A list a statistical definition for the CICM Fellowship exam

2019.2

Stratification: the process of assigning subjects to groups based on factors other than that of the treatment. This
protects against imbalance of cofounding factors at the time of randomization. Eg: knowing that weight is a
contributing factor to heart disease, one with attempt to balance the BMI of treatment of control groups when
testing a new drug treatment

Minimization algorithm: A form of stratification whereby assignment of subjects are ‘weighted’ by a algorithm
based on the pre-existing composition of several prognostic variables/cofounders of the groups. Eg if all else being
equal, if the BMI is 25 for group 1 and 15 for group two, the algorithm will assign a subject of BMI 30 to group 1 in an
attempt to minimize the difference in BMI due to it being a confound.

Block randomization: Assignment of a number of patients to groups all at once rather than an single individual in
order to minimize the difference in sample size. Eg instead of assignment patient 1 to group A, pt 2 to group A, pt 3
to group A, pt 4 to group A, pt 5 to group B, pt 6 to group A sequentially as they get enrolled thus leading to unequal
group sizes, they are assigned as a whole block so that three will go to group A and three to group B randomly.

Blinding: process which participants data collectors and investigators are unaware which group subjects are assigned
to. This prevents detection, observer and recall bias. There is a recent shift away from using the terms ‘single, double
or triple blinded’ those terms are not well defined’ (2010 CONSORT statement)

Allocation concealment: the assignment to groups are blinded to the clinical investigators and those involved to
minimize selection bias. This is usually done via online algorithms/pre-sealed envelopes.

Intention to treat: Statistical analysis of groups based on their initial assignment, irrespective of whether then
completed the treatment, withdrew, crossed over or deviated from protocol. Minimises type 1 false positive error
and increases validity of the study.

Sensitivity analysis: Statistical analysis whereby the assumptions within the study are deliberately changed to see
how the outcomes are affected. Improves external validity and power of the study. Eg instead of assuming all
subjects who dropped out of the treatment group survived, how does it affect the efficacy of the drug if instead they
were classified as dead?

Kaplan-Meir curve: A plot of percentage survival over time, with time measured in set intevals

ROC received operative curve: plot of True positive (sensitivity) on the y axis and false positive (1-specificity) on the
x axis. All tests require some form of cut-off/threshold value, and this value is a balance between sensitivity (how
likely it is to detect the disease) and specificity (without being too sensitive and generating too many false positives).
A good test will have high AUC such that it has high sensitivity and specificity (value of 1-specificity is therefore low)

https://towardsdatascience.com/understanding-auc-roc-curve-68b2303cc9c5
Analysis of completing risk: An event that influences the chance of the event of interest occurs. Eg someone dies
from a heart attack during the study of the need for CRRT after FST

SMR: ratio of observed deaths in the study group to expected deaths in the general population based on APACHE or
other severity of illness over a ser time interval. Part of a KPI to measure quality of ICU. Can someone expand this?

Exposure+ Exposure -
Disease + A C
Disease - B D

Prevalence: The percentage of the population that has the disease (A+C)/(A+B+C+D)

Relative risk/Risk ratio: The ratio of probability of the event occurring in the exposed group vs the probability of
event occurring in the control group. [A/(A+B)] / [C/(C+D)]

Attributable risk: The difference between the probability of the event occurring in the exposed group and the
probability in the control group [A/(A+B)] - [C/(C+D)]

Odds ratio: The ratio of odds between the treatment groups. [A/B] / [C/D]. Note that this is slightly different to
RR but is approximately the same in large samples.

Eg:

Smoking+ No smoking
Has lung cancer 10 5
No lung cancer 90 95

The difference between relative and attributable risk can be thought of this way. 10% of those in the smoking group
gets cancer, but only 5% of those who don’t smoke gets cancer. Those who smoke are 2 times more likely to get
cancer (relative risk) compared to those who don’t. But it’s only a 5% increase in (attributable) risk. The odds ratio is
2.11
Absolute risk: The actual event rate in the treatment or placebo group

A good rule of thumb is that ‘absolute’ means then difference, while ‘relative’ means ratio

Power of the study: The probability (normally 80%) that a study will detect the difference between the placebo and
treatment group and therefore reject the null hypothesis. It is affected by the magnitude of the effect size, the
statistical significant (usually aimed at 0.05), variance within the population, and will determine the sample size
required. Eg the smaller the effect size, the larger the sample population is needed to because the effect is so small.
The higher the power, the larger the sample size, The lower the power, the higher the false negative rate.
 Types of plots

Forest plot

Position of square is the point estimate of each study.

Horizontal lines are the 95% confidence intervals or the odds ratio
Size of the square indicate the weight of the study
The diamond is the aggregate of the meta-analysis. Its accuracy will depend on
Definition of research question
Definition of inclusion criteria for studies
Adequate search protocol
Assessment of methodological quality
Measurement of heterogeneity
Assessment of publication bias, collection of unpublished results

Used in meta analysis (a form of systematic review that uses statistical methods to combine the results from
different studies)
Improves statistical power by increasing sample size
Reduces false negative
Resolves uncertainty when studies disagree
Improves point estimates
Establish questions for future RCTs
 Funnel plot

Plot of standard error and their odds ration. Larger studies (and therefore smaller errors) are higher. Consequently
their 95% confidence interval (diagonal lines) are less spread apart. 1 shows no difference.

ERRORS
Type 1 FP error: The null hypothesis is falsely rejected. Finding treatment effect, when in true there isn’t any

Type 2 FN error: The null hypothesis is falsely accepted. Finding no treatment effect, when in truth there is one