+ - 0:00:00
Notes for current slide
Notes for next slide

Eliminating Accidental Deviations in Human Connectomics

Eric W. Bridgeford | ericwb.me

1 / 52

What is Reproducibility?

  • Reproducibility: ability to replicate, or reproduce, a conclusion
  • serves as a "first-pass" check for scientific utility
  • currently in a "reproducibility crisis"
4 / 52

How do we address the Reproducibility Crisis?

  • fix post hoc analyses (e.g., pp-values)?
  • fix measurements (e.g., measurement reproducibility)?

Proposal: design experiments to maximize inter-item discriminability, rather than simply checking reproducibility after conducting the experiment

5 / 52

Outline

Additional Content

6 / 52

What do we want of our data?

If we measure a sample multiple times, then each measurement of that sample is closer to all the other measurements of that sample, as compared to any of the measurements of other samples.

Perfect discriminability

7 / 52

What do we want of our data?

Imperfect discriminability

8 / 52

What do we want of our statistic?

Discriminability is the probability of a measurement from the same item being closer than a measurement from a different item.

9 / 52

Discriminability Statistic: Step 1

  • Compute N×NN \times N pairwise distance matrix between all measurements

10 / 52

Discriminability Statistic: Step 2

  • For each measurement, identify which measurements are from the same individual (green boxes)
  • let g\color{green}g be the total number of green boxes = 20

11 / 52

Discriminability Statistic: Step 3

  • For each measurement, identify measurements from other individuals that are more similar than the measurement from the same individual (orange boxes)
  • let f\color{orange}f be the total number of orange boxes = 84

12 / 52

Discriminability Statistic

  • Discr = 1fN(N1)g=184201920.771 - \frac{\color{orange}f}{N(N-1) - \color{green}g} = 1 - \frac{\color{orange}{84}}{20\cdot 19 - \color{green}{20}} \approx .77

High discriminability: same-item measurements are more similar than across-item measurements

13 / 52

Discriminability is Construct Valid

14 / 52

What data will we be using?

  • CoRR metadataset
  • N>1,700N>1,700 individuals imaged across 2626 different datasets
    • anatomical MRI and fMRI scans for each
    • Individuals are measured at least twice
16 / 52

Analysis Procedure

Process each measurement using 192192 different pipelines

  1. Brain alignment (ANTs/FSL)
  2. Frequency filtering (Y/N)
  3. Scrubbing (Y/N)
  4. Global Signal Regression (Y/N)
  5. Parcellation (4 options)
  6. Rescaling connectomes (Raw, Log, Pass-to-Rank)

192=2×2×2×2×4×3192 = 2 \times 2 \times 2 \times 2 \times 4 \times 3

All options represent strategies experts consider useful

17 / 52

Pipeline impacts discriminability

18 / 52

Marginally most discriminabile options tend to be best global options

  • Each point is the pairwise difference holding other options fixed (e.g., FNNGCP - ANNGCP)
  • Best pipeline marginally (FNNGCP) is second best pipeline overall, and not much worse (2-sample test, p=.14) than the best pipeline FNNNCP
  • We may not need to always try every pre-processing strategy every time
19 / 52

Selection via Discriminability improves inference

For each pre-processing strategy, for each dataset, compute:

  1. Within-dataset Discr.
  2. Demographic effects (sex and age) within the dataset via Distance Correlation (DCorr)
  3. Within a single dataset, regress demographic effect on Discr.

Question: does a higher discriminability tend to yield larger effects for known biological signals?

20 / 52

Selection via Discriminability improves inference

21 / 52

Contributions

  1. Discriminability quantifies the contributions of systematic and accidental deviations
  2. Provide theoretical motivation for discriminability in connection with predictive accuracy
  3. Formalize tests for assessing and comparing discriminabilities within and between collection strategies
  4. Illustrate the value of discriminability for neuroscience and genomics (not discussed) data
  5. Code implementations in python and R
23 / 52

Acknowledgements

Josh Vogelstein
Shangsi Wang
Zhi Yang
Zeyi Wang
Ting Xu
Cameron Craddock
Jayanta Dey
Greg Kiar
William Gray-Roncal
Carlo Colantuoni
Christopher Douville
Stephanie Noble
Carey Priebe
Brian Caffo
Michael Milham
Xinian Zuo

24 / 52

Population Discriminability

  • population discriminability DD is a property of the distribution of measurements D=P(δ(xik,xik)<δ(xik,xjk))D = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}))
  • Probability of within-individual measurements being more similar than between-individual measurements
27 / 52

Discriminability: unbiased and consistent

  • Sample Discr. == fraction of times δ(xik,xik)<δ(xik,xjk)\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''})
  • i,j=1,...,ni, j = 1, ..., n and iji \neq j for nn individuals
  • k,k,k=1,...,sk, k', k'' = 1, ..., s and kkk \neq k' for ss sessions
  • Sample Discr. is an unbiased estimator of DD
  • Sample Discr. converges to DD (asymptotically consistent in nn)
28 / 52

Connecting Discriminability to Downstream Inference

Assumption: Multivariate Additive Noise Setting

  • yiBern(π)i.i.d.y_i \sim Bern(\pi)\;i.i.d.,
  • θiN(μ(yi),Σt)ind.\theta_i \sim \mathcal N(\mu(y_i), \Sigma_t)\;ind.,
  • (the individual means have a center which depends on the class)
29 / 52

Connecting Discriminability to Downstream Inference

Assumption: Multivariate Additive Noise Setting

  • yiBern(π)i.i.d.y_i \sim Bern(\pi)\;i.i.d.,
  • θiN(μ(yi),Σt)ind.\theta_i \sim \mathcal N(\mu(y_i), \Sigma_t)\;ind.,
  • (the individual means have a center which depends on the class)
  • ϵikN(c,Σe)i.i.d.\epsilon_{i}^k \sim \mathcal N(c, \Sigma_e)\;i.i.d. and ind.ind. of θi\theta_i,
  • xik=θi+ϵikx_{i}^k = \theta_i + \epsilon_i^k.
  • (the measurements xikx_i^k are normally dispersed about the individual means)
30 / 52

Connecting Discriminability to Downstream Inference

Suppose (xik,yi)(x_i^k, y_i) follow the Multivar. Additive Noise Setting, where i=1,...,ni=1, ..., n and k=1,...,sk=1,...,s.

Theorem 1

There exists an increasing function of DD, f(D)f(D), which provides a lower bound on the predictive accuracy of a subsequent classification task

  • f(D)Af(D) \leq A, where AA is the Bayes Accuracy of the classification task

Consequence

  • Df(D)D \uparrow \Rightarrow f(D) \uparrow
31 / 52

Corollary 2

A strategy with a higher DD provably provides a higher bound on predictive accuracy than a strategy with a lower DD

Consequence

Suppose D1<D2D_1 < D_2, then since ff is increasing, f(D1)<f(D2)f(D_1) < f(D_2)

Implication

We should use strategies with higher discriminability, as the worst-case for subsequent inference is better than a generic strategy with a lower discriminability

32 / 52

Simulation Setup

33 / 52

Discriminability and Accuracy

Discr. decreases proportionally with accuracy

34 / 52

Are data discriminable?

35 / 52

Is one dataset more discriminable than another?

36 / 52

Outline

Additional Content

37 / 52

Intraclass Correlation Coefficient (ICC)

  • can be thought of as looking at the "relative size" of the within-group vs total variance
  • yik=μ+μi+ϵiky_i^k = \mu + \mu_i + \epsilon_i^k
  • let μiN(0,σb2)\mu_i \sim \mathcal N(0, \sigma_b^2), and ϵikN(0,σe2)\epsilon_i^k \sim \mathcal N(0, \sigma_e^2)
  • ICC=σb2σe2+σb2ICC = \frac{\sigma_b^2}{\sigma_e^2 + \sigma_b^2}
  • ICCICC \uparrow \Rightarrow between-group variance "contains" most of the total variance
  • negative ICC? mean squared error-based estimator
38 / 52

Image Intraclass Correlation Coefficient (I2C2)

  • simplest "multivariate extension" of ICC
  • yik=μ+μi+ϵiky_i^k = \mu + \mu_i + \epsilon_i^k
  • let μN(0,Σb)\mu \sim \mathcal N(0, \Sigma_b) and ϵikN(0,Σe)\epsilon_i^k \sim \mathcal N(0, \Sigma_e)
  • Wilk's Λ=det(Σb)det(Σb)+det(Σe)\Lambda = \frac{\det(\Sigma_b)}{\det(\Sigma_b) + \det(\Sigma_e)}

  • I2C2=tr(Σb)tr(Σb)+tr(Σe)I2C2 = \frac{tr(\Sigma_b)}{tr(\Sigma_b) + tr(\Sigma_e)}

  • "ratio of total variability accounted for between groups"

  • Why I2C2 over Wilk's Λ\Lambda? Ease-of-use for high-dimensional data
39 / 52

Fingerprinting Index (Finger.)

  • "greedy discriminability"
  • Finger.=P(δ(xi1,xi2)<δ(xi1,xj2)ij)Finger. = \mathbb P(\delta(x_i^1, x_i^2) < \delta(x_i^1, x_j^2) \;\forall\; i \neq j)
  • ij\forall\; i \neq j: this property must occur for every measurement in the second session
40 / 52

Distance Components (Kernel)

  • "non-parametric ANOVA"
  • total dispersion is the sum of between and within-sample dispersions (BB and WW)

  • DISCO=Bn1WnsnDISCO = \frac{\frac{B}{n - 1}}{\frac{W}{n\cdot s - n}}

  • "pseudo F" statistic

41 / 52

Limitations

  • experimental design is not "one-size-fits-all"
  • Discriminability is not sufficient for practical utility
    • categorical covariates are meaningful but not discriminable
    • fingerprints are discriminable but not typically biological useful
  • These statistics are not immune to sample characteristics
    • confounds such as age may inflate discriminability
43 / 52

Outline

Additional Content

44 / 52

Extension: Discriminability Decomposition

Setting

(xik,yi)(x_{i}^k, y_i) i=1,...,ni=1, ..., n, k=1,...,sk=1,...,s, yiy_i \in {1,...,Y1, ..., Y}

  • associated with each individual, I have some other categorical covariate of interest, yiy_i, taking one of YY possible values
  • Can the population discriminability be decomposed as a sum of the within-group discriminabilities?
45 / 52

Within-Group Discriminability

  • Let D(y)=P(δ(xik,xik)<δ(xik,xjk)yi,yj=y)D(y) = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i, y_j = y)
  • D(y)D(y) is the group discriminability for group yy
  • "How discriminable are samples from group yy?"
46 / 52

Within-Group Discriminability

  • Let D(y)=P(δ(xik,xik)<δ(xik,xjk)yi,yj=y)D(y) = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i, y_j = y)
  • D(y)D(y) is the group discriminability for group yy
  • "How discriminable are samples from group yy?"
  • Note that W=P(δ(xik,xik)<δ(xik,xjk)yi=yj)W = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i= y_j)= P(δ(xik,xik)<δ(xik,xjk),yi=yj)P(yi=yj)\frac{\mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) , y_i = y_j)}{\mathbb P(y_i = y_j)} by def conditional probability
47 / 52

Within-Group Discriminability

  • Let D(y)=P(δ(xik,xik)<δ(xik,xjk)yi,yj=y)D(y) = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i, y_j = y)
  • D(y)D(y) is the group discriminability for group yy
  • "How discriminable are samples from group yy?"
  • Note that W=P(δ(xik,xik)<δ(xik,xjk)yi=yj)W = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i= y_j)= P(δ(xik,xik)<δ(xik,xjk),yi=yj)P(yi=yj)\frac{\mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) , y_i = y_j)}{\mathbb P(y_i = y_j)} by def conditional probability

  • Let w(y)=P(yi=yj=y)w(y) = \mathbb P(y_i=y_j = y) denote the within-group weights

  • With ω=yw(y)\omega = \sum_y w(y), then:

    W=1ωyw(y)D(y)W = \frac{1}{\omega}\sum_y w(y) D(y) is the within-group Discriminability

48 / 52

Between-Group Discriminability

  • Let D(y,y)=P(δ(xik,xik)<δ(xik,xjk)yi=y,yj=y)D(y, y') = P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i = y, y_j = y')
  • D(y,y)D(y, y') is the between-group discriminability for groups yy and yy'
  • "How discriminable are samples from group yy vs group yy', and vice versa?"
49 / 52

Between-Group Discriminability

  • Let D(y,y)=P(δ(xik,xik)<δ(xik,xjk)yi=y,yj=y)D(y, y') = P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i = y, y_j = y')
  • D(y,y)D(y, y') is the between-group discriminability for groups yy and yy'
  • "How discriminable are samples from group yy vs group yy', and vice versa?"

  • Note that B=P(δ(xik,xik)<δ(xik,xjk)yiyj)B = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i\neq y_j)= P(δ(xik,xik)<δ(xik,xjk),yiyj)P(yiyj)\frac{\mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) , y_i \neq y_j)}{\mathbb P(y_i \neq y_j)} by def conditional probability

50 / 52

Between-Group Discriminability

  • Let D(y,y)=P(δ(xik,xik)<δ(xik,xjk)yi=y,yj=y)D(y, y') = P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i = y, y_j = y')
  • D(y,y)D(y, y') is the between-group discriminability for groups yy and yy'
  • "How discriminable are samples from group yy vs group yy', and vice versa?"

  • Note that B=P(δ(xik,xik)<δ(xik,xjk)yiyj)B = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) | y_i\neq y_j)= P(δ(xik,xik)<δ(xik,xjk),yiyj)P(yiyj)\frac{\mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}) , y_i \neq y_j)}{\mathbb P(y_i \neq y_j)} by def conditional probability

  • Let b(y,y)=P(yi=y,yj=y)b(y, y') = \mathbb P(y_i = y, y_j = y') denote the between group weights

  • With β=yyb(y,y)\beta = \sum_{y\neq y'} b(y,y'), then:

B=1βyyb(y,y)D(y,y)B = \frac{1}{\beta}\sum_{y \neq y'}b(y,y')D(y,y') is the between-group Discriminability

51 / 52

Discriminability Decomposition

  • D=ωW+βBD = \omega W + \beta B
  • Population discriminability is a weighted sum of within and between-group Discriminabilities
  • Can look at how the within, or between, group discriminabilities compare
  • WD\frac{W}{D} ratio of within-group Discriminability and pop. discriminability
  • BD\frac{B}{D} ratio of between-group Discriminability and pop. discriminabillity
  • are certain groups more discriminable than others?
  • are certain between-group discriminabilities greater than others?
  • "ANOVA-esque" or DISCO-esque"
52 / 52
Paused

Help

Keyboard shortcuts

, , Pg Up, k Go to previous slide
, , Pg Dn, Space, j Go to next slide
Home Go to first slide
End Go to last slide
Number + Return Go to specific slide
b / m / f Toggle blackout / mirrored / fullscreen mode
c Clone slideshow
p Toggle presenter mode
t Restart the presentation timer
?, h Toggle this help
Esc Back to slideshow