Measures of Association_Prof Bhisma
description
Transcript of Measures of Association_Prof Bhisma
Measures of Association
Prof. Dr. Bhisma Murti, MPH, MSc, PhD
Overview for Today
• Purpose of measuring associations
• The 2x2 table
• Ratio comparisons (relative measures)
– Relative risk, risk/rate ratio, odds ratio
• Difference comparison (absolute measures)
– Risk/rate difference, population risk/rate difference, attributable
proportion among exposed and total pop.
• Formula, examples, interpretations
• Put it all together with an example
Why Estimate Comparisons?
• Quantitative comparison is an essential element of
epidemiology to identify disease determinants.
• Summarize relationship between exposure and disease
by comparing at least two frequencies in a single
summary estimate.
• Overall rate of disease in an exposed group says nothing
about whether exposure is a risk factor for or causes a
disease.
– This can only be evaluated by comparing disease occurrence in
an exposed group to another group that is usually not exposed.
– The latter group is usually called the comparison or reference
group.
Two Main Options for Comparison
Calculate ratio of two measures of disease frequency
1. Relative comparison
2. Strength of relationship between exposure and outcome
3. Magnitude of association between exposure and outcome
4. How much more likely one group is to develop outcome
Calculate difference between two measures of disease
frequency
1. Absolute comparison (attributable)
2. Public health impact of exposure
3. How much greater is the frequency of outcome in one group
Comparing Ratios
The “2-by-2” Table
Yes No Total
Yes a b a+b
No c d c+d
Total a+c b+d a+b+c+d
Outcome
Exposure
Prevalence or cumulative incidence
= a/(a+b) among exposed
= c/(c+d) among unexposed
= (a+c)/(a+b+c+d) among total population
Example #1
Yes No Total
Yes 23 304 327
No 133 2816 2949
Total 156 3120 3276
MI
OC use
5-year CI of myocardial infarction
= 23/327 = 7.0% among OC users
= 133/2949 = 4.5% among non-OC users
Data from a fixed cohort study of oral contraceptive use (OC)
and myocardial infarction (MI) in pre-menopausal women
followed for 5 years (adapted from Rosenberg et al AJE 1980).
The “2-by-2” Table for Person-Time Data
Yes No Total
Yes a - PTexp
No c - PTunexp
Total a+c - Total PT
Outcome
Exposure
Incidence rates
= a/(PTexp) among exposed
= c/(PTunexp) among unexposed
= (a+c)/(Total PT) among total population
Example #2 -- Using Person-Time Data
Yes No Total
Yes 30 --- 54,308 PY
No 60 --- 51,477 PY
Total 90 --- 105,786 PY
CHD
PH use
Incidence rate of CHD
= 30/54,306 PY = 55.2/100,000 PY among PH users
= 60/51,477 PY = 116.6/100,000 PY among non-PH users
Data from a dynamic cohort study of postmenopausal
hormone use (PH use) and coronary heart disease (CHD)
among women (Stampfer et al NEJM 1985)
Relative Comparisons
• Based on ratio of 2 measures of frequency
• Often referred to as “relative risk”
• Dimensionless and ranges 0 - infinity
Prevalence ratio (PR)
= Pexp / Punexp = [a / (a+b)] / [c / (c+d)]
Cumulative incidence ratio (CIR)
= CIexp / CIunexp = [a / (a+b)] / [c / (c+d)]
Incidence rate ratio (IRR)
= IRexp / IRunexp = [a / PTexp] / [c / PTunexp ]
Interpretations of Risk Ratios
• Gives information on the relative effect of the exposure on the disease.
• Tells you how many times higher or lower the disease risk is among the exposed as compared to the unexposed.
• Therefore, commonly used in etiologic research as a measure of risk
• Note:– Some people consider CIR and IRR better measures of “relative
risk” than PR because CIR and IRR include measures of time
Example #1 (continued)
Yes No Total
Yes 23 304 327
No 133 2816 2949
Total 156 3120 3276
MI
OC use
5-yr CIR of MI among OC users compared to non OC users:
5-yr CIR= CIe/CIu = [a/(a+b)]/[c/(c+d)] = (23/327)/(133/2949) = 1.6
Interpreted as relative risk:
1. Women who used OCs had 1.6 times the risk of having MI over 5
years compared to non-OC users (1.6-fold increased risk).
2. There is a 60% increase risk of MI among OC users over a 5-yr period
compared to non-users (60% more likely to have MI).
Example #2 (continued)
Yes No Total
Yes 30 --- 54,308 PY
No 60 --- 51,477 PY
Total 90 --- 105,786 PY
CHD
PH use
Incidence rate ratio of CHD among PH users compared to non-PH users:
IRR = IRe/IRu = [a/PTe]/[c/PTu] = 0.5
Interpreted as a relative risk:
• Women who used PH had 0.5 times, or half, the risk of developing
CHD compared to non-users.
Interpretation of Relative Risk
• RR=1
– Risk in exposed = risk in non-exposed
– No association
• RR>1
– Risk in exposed > risk in non-exposed
– Positive association, factor is associated with disease
– Larger RR stronger association
• RR<1
– Risk in exposed < risk in non-exposed
– Negative association, factor is “protective”
• RR for exposed = 1/RR for non-exposed
– Need to pay attention to referent group
– Risk for whom compared to whom?
Example #1: Extension to 2 x 2 Table
Yes No Total
< 1 year 4 31 35
1-4 years 5 107 112
5-9 years 7 127 134
10+ years 7 39 46
Never 133 2816 2949
Total 156 3120 3276
MI
OC use
Any Two Groups Can Be Compared
Applying Formulas on Appropriate Cells
Yes No Total
< 1 year 4 31 35
1-4 years 5 107 112
5-9 years 7 127 134
10+ years 7 39 46
Never 133 2816 2949
Total 156 3120 3276
The 5-yr CIR (“relative risk”) of MI among users of OC for greater than 10 years
compared to never users is (7/46) / (133/2949) = 3.4.
The 5-yr CIR (“relative risk”) of MI among users of OC for greater than 10 years
compared to users of 5-9 years is (7/49) / (7/134) = 2.7
Case-Control Studies
• Participants are selected for study participation on the
basis of pre-existing disease status
• Cannot estimate prevalence, CI, or IR
– Do not know the population at risk
• Cannot use previous formulas
• Relative risk can be estimated by odds ratio (OR)
– ratio of odds of exposure among cases to odds of exposure
among controls
OR = (a/c)/(b/d) = ad/bc
Disease
Yes
NoExposure
Exposure
Yes
Yes
No
No
Case-control: Looking Back
Example #3: Odds RatioExample adapted from Kehrberg et al 1981 AJE
Cases Controls Total
Rely 15 14 (29)
Other 9 45 (54)
Total 24 59 (83)
Toxic shock syndrome
• OR = (15*45)/(14*9) = 5.4
• Interpretations:
– The odds of using Rely tampons among women with TSS were 5.4 times higher than the odds of using Rely tampons among those without TSS (technical)
– Women who used Rely tampons were 5.4 times more likely to develop toxic shock syndrome than women who used other brands (loosely).
Tampon brand
used during
month of illness
Notes on Odds Ratios
• Can be thought of as “exposure odds ratio”
• These data provide an estimate of risk in some situations.
• They do not allow one to estimate incidence of TSS in the
population of women at risk.
• To do this, one would need to know the number of all cases
of TSS among women (numerator) and the number of all
women who are at risk (denominator).
• Usually, denominator data are not available and numerator
data may be incomplete in case-control studies.
Odds Ratio Approximates Relative Risk
• When disease is rare
– Proportion of cases in exposed and unexposed groups is low
– a<<b, so a+b ≈ b and c<<d, so c+d ≈ d
– RR = a/(a+b)/c/(c+d) ≈ a/b / c/d = ad/bc
• If disease is not rare:
– When cases are newly diagnosed
– When prevalent cases are excluded, making it more like a
cohort/ incidence study
Comparing Differences
Difference (Absolute) Comparisons
• Based on difference between 2 measures of frequency
• Comparing disease occurrence among the exposed with
the disease occurrence among the unexposed
comparison group by subtracting one from the other.
• Gives information on:
– the absolute effect of exposure on disease occurrence
– the excess disease risk, or disease burden, in the exposed
group compared to the unexposed group
– the public health impact of an exposure, that is, how much
disease would be prevented if the exposure were removed
• Note: this assumes that the exposure causes the
disease
Risk Difference
Risk difference (RD) = Rexp – Runexp
For Prev*: RD = Pexp - Punexp = a / (a+b) – c / (c+d)
For CI: RD = CIexp - CIunexp = a / (a+b) – c / (c+d)
For IR: RD = IRexp - IRunexp = a / PTexp – c / PTunexp
• * Some epidemiologists hesitate to use “risk” when referring to prevalence estimates
• More precisely called prevalence difference, cumulative incidence difference, and incidence rate difference
• Also called attributable risk, rate difference, attributable rate
– Note: “attributable” implies causality
– RD = 0 when there is no association between exposure and disease
• What we can hope to accomplish in reducing risk of disease among exposed if exposure were eliminated
Example #4(adapted from Boice 1977 JNCI)
IRD = 6.7 / 10,000 PY
• Interpretation:– Broad: there are 6.7 excess cases of breast cancer for every 10,000 PY
among those exposed to radiation compared to those not exposed.
– Narrow: Eliminating this radiation would prevent 6.7 cases of breast cancer for every 10,000 PY (“attribution”).
Breast
cancer casesPY
Rates per
10,000 PY
Radiation exposure 41 28,010 14.6
No radiation
exposure15 19,017 7.9
Total 56 47,027 11.9
Example #5: Comparison of RR and RD
Lung Cancer Coronary Heart
Disease
Cigarette Smoker 140 669
Non Smoker 10 413
RR 14.0 1.6
RD 130/100,000/YR 256/100,000/YR
Annual Mortality Rate Per 100,000
Conclusion: Cigarette smoking is a much stronger risk factor for lung
cancer but (assuming smoking is causally related to both diseases) the
elimination of cigarette smoking would prevent far more deaths from
coronary heart disease. Why is this so?
Death from CHD is much more common.
In Other Words
• Relative risk is a measure of strength of association
between exposure and disease and is useful in analytical
studies
– risk
• Relative difference is a measure of how much disease
incidence is attributable to exposure, and is useful in
assessing exposure’s public health importance
– burden
Population Risk Difference (PRD)
• Measures excess disease occurrence among the total
population that is associated with the exposure.
• Helps to evaluate which exposures are most relevant to the
health of a target population.
– Describes impact of exposure on total population
– Number of cases that would be eliminated in total population if
exposure was removed (assuming causality)
• Depends on prevalence of exposure in population
– For ex, PRD will be low if exposure is rare, even if RR is high
• Also called population attributable risk
Calculating Population Risk Difference
• Two formulas for PRD:
PRD = Rt - Ru where Rt = risk total population and
Ru = risk among unexposed
PRD = (RD)*(PPexp) where PPexp is % of population
that is exposed
• Note (as always) that “risk” may refer to IR or CI (more
accurately) or prevalence (less accurately)
Example #4 Revisited
PRD = 11.9 – 7.9 = 4
• Interpretation:
– 4 excess breast cancer cases for every 10,000 PY of observation can be attributed to radiation exposure.
– If radiation causes breast cancer, then 4 cases of breast cancer for every 10,000 person-years of observation could be prevented if the radiation exposure were removed.
Breast cancer
casesPY Rates/10,000 PY
Radiation exposure 41 28,010 14.6
No radiation exposure 15 19,017 7.9
Total 56 47,027 11.9
Attributable Proportion among Exposed
• APe describes the proportion of disease among exposed that is due (attributable) to exposure or that would be prevented if exposure were eliminated
• Risk in non-exposed group can be considered “background” incidence that would occur regardless of exposure
• Interpretation may assume causal relationship
APe = [(Re – Ru)/Re] * 100
= RD / Re * 100
Re = “risk” (IR, CI, P) among exposed
Ru = “risk” (IR, CI, P) among unexposed
• Also called etiologic fraction, attributable risk percent, attributable risk among exposed
Alternative formula
Again, we are interested in the difference between the risk
in the exposed and risk in the unexposed:
Divide numerator and denominator by Ru
APe = [(Re – Ru)/Re] * 100
APe = [(RR-1)/RR] * 100
(Different formulas may be helpful depending on what
information you have.)32
Note on alternative formula:
• Especially useful for case-control data
– You cannot estimate risk directly
– Can use OR to estimate RR under certain conditions.
– In which case:
APe = [(OR-1)/OR] * 100
33
Example(continued)
APe = [(14.6 – 7.9)/14.6] * 100 = 46%
• Interpretation: 46% of cases of breast cancer among those exposed
to radiation may be attributed to radiation exposure and could be
eliminated if exposure were removed.
Breast cancer
casesPY Rates/10,000 PY
Radiation exposure 41 28,010 14.6
No radiation exposure 15 19,017 7.9
Total 56 47,027 11.9
Attributable proportion among total population
• APt describes the proportion of disease among total population that would be eliminated if exposure were eliminated
• What percent of disease in total population is due to exposure
• Useful for setting priorities for public health action
– Elimination of exposure lead to what impact on population?
• Assumes causal relationship
APt = [(Rt – Ru)/Rt] * 100
= PRD / Rt * 100
Rt = “risk” (IR, CI, P) among total population
Ru = “risk” (IR, CI, P) among unexposed
• Also known as population attributable risk percent
Alternative formulas
APt = [1 - (Ru / Rt)] * 100
APt = [(Pe)(RR-1)]/[(Pe)(RR-1)+1] * 100
For case-control studies:
APt = [(Pe)(OR-1)]/[(Pe)(OR-1)+1] * 100
where Pe = proportion of exposed controls
(Convince yourself with some algebra if you like.)
36
Example(continued)
APt = [(11.9 – 7.9)/11.9] * 100 = 34%
• Interpretation:
– 34% of breast cancer cases in total study population may be attributed
to radiation exposure and could be eliminated if exposure were
removed.
Breast cancer
casesPY Rates/10,000 PY
Radiation exposure 41 28,010 14.6
No radiation exposure 15 19,017 7.9
Total 56 47,027 11.9
37
Key points
• Relative and absolute measures of comparison tell us
different things
– Relative risk is a measure of strength of association; often of
interest to epidemiologists who do etiologic research
– Absolute risk then becomes important (assuming causality) to
public health planners, policy makers, etc. for estimating public
health impact of exposures on communities
• Measures go by different names and multiple formulas
are sometimes available. It is important to
– know what you are interesting in measuring,
– know what data are available and in what form, and
– if data are not available, know how to collect appropriate data.