Epidemiology Flashcards
epidemiology
study of diseases in populations (compare between groups) (incidence and distribution)
why do we study edipemiology
calculate mean of population so can infer what’s going on in patient by looking at population data - diagnosis, prognosis (how long live e.g. compare CD4 levels to population mortality)
identify risk factors
quantify incidence/prevalence so targeted intervention
epidemiologic approach
find association between exposure to variable and disease (risk factors)
is the difference real
why has is occurred (cause)
prevention and control
target population
population of interest, maybe who is at risk e.g. gender/age
study population
subset of target population that you can access e.g. women gone to doctors in 2009
study sample
subset of study population, actual number of people you study, needs to be representative
proportion vs ratio equation
proportion is a/(a+b)
ratio is a/b
incidence
number of new cases of disease in time period (rate)
(new cases)/(at risk in that time period) so measures risk
already infected or not in target population are not counted because not in ‘at risk’ group
prevalence
disease in population at any one time (snapshot), point prevalence
(cases at specific time)/(no. in population at time)
period prevalence is over defined time because hard to sample at 1 time
incidence vs prevalence
incidence is a measure of risk and prevalence isn’t
incidence is prevalence / duration of infection
prevalence is incidence x duration
(like speed distance time equation)
risk ratio (relative risk)
risk of disease in exposed / risk in unexposed
(no. of disease in exposed / total exposed) divided by (no. of disease in unexposed / total unexposed)
RR 1 means same risk in both so no association
RR more than 1 means risk in exposed is higher
RR less than 1 means risk in unexposed is higher
odds ratio
odds of disease in exposed / odds in unexposed
(no. disease in exposed / no. of healthy in exposed) divided by (no. disease in unexposed / no. healthy in unexposed)
OR 1 means both same
OR more than 1 means odds higher in exposed
OR less than 1 means odds higher in unexposed
attributable risk
relative importance of different risk factors
risk in exposed - risk in unexposed
(no. disease in exposed / total exposed) take away (no. unexposed w/ diseases / total unexposed)
population attributable risk (PAR)
amount of disease attributable to exposure in population, relevant to public health decisions because if risk is important or not and should intervene
AR x prevalence of exposure in population
attributable proportion
proportion of disease that would be eliminated in population if risks reduced to that of unexposed, so number of lives saved
types of epidemiological studies
experimental - randomised controlled
observational - cohort, less case-control, cross-sectional, ecological (population based)
experimental studies
clinical trials to observe effect independent of others, compare treatments experimentally
strongest inference because randomly assigned groups and control other risk factors
similar participants
look forward in time
conclude that difference is due to risk factor
some bias if not random groups
observational studies (descriptive vs analytical)
compare natural exposure so no randomisation so hard to identify cause/effect
descriptive - estimate w/o comparison so don’t identify cause e.g. case studies
analytical - statistical comparisons, cohort study, case-control study, cross-sectional study
cohort study
best observational study
follow exposed/unexposed through time (prospective)
exposures not randomised so bias
can identify cause because temporal sequence (change occurs after risk)
no recall bias because data gathered as happens
expensive and time consuming and can’t use key tests to analyse
case-control study
compare with and without disease and who has exposed (retrospective so back in time) and weaker inference
non randomly assigned
use medical records/question
compare lifestyle and risk
cheap, quick
can’t identify cause, rely on recall
cross-sectional study
measured at 1 point in time like questionnaire/sample once
generate hypotheses but then need experiment
no time sequence so don’t know which cause/effect
quick, cheap
case studies
describes without comparison
understand disease/treatment before experiment
order of studies used
clinical observation then identify available data then case-control study then cohort study then randomised trial
order of studies in best inference
RCT (randomised controlled trial), cohort study, case-control study, cross-sectional study
multiple causal pathway
risk factor may not be associated directly with disease but may need both risks or both cause but don’t need them together
guidelines to judging whether an association is causal (when we can’t do experiment) - Hills criteria
strength of association dose-response relationship correct temporal relationship independent of recognised confounders consistency with other knowledge biologically plausible reversible
strength of association
high relative risk/odds ratio in a cohort/case control study
dose-response relationship
disease increases as exposure increases
correct temporal relationship
if exposure before disease then can establish sequence of events from prospective study
independent of recognised confounders
if known causes are accounted for and there is still a significant association
consistency with other knowledge
if same pattern in many studies
biologically plausible
coherence with biological literature and makes sense
reversible
disease goes away if remove risk factor
3 types of error
random error
measurement error
systematically (bias)
random error
mean not representative of population because of sampling
measurement error
type of random error from data collection
systematically (bias)
if allocations not random it lowers significance of associations
random vs bias error
random decreases chance of detecting difference while bias can modify association so get a wrong conclusion
type I error (alpha)
say significantly difference but truly they are not
happens by chance
allow 5% chance they are false
type II error (beta)
don’t detect association where there truly is one
dependent on ability of study to detect association
aim for 20% chance of error
types of bias
selection bias
recall bias
funding bias
bias due to confounders
selection bias
individuals not representative of population
concern for all types of studies
3 types:
case-control studies - if select on basis of exposure
clinical trials - allocation like double blind
cohort studies - don’t follow up, group less likely to come back
what is done to control errors and bias
hypothesis and alternative hypotheses
appropriate study design and good sample
avoid confounders
recall bias
rely on memory
may recall differently with disease
some groups less likely to report
try to ask Qs in multiple ways
funding bias
sponsored by bias company
bias due to confounders
another variable associated
more likely in observational studies
e.g. sex, age, race
biological variation
diagnosis involves comparing to what know at population level - with distributions
types of distribution
bimodal distribution - 2 peaks so 2 means
unimodal distribution - 1 peak, harder to say if patient infected/not
case definition (+limitations and good enough)
criteria used to decide if individual has disease
clinical signs, diagnostic tests, history
tests often not 100% accurate
not 1 unique criteria and sometimes no known cause
some can only diagnose after death
definitions and symptoms can chnage
but epidemiology is based on groups so some error is okay if we recognise and account for them
dichotomous results
only positive or negative and no distribution
sensitivity (Se)
probability of +ve test result if have disease so how good at identifying disease
(true positive / total with disease)
specificity (Sp)
probability of -ve test result if don’t have disease
true negative / total without disease