To Add
Make it clearer that excluding variables from a model because it is not “predictive” removes all meaning from a CI since this is infinite repetitions
Talk about stats being based on sampling variability – assumes sample is a random sample of some super-population (even if narrowly defined), but they are not a random sample, they self-select, so we couldn’t have infinite samples
Random Error I: p-values, confidence intervals, hypothesis
testing, etc.
Matthew Fox
Advanced Epidemiology
Do you like/use p-values?
What is a relative risk?
What is a pvalue?
Table 1 of a randomized trial of asbestos and LC
Factor Asbestos No Asbestos
pvalue
Female 10% 25% 0.032
Smoking 60% 40% 0.351
>60 yrs 5% 7% 0.832
HBP 25% 24% 0.765
Alcohol use 37% 45% 0.152
Which result is more precise?RR 2.0 (95% CI: 1.0 – 4.0)
RR 5.0 (95% CI: 2.5 – 10.0)
RR 2.0 (95% CI: 1.0 – 4.0)What are the chances the true results is between 1.0 and 4.0?
If yes, what does it mean to be “by chance?” What is it that is caused by chance?
In a randomized trial, could the finding be by change?
This Morning
Randomization– Why do we do it?
P-values– What are they?– How do we calculate them?– What do they mean?
Confidence Intervals
Last Session
Selection bias– Results from selection into our out of study related to
both exposure and outcome– Structural: conditioning on common effects– Adjustment for selection proportions– Weighting for LTFU
Matching– In a case control study, creates selection bias by
design, must be controlled in analysis
“There’s a certain feeling of ease and pleasure for me as a scientist that any way you slice the data, it’s statistically significant,” said Dr. Anthony S. Fauci, a top AIDS expert in the United States government, which paid most of the trial’s costs.
Randomization
Randomization lends meaning to likelihoods, p-values and confidence intervals
It can reduce the probability of severe confounding to an acceptable level
But randomization does not prevent confounding
Greenland: Randomization, statistics and causal inference
Objective: – Clarify the meaning and limitations of inferential
statistics in the absence of randomization
Example — lidocaine therapy after acute MI– Patient 1: doomed– Patient 2: immune– lidocaine therapy assigned at random– two results are equally likely
Greenland: Randomization, statistics and causal inference
True RD = 0, so both possible results are confounded Expectation = 0 = (1 + -1)/2
– Statistically unbiased (expectation equals truth)
Conclusions – Randomization does not prevent confounding– Randomization does provide a known probability distribution for
the possible results under a specified hypothesis about the effect– Statistical unbiasedness of randomized exposure corresponds to
an average confounding of zero over the distribution of results
Result 1 Result 2
lidocaine placebo lidocaine placebo
Patient 1 Patient 2 Patient 2 Patient 1
RD 1 - 0 = 1 0 – 1 = -1
Probability Theory
With an assigned probability distribution, can calculate expectation
The expectation does not have to be in the set of possible outcomes– Here, the expectation equals zero
Probability
If we randomize and assume null is true (as we do when calculating p-values) – We expect half of the subjects to be exposed and half
the events to be among the exposed If truly no effect of exposure, all data
combinations, permutations are possible– Everyone was either type 1 or 4– All the events (deaths) would occur regardless of
whether assigned the exposure or not
Probability Theory
The probability of each possible data result in a 2x2 table is:– A function of the number of combinations
(permutations implies order matters)– Probability of each event is number of ways to
assign X subjects to exposure out of Y and A events out of a total of B total events
– Assumes the margins are fixed
E+ E- TotalD+ ? ? 100D- ? ? 900Total 500 500 1000
Fixed margins, how many parameters (cells) do I need to estimate to fill in the entire table?
Greenland: Randomization, statistics and causal inference What comfort does this provide scientists
trying to interpret a single result? Can make probability of severe confounding
small by increasing the sample size
E+ E- Total
D+ 30 70 100
D- 470 430 900
Total 500 500 1000
Risk 0.06 0.14
RR 0.43 RD -0.08
Greenland: Randomization, statistics and causal inference
Given there were 100 cases and an even distribution of exposed and unexposed, how many cases would we expect to be exposed?
E+ E- Total
D+ 30 70 100
D- 470 430 900
Total 500 500 1000
Risk 0.06 0.14
RR 0.43 RD -0.08
Greenland: Randomization, statistics and causal inference What comfort does this provide scientists
trying to interpret a single result? Can make probability of severe confounding
small by increasing the sample size
0001.0
500
1000
500
900100
)cases 1000 in deaths 10030(
30
0
i
ii
XPProbability under the null that randomization would yield a result with at least as much downward confounding as the observed result
Back to the counterfactual
If association we measure differs from the truth, even if by chance, what explains it?– Unexposed can’t stand in for what would have
happened to exposed had they been unexposed This is confounding
– But on average, zero confounding– This gives us a probability distribution to calculate the
probability of confounding explaining the results This is a p-value
Randomized trial of E on D in 4 patients
We find:
E+ E-
D+ 2 0
D- 0 2
Total 2 2
Randomized trial of E on D in 4 patients
If the null is true, what CST types must they be?
E+ E-
D+ 2 0
D- 0 2
Total 2 2
Hypergeometric distribution The hypergeometric distribution:
Where X = random variable, x = exposed cases,
n = exposed population, M is = total cases, and
N = total population.
n
N
xn
MN
x
M
xXP
M!/x!(M-x)!
Spreadsheet
Greenland: Randomization, statistics and causal inference
When treatment is assigned by the physician, Expectation depends on physician behavior – Expectation does not necessarily equal truth
For observational data we DON’T have probability distribution for confounding– When E isn’t randomized, statistics don’t provide valid
probability statements about exposure effects because– p-values, CIs, & likelihoods calculated with assumption
all data interchanges are equally likely
Result 1 Result 2
lidocaine placebo lidocaine placebo
Patient 1 Patient 2 Patient 2 Patient 1
RD 1 - 0 = 1 0 – 1 = -1
Greenland: Randomization, statistics and causal inference
Alternatives– Limit statistics to data description (e.g., visual
summaries, tables of risks or rates, etc.)– Influence analysis: explore degree to which effect
estimates would change under small perturbations of the data, such as interchanging a few subjects
– Employ more elaborate statistical models– Sensitivity analysis– At the very least, interpret conventional statistics as
minimum estimates of the error
(1) The p-value is:
Probability under the test hypothesis (usually the null) that a test statistic would be ≥ to its observed value, assuming no bias in data collection or analysis– Why the null? Our job is to measure– 1-sided upper p-value is test stat ≥ observed value– 1-sided lower p-value is test stat ≤ observed value– Mid-p assigns only half probability of the observation
to the 1-sided upper p-value– 2-sided p-value is twice the smaller of the 1-sideds
(2) The p-value is not:
Probability that a test hypothesis (null hypothesis) is true– Calculated assuming that test hypothesis is true. – Cannot calculate probability of an event that is
assumed in the calculation
Probability of observing the result under the test hypothesis (null) [likelihood]– Also includes probability of results more extreme
(3) The p-value is not:
An -level (the Type 1 error rate)– More on that later
A significance level– Used to refer to both p-values and Type 1
error rates– Should be avoided to prevent confusion
(4) The 2-sided p-value is not:
Because 2-sided p-value is twice smaller of lower and upper 1-sided p-values, which may not be same and may be > 1, it is not the:– Probability that the data would show as
strong an association as observed or stronger if the null hypothesis were true;
– Probability that a point estimate would be as far or further from the test value as observed
Significance testing:
Compares p-value to an arbitrary or conventional Type 1 error rate =0.05
Emphasizes decision making, not measurement– Derives from agricultural and industrial
applications of statistics– Reflects the roots of epidemiology as the
union of statistics and medicine
JNCI announces materials to “help journalists get it right”
Response
They acknowledge the definitions were incorrect, however:
“We were not convinced that working journalists would find these definitions user-friendly, so we sacrificed precision for utility. We will add references to standard textbooks for journalists who want to learn more.”
Frequentist Statistics
Alternatives to pvalues
Two studies which is more precise?– RR 10.0, p = 0.039– RR 1.3, p = 0.062
The pvalue conflates the size of the effect and its precision– RR 10.0, p = 0.039, 95% CI: 1.5-66.7– RR 1.3, p = 0.062, 95% CI: 0.99-1.7
Frequentist intervals (1)
Definition: – If the statistical model is correct and no bias, a
confidence interval derived from a valid test will, over unlimited repetitions of the study, contain the true parameter with a frequency no less than its confidence level (e.g. 95%).
But the statistical model is only correct under randomization
CAN’T say that the probability the interval includes the truth equals the interval’s coverage probability (e.g., 95%).
Confidence Interval Simulation
Frequentist intervals (2)
Advantages– Provides more information than significance tests
or p-values: direction, magnitude, and variability– Economical compared with p-value function
Disadvantages– Less information than the p-value function– Underlying assumptions (valid statistical model, no
bias, repeated experiments)
Approximations: Test-based
)1(ˆ
)]ˆ[ln(
)ˆln(
z
RRCI
RRSE
RR
)1(ˆ
)ˆ(
ˆ
zDRCI
DRSE
DR
Approximations: Wald
DRSEzDRCI
eCIRRSEzRR
ˆˆ
)ˆln()ˆln(
Standard Errors (basic over i strata):
i ii
ii
i ii
i
ii
i
idcba
ba
baRIR
Nb
d
Na
cRR
ROiiii
ˆlnvar
ˆlnvar
ˆlnvar
01
1111
How do we measure precision?
Width of the confidence interval Measured how?
– If I tell you the 95% CI for an RR is 2 to 8, can you tell me the point estimate?
– Sqrt(U*L) – Difference measures, just subtract
Remember relative measures are on the log scale, so width of a CI is measured by the RATIO of the upper to the lower CI
Frequentist Intervals (4): Interpretation
Conclusion about confidence intervals
A CI used for hypothesis testing is an abuse of the CI– The goal is precision, not significance
The goal of epi is precision, not significance– A precise null estimate is just as important as a
precise significant estimate– An imprecise, statistically significant estimate is as
useless as a non-statistically significant, imprecise estimate
What results are published or highlighted in publications?
Find a publication with multiple results Rank them in order of precision Then see what is highlighted in the abstract
Maternal age <35 Maternal age >= 35Parity OR LCL95 UCL95 width rank OR LCL95 UCL95 width rank
0 1 11 1.08 0.86 1.35 1.57 1 1.11 0.65 1.88 2.89 62 1.14 0.87 1.49 1.71 2 1.67 0.99 2.82 2.85 53 1.65 1.13 2.4 2.12 3 1.24 0.69 2.23 3.23 8
>=4 1.45 0.93 2.26 2.43 4 2.41 1.41 4.12 2.92 7
Maternal age <35 Maternal age >= 35Parity OR LCL95 UCL95 width rank OR LCL95 UCL95 width rank
0 1 11 1.08 0.86 1.35 1.57 1 1.11 0.65 1.88 2.89 62 1.14 0.87 1.49 1.71 2 1.67 0.99 2.82 2.85 53 1.65 1.13 2.4 2.12 3 1.24 0.69 2.23 3.23 8
>=4 1.45 0.93 2.26 2.43 4 2.41 1.41 4.12 2.92 7
CIPRA Trial
Trial of nurse vs. Doctor managed HIV care For primary results, co-investigators wanted
pvalues and confidence intervals Didn’t want hypothesis testing even though
was aware people would do it anyway I fought initially, and lost the debate Put in both
CIPRA Trial
Reviewer comment:
Table 3: Column with p-values can be dropped given that 95% confidence intervals are presented;
perhaps mark significance as * (e.g. for p<0.025) and ** (e.g. for p<0.005) after the 95% CI's.
Summary
Randomization gives meaning to statistics– Gives a probability distribution for confounding
When randomization doesn’t hold, we have no probability distribution
Pvalues aren’t probability of chance, null, etc. CIs allows us to assess precision
– But are based on infinite repetitions– Do not contain the true value with 95% probability