Hadpop Flashcards
What is the point of critical appraisal?
To be able to distinguish between good and bad evidence to provide optimal care for patients
It is necessary to critically appraise evidence to decide about the causality
Deterministic or stochastic
Lab based evidence vs pop based evidence
Why do we need a population perspective of medicine?
So that we can study large groups of people to discover the causes of disease and evaluate preventative/ curative measures, using the experience of many doctors
Define deterministic evidence
Validation of hypothesis by systematic observation
Can be used to predict future events with CERTAINTY
Causality
Define stochastic evidence
Assessment of hypothesis by systematic observation
Predict the LIKELIHOOD OF OCCURRENCE of future events
Association
Which evidence, stochastic or deterministic is used in HADPop?
Stochastic as its an association
In which two ways can evidence for causality be obtained?
By lab based evidence
By pop based evidence
Why is lab based evidence not very helpful?
It’s provides contributory evidence but is neither necessary nor sufficient to prove causality
Why is pop based evidence favoured in HADPop?
Association is not equal to causality
More definitive in HADPop than lab based evidence
More flirting with Hadpops stochastic approach
What is the paradox of the commons?
It is the principle that the optimum strategy for an individual is not the optimum strategy for the community
What is a census?
Simultaneous recording of data by the government at a particular time pertaining to all the persons living in a particular territory
What is a census useful for?
Allocation of resources
Trends in populations (age and ethnicity)
Population projections
Tells us about population size, structure and characteristics
What are some features of a census?
Universal coverage Occurs at regular intervals (every 10 yrs in UK) Ran by government Simultaneous Particular time Particular territory
What three things affect population size and structure?
Births deaths and migration
In what 3 ways can births be measured?
Crude birth rate
General fertility rate
Total period fertility rate
What is the crude birth rate?
The number of live births per 1000 of the population
What does the CBR describe?
Impact of births on a population
What is CBR affected by?
Age and gender
Varies directionally proportionally with GFR
What is the general fertility rate?
Number of live births per 1000 fertile women aged 15-44
What does the GFR describe?
Compares the fertility of fertile female populations (eliminates gender - so no longer affected by gender)
Takes into account that half the population (males) cannot give birth
What affects GFR?
Affected by age
Affected by ASFR and age distribution in 15-44 year olds
What is the total period fertility rate?
Average number of children that would be born to hypothetical woman aged 15-44
Why is the TPFR advantageous?
It gives each age equal weighting in its calculation
Standardise age and gender
What does the TPFR describe?
Compared fertility of fertile females without the influence of age (eliminates gender and age)
Sum of age specific fertility rates
What are the determinants of fertility?
Fertility, fecundity and conception
Define fecundity
The physical ability to reproduce
What can cause fecundity to decrease?
Increase in sterilisations and hysterectomies
Define fertility
The realisation of fecundity as live births
What can cause fertility to decrease?
Increased use of contraception and abortions
What can cause fertility to increase?
Increased sexual activity and improvements in economic climate
What is conception?
Live births + miscarriages + abortions
In what 3 ways can deaths be measured?
Crude death rate
Age specific death rates
Standardised mortality ratio
What is crude death rate?
Number of deaths per 1000 population
What is age specific death rate?
Number of deaths per 1000 people in an age group
What is ASDR affected by?
Affected by gender
What is the standardised mortality ratio?
Summative figure comparing the observed deaths in an age group with the expected deaths if the age sex distribution of the population were identical
Adjusts for age sex distribution
Why is mortality data useful?
Classify cause of death
Analyse patterns in mortality rates
Identify health problems
Inform service needs
How can mortality data be useful in diseases with a rapid progression?
Mortality statistics can be a measure of incidence
Define migration
Movement of people into or out of a population in a given time period
What are population estimates?
Apply what is known about births, deaths and migration to the PRESENT to calculate an estimate of population size
Any difference in the population estimate compared to census data is dependent on…?
Migration only
As birth and death rates are known
What are population projections?
FUTURE orientated estimates which rely on assumptions about births, deaths and migration in the future
Any difference in the population projection compared to census data is dependent on…?
Fertility rates and migration
As death rates remain relatively stable and so are generally close to true values
What can health information be used for?
Identify health and health care needs Monitor trends in disease Monitor performance in healthcare Monitor contentious issues associated with use of health info- -completeness and or duplication -accuracy -confidentiality -numerator/denominator mismatch -varying diagnosis of disease
What is variation?
Where there is a difference between the observed and actual value
Variation is ALWAYS present due to chance (random variation)
Additional variation may be present due to bias and confounding (systematic variation)
What causes variation in results?
Random variation- chance
Systematic variation- confounding and bias
Define a confounder
A characteristic of the population that is associated with both the outcome and the exposure of interest, but is not on the causal pathway between exposure and outcome
How do confounders affect results?
Confounders can distort results and give misleading results
They can show potential causal links which are actually unfounded
Define bias
Systematic error in the design of a study that results in an incorrect estimate of the association between exposure and risk of outcome
What is the difference between bias and confounding?
Both are types of systematic variation
Bias- Systematic error in the design of a study that results in an incorrect estimate of the association between exposure and outcome
Confounder- characteristic of the population that affects both exposure and outcome, but is not on the causal pathway between exposure and outcome
B: So systematic error in study design vs C:population characteristic
Bias affects relationship between e and o vs confounders affect e and o individually
What are some types of bias and where can they be found?
Selection bias- allocation bias (RCT’s) and healthy worker effect (external cohort trials) survivor bias (cohort studies) selection bias (case control studies)
Information bias- recall bias (conventional case control studies) and publication bias (systematic reviews)
Define selection bias
Error due to systematic differences in the ways in which the two groups were collected
Define information bias
Error due to systematic misclassification of subjects in the group
What is incidence?
Number of new cases arising in a given time
Rate
What is prevalence?
Number of people affected by a disease at any given moment
Proportion
Draw the prevalence bath
Incidence coming in, death and cure coming out
P=~ I x L
How do you increase prevalence?
Use bath
By increasing incidence and keeping people alive longer
How do you decrease prevalence?
Use bath
Decrease prevalence by curing more patients/ killing more patients
What is absolute risk?
Any rates are measures of absolute risk
Point prevalence
Incidence rate
How do you calculate point prevalence?
No. of sufferers / No. at risk
No time frame
E.g 30 people per 1000 have diagnosed colorectal carcinoma
How do you calculate incidence rate?
Number of new cases/ Person years
E.g. 3 cases of colorectal carcinoma diagnosed per 1000 people per year
1000 Person years = number of people x number of years x 1000
Usuall given in 1000 person years
What is relative risk?
Any ratios are measures of relative risk as they compare groups (e.g exposed and unexposed groups)
Incidence rate ratio
Standardised mortality ratio
What is IRR used for?
Comparing incidence rates of two separate populations, varying in exposures to establish whether exposure causes disease
Comparing incidence rates of groups exposed to a new and old treatment to establish the efficacy of a new treatment
How do you calculate IRR?
E.g. Comparing A (study pop) with B (standard ref pop)
Incidence rate of group A / incidence rate of group B
How do you interpret an IRR?
E.g. Comparing A (study pop) with B (standard ref pop)
IRR is 2
A is 2 times likely to get the outcome when exposed to the exposure compared to B
IRR is 0.5
A is 05 time likely to get the outcome when exposed to the exposure compared to B
What are some limitations of IRR?
Age and sex are both factors that can confound the risk of disease
So at risk of confounding
What can occur to data to eliminate confounding?
Age sex standardisation
Why are separate mortality rate ratios for each age band not used?
They could be difficult to interpret and require lots of time
What is the standardised mortality ratio?
Summative figure describing mortality experienced by a local population compared with a general standard population which takes into account confounding factors (age and sex)
Type of relative risk
What is SMR used for?
Compare observed deaths in an age group with the expected deaths if the age sex distribution of the population were identical
How do you calculate SMR?
100 x Total no. of observed deaths in a study pop / total no. of expected deaths in a pop
What observed quantities can be used for the numerator of the SMR?
Incidence, prevalence, IRR
Which depart from true values via random variation
What is the relevance of an observed value?
It is the best estimate we have of the true value and enables us to test hypothesis about true values and calculate confidence intervals containing the true value
Define a hypothesis
A statement that an underlying tendency of scientific interest takes a particular value
What can be calculated to allow for variations in an epidemiological study?
Error factor and confidence intervals
What does it mean if a null hypothesis lies within the CIs?
So if the null hypothesis lies within the CIs:
not statistically significant
cannot reject the null hypothesis (it may be true that there is equal likelihood of outcome in A and B)
data is consistent with stated hypothesis
Can’t say with confidence that you are IRR times likely to get an outcome in A than in B
P>0.05
Cannot exclude that any variation between observed and true value is most likely due to random variation (chance)
What does it mean if the null hypothesis does not lie within the CIs?
So if the null hypothesis does not lie within the CIs:
statistically significant
can reject the null hypothesis (it is not true that there is equal likelihood of outcome- ie. there is a difference)
data inconsistent with stated hypothesis
so there is a difference/ two groups are not the same
you can state that you are IRR times likely to get an outcome in A than in B
P<0.05
any variation between observed and true value is most likely due to systematic variation (bias or confounding)/ unlikely to be due to chance
What is the p value?
Probability of obtaining a test statistic
Universally accepted at 5 %
The p-value states how likely the results in the study would have occurred by chance if the null hypothesis was true.
What are confidence intervals?
Range of values that we can say with confidence that the actual value will lie in between this range in 95% of cases
How do you calculate confidence intervals?
Upper bound confidence interval = observed value x error factor
Lower bound confidence interval = observed value / error factor
What values are used in calculating error factors?
d = events observed in the population
So e.g IR 5 new cases in 1000 use 5
So e.g. IRR IR 5 new cases in 1000 and 8 new case in 2000 use 5 and 8
What is the relationship between size of a sample and error factors and confidence intervals?
Large sample will have small error factor and hence narrow confidence intervals
Small sample will have larger error factor and hence wider confidence intervals
Ultimately we want smallest error factor and narrowest confidence intervals as possible provided that its not costly and not time consuming
How do you interpret SMR?
Null is 100
SMR > 100 suggests excess mortality
SMR < 100 suggests less mortality…
…with confounders accounted for
When is a null hypothesis 0/1?
Null hypothesis is 0 when comparing differences (i.e. there is no difference)
Null hypothesis is 1 when comparing rate ratios (i.e. incidence rates are the same)
What is a cohort study?
Type of observational analytical study that classifies patients based on EXPOSURE and then follows patient up, monitoring the progression of the OUTCOME (disease)
How do you lay out a cohort study table?
First column: exposed and unexposed First row: disease and no disease AB CD IRR= A(C+D)/C(A+B) = A/(A+B) / C/(C+D)
What are internal and external cohort studies?
Internal cohort study- compares incidence between 2 cohorts in a population; calculate IRR
External cohort study- compares incidence within a cohort with a reference population; calculate SMR
What is the effect of internal and external cohort studies on the error factors and hence CIs?
Internal- IRR: smaller cohort –> larger ef –> larger CIs
External- SMR: ef will be smaller than with internal as larger cohort–> smaller ef –> smaller CIs
Hence due to there being smaller ef with SMR external, external cohorts are preferred
What are some advantages of external cohorts over internal cohorts?
Smaller ef and hence smaller CIs with external cohort which uses SMR
What are some limitations of external cohort studies?
Limited data is available for the reference population so mortality data is used- may not be detailed enough
Study and reference population may not be comparable as a result of selection bias and the healthy worker effect
Describe the two types of cohort studies
Concurrent- prospective data collection- waiting for events to happen
Historical- retrospective data collection- from pre existing records
What are some advantages of cohort studies?
Can study a range of outcomes for one exposure
Good for Rare exposures
Good at establishing that exposure precedes outcome
What are some disadvantages of cohort studies?
Bad for rare outcomes/ rare diseases Large Time consuming Expensive Risk of high no. of losses to follow up = survivor bias Difficulty with confounding variables
What is a case control study?
Type of observational analytical study that classifies patients on the basis of OUTCOME (disease) and then queries about previous EXPOSURE
How do you lay out a case control study table?
First column: exposed and unexposed First row: case and control ab cd OR= ad/bc
What is the rare disease assumption?
In the case of rare diseases, in a cohort study:
A (exposed and develops disease) and C (unexposed and develops the disease) –> 0
Which suggests that IRR –> AD / BC
In a case control study:
OR= ad/ bc
Therefore under the rare disease assumption IRR = OR
How many more controls should you recruit compared to cases in a case control trial?
You should recruit 4-6 times the number of cases as controls
Since the precision of the OR is affected by the number of controls (healthy people) as well as the number of cases (diseased)
Recruitment of 4-6 times the cases decreases the error factor
Why is not sensible to recruit more than 4-6 for the number of controls?
Beyond 4-6 is not worth it as the decrease in the error factor is disproportionate to the increase in cost/ time/ expense of controls
What are the two type of case control studies?
Conventional- retrospective collection of data (recall)
Nested- collection of data from evolving outcomes and exposure database of a concurrent/ prospective cohort study
Why are nested case controls considered to be better than conventional case control studies?
Incidence rates can be calculated as sampling fraction is known
In a conventional case control the cases and controls are sampled from a source population with unknown size, whereas source population size is known in a nested case control study
In a nested case control the cases emerge from a well defined source population and the controls are sampled from that same population. In conventional it is difficult to ensure that cases and controls are sampled from the same source population.
Why are nested case control studies considered to be better that cohort studies?
Can collect more detailed information for a minority of participants e.g. Blood samples
What are some advantages of case control studies?
Good for rare outcomes/diseases
Can study a range of exposures for a single outcome
Quicker than cohort
Cheaper that cohort
Can be used for diseases which have a longer developmental time
What are some disadvantages of case control studies?
Bad for rare exposures
Difficult to establish that exposure precedes outcome (temporal sequence)
Information (Recall) bias- non differential misclassification (randomly inaccurate measurement) or systematic misclassification (assessor bias and differing data collection methods)
Selection bias- cases should be representative of all cases; controls should be representative of a population (e.g getting all cases and controls from a respiratory ward where patients are more likely to have respiratory diseases)
Confounding- can be minimised by matching; adjusted for by analysing with logistic regression
What is the relevance of information bias in case control studies?
- non differentiated misclassification (randomly inaccurate measurement)
- 10% non differentiated misclassification- where cases and controls lose 10% of themselves and gain 10% of other exposed or unexposed [OR –> 1 (null)]
- systematic misclassification (assessor bias and differing data collection methods)
- systematic misclassification of 10% of unexposed cases- exposed cases gain 10% and unexposed cases lose 10% [OR –> AWAY FROM 1 (null)]
- systematic misclassification of 10% of unexposed controls- exposed controls gain 10% and unexposed controls lose 10% [OR –> 1 (null)]
What is the limitation of selection bias in case control studies?
Cases should be representative of all cases
Controls should be representative of the entire population
In case control trials this can sometimes be overlooked in the form of selection bias:
E.g. getting cases and controls from a respiratory ward means that the controls are also like to suffer from other respiratory diseases too which means that the controls won’t be representative of the entire population
How can confounding be minimised in case control studies?
By MATCHING using important confounders
Or
By adjusting for the confounders by analysing with logistic regression
E.g.
In a study to look at the risk of myocardial infarction in users of oral contraceptive pills, age is a very important confounder
Therefore, cases and controls were matched by their date of birth
And the analysis involved logistic regression to give ORs adjusted for other potential confounders
What is a clinical trial?
Any form of planned experiment which involves patients and is designed to elucidate the most appropriate method of treatment of future patients with a given medical condition
So a new treatment and standard treatment (or placebo) is allocated by a trial investigator and outcomes are measure over time
What is the purpose of a clinical trial?
To provide reliable evidence of treatment efficacy and safety
Define efficacy
Ability of health care intervention to improve the health of a defined group under specific conditions
Define safety
Ability of health care intervention not to harm a defined group under specific conditions
What three things do clinical trials need to be?
Fair
Controlled
Reproducible
What is a non randomised controlled trial? And what is it prone to?
Where patients with new treatment are compared with patients with a standard treatment
Prone to selection bias/ known and unknown confounding
What is a comparison with historical controls clinical trial? And what is it prone to?
Where patients who had previously received standard treatment are compared with patients who presently receive new treatment
Prone to: selection criteria being less well defined/ standard and new treatment being treated differently/ less info about potential confounders/ bias/ unable to control for confounders
Why are non randomised controlled trials and historical control clinical trials not good?
Always overestimate benefits of the new treatment
What are randomised controlled trials?
Experimental studies that answer the question- is the new treatment BETTER OR WORSE than the usual standard treatment?
What are the 3 broad steps of RCT’s?
Definition of factors
Conduct of trial
Comparison of outcomes
What is involved in the defining of factors in a RCT?
Disease Treatment Eligible patients Patients that will be excluded Outcomes Potential bias and confounders
What is involved in the conduct of trial in a RCT?
Identifying and recruiting patients Gaining consent of patients Allocating patients randomly Follow up Minimising losses and maximising compliance
Why are outcomes compared in RCT’s?
To see if any difference in response to treatment is statistically significant and clinically important
Why are outcomes predefined in RCT’s?
To prevent data dredging and repeat analyses
To set data collection protocol
To provide agreed criteria for measurement and assessment of outcomes
What are primary and secondary outcomes in RCT’s?
Primary outcome- usually 1 to calculate a sample size
Secondary outcome- side effects
What are three categories of outcomes in RCT’s?
Pathophysiological
Clinically defined
Patient focused
When are outcomes measured in RCT’s?
Baseline
During trials
Final
Describe an ideal outcome of an RCT
Appropriate and relevant Sensitive and specific Cheap and timely Reliable and robust Valid and attributable Simple and sustainable
How is allocation carried out in RCT’s?
Through randomisation and blinding
- minimises selection bias and confounding
- can give strong evidence of causality
What is randomisation?
In RCT’s where treatment is allocated randomly using random number tables/ generators
What is blinding?
Where one(single), two(double), or three(triple) of the patient, clinical consultant and investigator do not know the destination of treatment - avoids bias by e.g. Labelling bottles the same
Where can blinding be difficult?
Surgical procedures
Lifestyle interventions
How can we minimise losses to follow up of a RCT?
Maintain contact with participants Practical follow up Minimise inconvenience Be honest about commitment Avoid coercion/ inducements
How can we maximise compliance with treatments in RCT’s?
Simplify instructions
Ask about compliance, effects, side effects
Monitor compliance
In what two ways can results from RCT’s be analysed?
Explanatory trial/ as treated
Pragmatic trial/ intention to treat
What is an explanatory / as treated analysis?
ONLY analyses those which have COMPLETED FOLLOW UP AND COMPLIED WITH TREATMENTS
Compares the physiological effects of treatments
What is bad about explanatory / as treated analysis of RCT’s?
It loses the effect of randomisation
Non compilers are likely to be systematically different from compilers
Selection bias and confounding present
Less realistic
Tends to overemphasise the effects of treatments
What is good about pragmatic / intention to treat analysis of RCT’s?
Preserves effects of randomisation
Minimal selection bias and confounding
More realistic
What is a pragmatic / intention to treat analysis?
Analyses those allocated to original treatment groups (REGARDLESS OF WHETHER THEY COMPLETED FOLLOW UP OR COMPLIED WITH TREATMENT)
Compares likely effect of using treatments in routine clinical practice
What are the key points about RCT’s and ethics?
Collective ethic- all patients have the right to a safe and effective treatment
Individual ethic- RCT’s go against beneficence, maleficence, autonomy and justice
RCT’s continue though BECAUSE THEY ARE FOR THE BENEFIT OF FUTURE PATIENTS
How can you increase ethics in clinical trials?
Clinical equipoise- when whether a treatment is better or worse is genuinely not known- unethical to allow someone to have treatment that is known to be bad)
Scientifically robust- relevant issue addressed with a valid question (unethical to be investigating a pointless question)
Ethical recruitment- inappropriate inclusion of people who won’t benefit/ inappropriate exclusion of people who differ from an ideal homogenous group
Valid consent- from a knowledgable informant and appropriate info
Voluntariness- valid consent required; perceived or actual coercion invalidates the consent
Clinical equipose
When whether a treatment is better or not is genuinely not known
It would be unethical to make someone go through with a treatment which you knew was bad
Scientifically robust
When a relevant issue is being addressed with a valid question
It would be unethical to be investigating a pointless question
Ethical recruitment
Include anyone who is capable of being in the trial
Unethical:
Inappropriate inclusion of individuals who won’t benefit from the treatment
Inappropriate exclusion of individuals who differ from an ideal homogenous group
Valid consent
From a knowledgeable informant and appropriate info
Voluntariness
Gain valid consent
Unethical if perceived or actual coercion is present
What are the henle Koch postulates?
• The agent must be shown to be present in every case of the disease by isolation in pure culture.
• The agent must not be found in cases of other disease.
• Once isolated, the agent must be capable of reproducing the disease in
experimental animals, and must be recovered from the experimental disease produced.
Isolate an organism
Propagate the organism in a pure culture in vitro
Reproduce the disease by injecting it into a susceptible recipient
Re-isolating the organism
Necessary (cause must be present before for disease to occur),
sufficient (cause alone results in disease),
specific (disease only results from this cause)
More fitting to a disease: cause effect relationship
What are HAI’s?
Hospital acquired infections
What are problems with HAI’s?
Increase the length of hospital stay
Cause deaths
Incur extra care costs
Problem is greatest in intensive care
How can we prevent HAI’s?
Hand washing
Restrict antibiotic usage
Cohort colonised and non colonised staff and patients together
In epidemiology- what causes disease to occur?
Epidemiology assumes that disease does not occur at random and has causal and preventable factors
What is an association?
Statistical dependence between two or more events, characteristics or other variables
Doesn’t necessarily imply a causal relationship
What does it mean for cause to be necessary and sufficient?
Necessary- cause must always precede the outcome
Sufficient- cause can cause disease on its own
Cause can be necessary, sufficient or both
Koch henle
What is a cause?
Exposure or factor that increases risk of disease
What are the Bradford hill criteria that confer causality?
Stronger arguments for causality (So Sick & Tired Can't Do Revision): Strength of association Specificity of association Temporal sequence Consistency of association Dose response Reversibility
Weaker arguments for causality (Can’t Be Asked):
Coherence of theory
Biological plausibility
Analogy
How is causality assessed?
Internal validity
- eliminate non causal associations
- are the features of causality present
External validity
-generalisation of results with: source population, other populations and your population
Compare results with other studies’ evidence
What is a systematic review?
An overview of primary studies that used explicit and reproducible methods
- An overview of primary studies usually RCT’s
- Unbiased and objective
- Synthesis - possibly including a meta-analysis
- Extremely Credible - explicit, transparent and reproducible
- Facilitates the synthesis of a large number of study results
- Might be done badly so should be appraised
- Inappropriate aggregation of studies
What is the relevance of a meta analysis?
Allows for a QUANTITATIVE comparison between primary studies that addressed the same hypothesis in the same way
How is a meta analysis performed?
Odds ratio and 95% CI’s for all studies are calculated, combined to give a pooled estimate odds ratio
Studies are weighted according to their size and uncertainty of Odds ratio (smaller e.f. gives a greater weight)
2 methods of calculating the pooled estimate odds ratio and its 95% CI:
- Fixed effect model - assumes studies estimate exactly the same effect size (any variation between data comes from within study variation)
- Random effect model - assumes studies are estimating similar (not same) effect size (any variation between data comes from within study variation and between study variation)
What are the two methods of calculating a pooled odds ratio in a meta analysis?
2 methods of calculating the pooled estimate odds ratio and its 95% CI:
- Fixed effect model - assumes studies estimate exactly the same effect size (any variation between data comes from within study variation) LESS VARIATION (smaller CIs)
- Random effect model - assumes studies are estimating similar (not same) effect size (any variation between data comes from within study variation and between study variation) MORE VARIATION (larger CIs)
Describe a forest plot
Graphical representation of a meta analysis
Squares (size proportional to weight)- individual OR’s Lines - 95% CI’s Diamond - pooled estimate Centre - pooled OR with dotted line Width - pooled CI Solid Line - Null Hypothesis value
What does a Forrest plot showing lots of RCT’s with an OR>1 suggest?
Indicates greater odds of survival
What are some problems with systematic reviews and meta analysis?
Heterogeneity
Publication bias
How can you test for heterogeneity?
Statistical test applied to results of trials being reviewed to assess whether the differences in the individual trial results are consistent with random variation or not - Chi squared statistical analysis (10% ci)
If not consistent with random variation then ps and weighting of studies is more equal
Subgroup analysis can help to explain heterogeneity which may provide further insight into the effect of a treatment or exposure
How can you test for publication bias?
Funnel plot
If there is no publication bias- small trials will be along the bottom of the graph and if there was no bias of publishing small trials with a positive effect, then there should be an even distribution across the bottom centred around the results of the larger trials higher up the graph
Well balanced systematic review will show a funnel shape in its studies when plotted
A biased systematic review will vary in shape
What is publication bias?
Where studies are only likely to be published if results are statistically significant or have a large sample size- if systematic studies only include published studies, results will be biased
So systematic reviews should include published and unpublished trials
What is heterogeneity?
The variation between different studies within a systematic review
Tested for using statistical tests (chi squared etc)
If systematic variation is found (p<0.05), can be accounted for by using a random effects model (which allows for more variation) rather than the fixed effect model
What is a funnel plot? How can it be interpreted?
Inverted funnel being superimposed on a graph with standard error (size of trial) on the y axis and treatment effect (result of trial)on the x axis
Poor test for publication bias
Poor method of showing heterogeneity
What are the advantages of systematic reviews?
Explicit methods can reduce bias
Exclusion of poor quality studies
Meta analysis provides an overall figure for the study
Large amounts of info can be assimilated quickly by healthcare professionals
Reduction in time between research discovery and implementation for clinical use
Used in Evidence Based Practice guidelines
How do you establish causality?
• Start with a hypothesis.
• Design a study to test the hypothesis.
• Validate any associations found by excluding possible alternative
explanations, e.g. – chance
– bias
– confounding
– reverse causality.
• Consider whether the statistical associations represent a cause-effect
relationship between exposure and disease.
Bradford hill criteria
Consider a case-control study of heart disease and smoking. What would be the effect on the odds ratio if:
a.the cases understated more than the controls how much they smoked?
This would bias the estimated odds ratio towards underestimating the true effect of smoking on heart disease.
Consider a case-control study of heart disease and smoking. What would be the effect on the odds ratio if:
b. the controls understated more than the cases how much they smoked?
This would bias the estimated odds ratio towards overestimating the true effect of smoking on heart disease.
Consider a case-control study of heart disease and smoking. What would be the effect on the odds ratio if:
c. both cases and controls understated randomly how much they smoked?
Although you would expect the odds ratio to be unaffected, it would in fact ‘shrink to the null’, i.e. be closer to 1.0 than it should be, as the random understatement of exposure by cases and controls converge to same level
Can you explain in your own words what the difference is between bias and confounding?
Bias is a characteristic of a flawed study – either the population studied is unrepresentative (selection bias: correct information from the wrong people) or the information gathered is systematically wrong (information bias: incorrect information from the right people).
Confounding is a characteristic of the population – e.g. drinkers throughout the population are generally more likely to smoke than non-drinkers – and this would affect even a perfect study.
What is an ecological type of comparison?
When exposure and outcome are not linked on an individual basis
What is subgroup analysis?
It can help to explain heterogeneity which may provide further insight into the effect of a treatment or exposure
Subgroup analysis
Stratification by study characteristics: where subset of ‘whole’ studies defined by:
Study design
Participant profile
Stratification by participant profile:where data is analysed by types of participants:
Although this has greater statistical power than for individual studies, data is often unavailable
Why are studies harder to assess in systematic reviews?
Poor study design
Poor design protocol
Poor protocol implementation
How susceptible are different studies to bias and confounding?
Less susceptible • Nonrandomised control studies • Cohort studies • Case control studies Most susceptible