lectures 11-23 Flashcards
what is randomisation and what does it achieve
randomly splitting participants into groups. Eliminates confounding because known and unknown confounders should be balanced
cluster randomisation
randomise groups (clusters) of participants instead of individuals, which may be difficult. E.g. GP practices, hospital wards
stratified (block) randomisation
If you want to be certain that important confounders are eliminated. Randomise individuals within each age group, sex, or hospital
cross over studies
Each person gets both treatments -confounding is effectively eliminated. Can only be done for long-term conditions and treatments that do not cure disease
protecting randomisation
- concealment of allocation - Make sure that people can’t cheat and pick the treatment that they prefer
- intention-to-treat analysis - Analyse participants as randomised, this reflects real world
- blinding - of researchers and participants
- complete follow-up
- use large numbers - balance confounding
per-protocol analysis
Analyse as treated (not necessarily as randomised)
Lose the benefit of randomisation
potential sources of bias
- lack of blinding - participants or researchers may act differently if they know the groups
- loss to follow up - cant analyse their results
- non-adherence - e.g. stop taking treatment
4 strengths of RCTs
- The best study design to test an intervention
- Well conducted studies should eliminate confounding and bias
- You can calculate Incidence, Relative Risks, and Risk Differences
- The strongest design for testing cause-and-effect associations
clinical equipoise
must have genuine uncertainty about benefit or harm of intervention in RCTs
practical issues of RCTs
- can be expensive - need many participants, long time
- often funded by pharmaceutical companies - unlikely to fund studies for cheap treatments
- Participants in RCTs are often not representative - They need to meet all the inclusion criteria
- RCTs are not efficient for rare outcomes
internal validity
whether or not there is a real association in the group you looked at, or if its due to chance, bias or confounding
external validity
can findings be generalised to broader population
effect of increasing sample size in random sample
- Makes it more likely to represent sample
- Reduces sample variability (standard deviation etc)
- Increases precision of parameter estimate - Confidence intervals get narrower
2 interpretations of confidence intervals
- A 95% Confidence Interval represents the range of values within which the parameter will lie between 95% of the time if we continue to repeat the study with new samples
- 95% CI = We are 95% confident that the true population value lies between the limits of the confidence interval
what does it mean if 0 in included in a CI for difference between two means
result could be due to chance
null hypothesis
There is no association between exposure and outcome
There is no difference between groups
Parameter equals null value
if this is true, any differences must be due to chance
alternative hypothesis
There is an association between exposure and outcome
Parameter does not equal null value
p value
The probability of getting an estimate as extreme as the one that you have observed if there is really no association
i.e. probability of it occurring by chance
statistically significant p value
P values < 0.05
Less than 5% probability of a result this extreme due to chance
We can reject the null hypothesis and accept alternative hypothesis
not statistically significant
if P value > 0.05
More than 5% probability of this result occurring by chance
We cannot reject the null hypothesis
type I error
false positive
occurs when we find a “statistically significant” result when there is no real difference
we reject H0 even though it is correct and the difference is due to chance
P(type I error) = alpha (significant level - usually 0.05)
type II errors
false negative
occurs when we don’t find a “statistically significant” result when there is a real difference
we fail to reject H0 even though it is false and the difference is real
More likely if we use smaller samples
Also more likely if we look for a smaller P value (e.g. ≤ 0.01)
clinical importance
if it will have a substantial effect/make a decent difference that makes it work funding
selection bias in case-control studies
Controls not representative of the population which gave rise to cases
If inclusion/exclusion criteria differ between cases and controls
selection bias in cohort studies
Loss to follow up
If comparison group selected separately from exposed group can lead to bias - healthy worker effect
how can measurement error occur
Participants provide inaccurate responses
Data is collected incorrectly/inaccurately
effect of measurement error
in a descriptive study could over/underestimate prevalence
In an analytic study can lead to misclassification
non-differential misclassification
When measurement error and any resulting misclassification occur equally in all groups being compared
differential misclassification in cross-sectional study
people with the outcome might report the exposure differently to those without the outcome
differential misclassification in a case-control study
cases might more accurately recall past exposures compared to controls
differential misclassification in a cohort study
interviewer aware of the exposure status may ask more probing questions about the outcome among those exposed compared with those in the comparison group - this is known as interviewer bias
minimising interviewer bias
Clearly defined study protocol and measures
Blinding
Training of interviewers
Structured questionnaire and standard prompts
recall bias in case-control studies
Cases would have had much more time actively thinking about their exposure than controls
If investigating effect of living by powerlines growing up on brain cancer, cases would be more aware of exposure status because they would have thought about it
minimising recall bias
objective measures
memory aids
what is confounding
A mixing or muddling of effects when the relationship we are interested in is confused by the effect of something else
confounders are: (3 things)
- associated with the exposure
- associated with the outcome (independent of exposure)
- not on the causal pathway
potential effects of confounding
- Reverse an association (“Simpson’s paradox”)
- Underestimate an association
- Overestimate an association
identifying confounding
Look at the associations before and after adjustment or stratification
If these are very different (i.e. more than 10% different), the factor was probably a confounder
common confounders
age, sex, socioeconomic status, lifestyle factors
minimising confounding
randomisation
restriction
matching
what is matching and advantages and disadvantages
Individual matching: match each case to one (or more) controls
Frequency matching: select controls to match age & sex distribution of cases i.e. same proportions
advantages:
- Good way to control for confounders – Can select multiple controls for each case
disadvantages:
- Individual matching for multiple confounders can be difficult
- Finding appropriate controls can be difficult
- It is possible to over-match: can miss a true association if controls are so similar that they share the same risk factors
what is restriction and disadvantages
Only include one group (stratum) of potential confounder e.g. only use men older than 35 for sample population
disadvantages:
- It is only practical to restrict for 1 or 2 known confounders
- Reduces generalisability to other groups
- Reduces number of potential participants
controlling confounding at an anaylsis level
- standardisation
- stratification
- multivariable analysis
what is stratification and disadvantages
Analyse in groups (“strata”) of potential confounders
disadvantages:
- Smaller numbers in each stratum
- Can usually only stratify on one (maybe 2) confounders
what is standardisation and disadvantages
Age standardisation: removes age as a confounder when age structure differ
Sometimes standardise for other confounders
disadvantages:
- Less suitable for multiple confounders
- Multivariable analysis is often more efficient
what is multivariable analysis and pros and cons
Mathematically adjust the analyses for multiple confounders
Provides “adjusted” measures of association
advantages:
- Greater sample size (more power) than stratifying into groups
- Can adjust for several confounders at the same time
- Easy to do with statistical software
disadvantages:
- Not always obvious what the data look like
- Unlike stratification, you may not detect effect modification
what is effect modification
A 3rd factor modifies the association between exposure and outcome. This factor is on the causal pathway
identifying effect modification
- Stratify the findings
- Associations in each group are very different –probably effect modification
- Similar associations each group –no effect modification (but there could still be confounding)
necessary cause
component cause which is necessary for disease to occur
Must be part be part of every sufficient cause (if there are multiple)
guidelines to judge whether association means causation
- Biological plausibility
- Experimental evidence
- Specificity
- Temporal sequencing (exposure must come before outcome)
- Consistency
- Dose-response relationship (risk of developing disease increases as dose of exposure increases)
- Strength of association
efficacy trial
Analyse the individuals as they are actually treated- not the with intention to treat
when carrying out randomisation
Equal chance of being in any group
Limit selection bias
Equal distribution of patient characteristics
Same proportion of confounders
factors to consider about research ethics
- balancing benefits and harms
- protecting potentially vulnerable groups
- conflict of interest
- informed consent
- justice (being fair)
balancing benefits and harms
clinical equipoise
Awareness of various costs or harms to participants and strategies to address these harms or costs
Evidence of scientific validity of the research
protecting potentially vulnerable groups
someone who is more at risk of exploitation because of social or physical disadvantage
- e.g. poorer people
- Those subject to racial pr religious discrimination
- Those who are less educated
- Those suffering cognitive impairment
- Older people
- Prisoners
- Children
conflict of interest
Situation where a person holds two or more potentially incompatible interests, which might compromise values of research
Can be managed by peer review, blinding, auditing
informed consent
Disclosure of purpose, risks and processes of study
Reasonable efforts from researcher to explain this info
That the person is competent to give consent
Absence of coercive factors
Must be in writing
critical appraisal
the process of evaluating research in the best way
systematic review
a way of putting a heap of different studies together and then analysing them to answer a question
advantages of systematic review
Reproducibility
Comprehensive
Transparent limits
Gaps in knowledge
Basis for decisions
challenges of systematic review
Doing all steps well
Publication bias
Heterogeneity
Poor quality trials
Conflicting reviews
Inconclusive results
steps of stratification
- calculate crude MoA between exposure and outcome
- divide data by level of suspected confounder (done for us)
- calculate MoA between exposure and outcome for each stratum
if stratum specific values are all similar to each other and to the crude value:
no evidence of confounding
stratum specific values are all similar to each other but different from crude value:
evidence of confounding
stratum specific values are different from each other
effect modification
looking for confounding in multivariable analysis
compare OR for unadjusted and adjusted data
if they are similar then no confounding
if they are different then confounding