triple: patient, physicians, outcome extractors
How can you blind if the outcome can’t be blinded e.g. surgical?
Use PROBE design
Prospective Randomised Open label Blinded endpoint adjudication
Whoever reviews the outcomes doesn’t know which arm of study the patient is in (e.g. radiologist when looking at stroke tx)
What do you need to look at in baseline?
Baseline characteristics of population and are they matched?
Besides blinding what else do you need to consider with outcome?
Is the choice of outcome important, adequate and important to patients?
Define power
the likelihood of detecting a difference when it exists (avoiding Type 2 error)
What 4 things should we look at when analysing the stats of a paper?
Power calculation and according recruitment target
Statistical models used for outcome measure (binary, continuous, time-dependent)
Effect size and significance level
Absolute risk reduction / Relative risk reduction / NNT
What 3 things does power calculation depend on?
- How many events are expected over follow-up time (incidence x time) - increase sample size
- Expected improvement (RRR of 50%) - increase effect size
- What probability you want to detect the difference when it exists (80%?) - increase precision of measurement
If a study is negative then think about the power!
Define p value
probability that the association detected has arisen by chance
How to strengthen the p value?
Size of association
Size of cohort
What test do you use to calculate difference between 2 groups with categorical variables when not adjusted? E.g. stroke vs no stroke
Chi-squared
What test do you use to calculate difference between 2 groups with categorical variables when adjusted? E.g AF X Stroke
Binary logistic regression
What test do you use for comparing 2 groups with a continuous outcome e.g. Gender x BMI if normally distributed?
T-test
What test do you use for comparing 2 groups with a continuous outcome e.g. Gender x BMI if not normally distributed?
Wilcoxon ranked test (dependent)/Mann Whitney U (independent)
What test do you use for comparing >2 groups with a continuous outcome e.g. number of children x BMI if adjusted?
ANOVA
What test do you use for comparing >2 groups with a continuous outcome e.g. number of children x BMI if not adjusted?
ANCOVA
What test do you use to find a correlation between 2 parametric continuous variables?
Pearson’s rank
What test do you use to find a correlation between 2 non-parametric continuous variables?
Spearman’s rank
What test to use for time to event data i.e. when time counts (e.g. mortality in cancer tx) to determine a direct association?
Kaplan Meier analysis
What test to use for time to event data i.e. when time counts (e.g. mortality in cancer tx) when adjusted for covariates?
Cox proportional hazards model
How to assess external validity?
Resources - equipment/cost
Population - can you generalise the demographics
What are the last 3 things we consider?
Funding - any conflicts of interest
Ethics - clinical equipoise, safety outcomes, data safety monitoring board
Conclusion
Define 95% Confidence interval
A range, between which the population mean value will lie 95% of the time
Define relative risk
Risk of developing disease in the exposed group compared to the risk of developing disease in the unexposed group
How do you calculate relative risk?
Those who got disease in exposed group/all exposed divided by those who got disease in non-exposed group/all non-exposed
Which studies is relative risk used in?
Prospective cohort
Define odds ratio?
Odds of something happening vs the odds of it not happening
How do you calculate odds ratio?
Exposed with disease/not exposed with disease divided by exposed without disease/not exposed without disease
AKA odds in disease group/odds in control
What studies is odds ratio used in?
Retrospective observational study (case-control study)
Define hazard ratio
Used to look at survival over time with a Kaplan Meier curve - equivalent to relative risk but risk is not constant with respect to time
What is a hazard ratio used in?
Prospective RCT
Define incidence rate
Number of new cases over a defined period of time
Define prevalence
Number of cases in a given population at any given time
Define Absolute risk reduction (ARR) and how to calculate it
Difference in the incidence of disease between two groups
= P(event occurring in group 1) – P(event occurring in group 2)
Define NNT and how to calculate it
Number of patients who need to be treated to prevent one event occurring
= 1/ARR
How to calculate RRR?
= ARR / P(event occurring in control group)
Express as %
Define Type I error
Probability of rejecting H0 when in fact H0 is true
i.e. probability of concluding there is a significant difference when actually there’s no difference.
False positive
Define Type II error
Probability of concluding H0 is true when in fact it is FALSE
i.e. false negative
How do we set a type II error?
β is the probability of making type II error – usually set at 0.8
How do we set a type I error?
Reflects p value cut-off i.e. usually = 0.05.
What does per protocol analysis mean?
What is the benefit?
What is the con?
Only participants who complied with study protocol completely are included in analysis.
More accurate representation of treatment effect
Susceptible to attrition bias and exclusion bias
What does intention to treat analysis involve?
What is the benefit?
What is the con?
All participants who have been randomised are included, regardless if they took the medicine / completed the study
More accurate representation of effect in clinical practice
Can’t determine what the drug does in optimal conditions
What are case control studies used for?
Identifying associations between exposure and outcome
Advantages of case control studies?
Rare diseases Quicker and cheaper to perform Can analyse multiple exposures at once Can calculate OR Good for dynamic populations where follow up is difficult
Disadvantages of case control studies?
Rely on quality of records
Selection bias and recall bias
Can’t demonstrate temporal association
What is a cohort study used for?
Evidence for causation and temporal association
Advantages of cohort studies?
Assess temporality Assess prognosis Can control data collection and quality Can calculate incidence – allows you to calculate RR, risk difference, NNT Good for common exposures
Disadvantages of cohort studies?
Expensive Take longer Not good for rare diseases Loss to follow-up and potential of attrition bias Can be affected by confounders
Define bias
Factors which cause systematic over or under-estimate of a particular result
What is selection bias?
Study sample does not represent entire population
What is volunteer bias?
People who volunteer are different from population as a whole
What is channelling bias?
Healthcare professionals may subconsciously select patients who would be good for study
What is Interview bias?
If interviewer already knows disease status of participant
What is recall bias?
Participants remember things differently
What is Hawthorne bias?
Participants alter behaviour because they are aware of monitoring
What is a confounder?
A factor which has an independent effect on both the exposure and the outcome
How can we minimise confounders?
Stratification of participant groups
Regression analysis
Randomisation and matching
What is external validity?
Generalisability of results
Difference between RR and HR
RR looks at if an event has occurred during the study
HR looks at if an event occurred and when it occurred
How to describe RR of 1.45?
45% more likely to have outcome X
How to describe OR of 1.6?
Odds of exposure to factor X is 1.6x higher
How to describe HR of 0.79?
At any particular point, group A were 21% less likely to have outcome X
What is a type 1 error?
Reject H0 (no difference) when actually H0 is true I.e. False positive
What is a type 2 error?
False negative
Define sensitivity and how to calculate it
Probability of correctly identifying those with disease
TP/TP + FN
Define specificity and how to calculate it
Probability of correctly identifying those without disease
TN/TN + FP
What is a QALY? What is the accepted cost?
Quality adjusted life year
<20k per QALY – ACCEPT. 20-30 – think about
What is a fixed effect meta analysis?
Assumes that studies are homogenous
All measuring same treatment effect
Calculate weighted average - give more weight to bigger studies
What is a random effects meta?
Assumes heterogeneity between studies
What methods are used to quantify heterogeneity?
Q – statistic - low p = heterogenetic
I2 – estimates proportion of total variance that is attributable to heterogeneity
How is heterogeneity dealt with?
Sub group analysis – separate meta analyses
Meta regression - quantify how treatment effect changes with study characteristic
What analysis approaches may be performed in an RCT?
Intention to treat
Per protocol
As treated
What is a sensitivity analysis?
Assesses the impact of different assumptions and methodological choices on the results of the primary analysis
Can looks at differences in analysis approach, inclusion criteria, outcome definition, missing data
May also be good for missing data – can account for it using imputation
What groups are responsible for overseeing a trial?
Trial management group – day to day
Data monitoring committee – independent – make recommendations. Perform formal interim analysis – stop trial if efficacy clear or serious adverse events
Trial steering committee – uses recommendation of DMC
What are the issues with a formal interim analysis?
May stop trial prematurely due to random variation in treatment effect over time
What is PPV?
Likelihood of people who test positive actually having disease
TP/TP + FP
What is NPV?
Likelihood of people who test negative actually not having disease
TN/TN + FN
What is heterogeneity?
the degree of difference between methodology in the studies and thus the treatment effect being measured
What does asymmetry in a funnel plot (meta analysis) suggest and what does it lead to?
publication bias (only studies that get good results are published) Leads to overestimation of the treatment effect
How do you overcome publication bias?
Register trial beforehand
Publish a protocol
What is clinical equipoise?
a state of genuine uncertainty on the part of the clinical investigator regarding the comparative therapeutic merits of each arm in a trial
What is Ecological fallacy and give an example
Assuming that the results of a cross sectional study relate to an individual. Eg areas with higher salt consumption are more at risk of heart attacks. Does amount of salt consumption actually increase likelihood in an individual?
How can you graphically represent a meta analysis?
Forest plot
What is the difference between a retrospective cohort study and a case control study?
case control takes patients based on outcome status whereas retrospective cohort doesn’t
Disadvantage of retrospective cohort study?
Potential to find reverse causality
Recall and interviewer bias
Disadvantage of using 2:1 allocation?
Undermines clinical equipoise and can cause therapeutic mis-estimation
Reduces power or requires 12% more participants
Affects internal validity
Advantages of using 2:1 allocation?
In early phase trials to determine clinical utility
If a tx is particularly expensive
Need for additional safety info - good for adverse events
How to interpret a confidence interval looking at the difference between 2 groups?
If it contains the value 0 then p>0.05
How to interpret a confidence interval when comparing groups using a ratio instead of a difference e.g. odds/RR?
If it contains the value 1 then p>0.05
Difference between OR and RR?
odds ratio is a ratio of two odds whereas the relative risk is a ratio of two probabilities
Where can you get research funding from?
Non-commercial: Research councils - HRA, MRC Government NIHR Charities - CRUK
Commercial:
Pharma companies
Industry companies
Where do you apply for ethical approval?
Health Research Authority Research Ethics Committee
Which studies is odds ratio used in?
case control
Which stats test to measure strength of association between 2 variables? (ND and NND)
ND: Pearson correlation
NND: Spearman rank
Which descriptive stats for ND vs NND?
ND: Mean, SD
NND: Median, IQR
Which stats test for comparison of one group to hypothetical value? ND, NND and categorical
ND: one sample T-test
NND: Wilcoxon
Categorical: Chi square/binomial
Difference between one sample T-test and independent T-test? What is a paired T-test?
One sample compares a group to a pre-determined value
Independent compares a group to another group
Paired compares values within a group at different intervals
Stats test for comparing 2 unpaired groups if ND, NND or categorical?
ND: Unpaired T-test
NND: Mann Whitney U
Categorical: Fisher’s exact test or Chi-squared if large groups
Stats test for comparing 2 paired groups if ND, NND or categorical?
ND: Paired T-test
NND: Wilcoxon
Categorical: McNemar’s
Stats test for comparing 3+ unmatched groups if ND, NND or categorical?
ND: one way ANOVA
NND: Kruskal Wallis
Categorical: Chi squared
Stats test for comparing 3+ matched groups if ND, NND or categorical?
ND: Repeated measures ANOVA
NND: Friedman test
Categorical: Cochrane Q
What stats test to use for prediction?
Single variable: Linear/logistic regression
Multiple variables: Multiple linear/logistic regressions
What outcome does logistic regression provide?
Adjusted odds ratio
What to use when the denominator is number of person-years?
Relative risk
What is the advantage of using parametric tests over non-parametric?
Provides more statistical power
What is the positive likelihood ratio?
The probability that a person with the disease tested positive for the disease divided by the probability that a person without the disease tested positive for the disease
AKA True positive / False positive
LR+ = sensitivity / (1 - specificity)
What is the negative likelihood ratio?
False negatives / True negatives
LR- = (1 - sensitivity) / specificity
Why is it important to do post hoc tests to investigate a significant ANOVA test?
They apply a correction for multiple testing to avoid a type 1 error
What are likelihood ratios used for?
They assess the potential use of a diagnostic test and the likelihood of the patient having the disease
A ratio of 0-1 reduces chance of disease
Ratio >1 increases probability of having disease
