#
35#
common features in great works of art (e.g. eastern promises):
1. icc is metaphor for the patriachal-hierarchical society such as russian mafia
2. our protagonists tattooes catalog his journey (intercept & slope)
3. this
will protect you or have your throat slitbut this isn’t all that’s common: watch headhunters, and generalize (see #43 below)
36#
epistemolgy
time: intercept, slope
identity (no deus ex machina, please!)
37#
identity -> variance/explanation -> outcomes
causal inference deals with identity -> outcomes
reflects influence of epidemiology
thus association is not causation dictum
black box is always present
but the hierarchial approach is more robust
reflects influence of statistics
actually explains variance in outcomes
no black box, which is seen with epidemiology
rct’s are the gold standard for causal inference
but they are expensive and time consuming
so we use observational studies
but they are prone to confounding
so we use propensity score matching
but they are prone to selection bias
so we use instrumental variables
but they are prone to endogeneity
so we use structural equation modeling
but they are prone to omitted variable bias
so we use bayesian networks
but they are prone to overfitting
so we use regularization
but they are prone to bias-variance tradeoff
so we use ensemble methods
but they are prone to overfitting
so we use bagging
but they are prone to overfitting
so we use boosting
but they are prone to overfitting
so we use stacking
but…
38#
bias-variance tradeoff is a fundamental tradeoff in machine learning
one is the error due to the model’s inability to capture the true relationship
variance is the error due to the model’s sensitivity to small fluctuations in the training set (ie its modelling the noise)
39#
hierarchical-patriachical approach to data analysis
why is it so important to understand the data generating process?
because we can structure the variance in the data & thus explain the outcomes
40#
process (hierarchical) vs. variance (assuming icc = 0)
fully sociological process (hierarchical) vs. fully biological process (variance)
will-to-power, domination, control (hierarchical: intra-class correlation)
biology, freedom, autonomy (intra-class variance)
also:
power/bias/control
vs.weakness/variance/explanation
explains why some folks intuitive “hate” apple products
Ah, if you’re interested in analytic decisions rather than raters’ reliability, the use of ICC takes on a different flavor. The decision to use hierarchical models or ordinary regression largely depends on the structure of your data and the specific research questions you are asking.
Hierarchical models are particularly useful when your data has multiple levels of grouping, such as patients within hospitals or repeated measurements on the same subjects over time. They allow you to explicitly model variance at each level of the hierarchy, which can give you deeper insights into the data. If the ICC is close to zero, this would suggest that the hierarchical structure might not be providing much information beyond what you’d get with an ordinary regression.
So, given that your interest lies in analytic decisions and not in rater reliability, an ICC of 0 could actually simplify your analytic strategy. It suggests that treating your data hierarchically may not provide additional benefits over a simpler ordinary regression model.
41
#
icc has brought us to the end of the road:
we can’t go any further with the data
need to go back to the data generating process
to understand the process that generated the data
this is where we reunite with counterfeiting!
42#
\( \Large \left\{ \begin{array}{ll} \text{Truth/Free/Agent} \\ \text{} \\ \text{Know/John 8:32/Verb} \ \ \left\{ \begin{array}{l} \text{Rigor} \text{} \\ \text{Error} \ \ \ \ \ \ \ \ \ \left\{ \begin{array}{l} \text{Variance} \\ \text{Bias} \end{array} \right. \\ \text{Sloppy} \end{array} \right. \left\{ \begin{array}{l} \text{Information/Control} \end{array} \right. \\ \text{} \\ \text{Morality/Slave/Object} \end{array} \right. \)
explain/control
literally the end of the road
we can’t go any further with the data
need to go back to the data generating process (rct at infinitely many levels of individual & group)
and therein we find that two roads diverged in a wood (truest approach unbiased, but v.imprecise)
one is the road of truth and the other is the road of fraud (selective, emotive reporting, p-hacking, compelling)
43#
come full circle these 43 years:
philosophy (explo85) 0-5
knowledge (student) 6-35
error (faculty) 36-43
icc (gtpci) 40-44
back-at-one (tenure) 45-50
truth; amoral (one is trained to treat ephors as gods; remove error of man)
fraud; moral (spartan ephors in 300; like magesterium in vatican)