P-value Hell

Viktor Petukhov

viktor.s.petuhov@ya.ru

 

BRIC, University of Copenhagen

Khodosevich lab

Table of content

  1. Motivation
  2. What is p-value?
  3. History of p-values
  4. What's wrong with p-values?
  5. How to live without p-values?
  6. How to live when everyone uses p-values?

Standard workflow

P=0.015

Significant!

The observed decrease in PV levels and synaptic contacts might indicate impaired maturation of PV+ interneurons.

What's that?

How did we understand?

Data

Statistics

P-value

Conclusion

What is p-value?

Comparison of means

Mice on drugs

Average weight: 215g

Mice without drugs

Average weight: 205g

Difference: -10g

Comparison of means

Hypothesis 0: single group, no difference

Comparison of means

Hypothesis 0: single group, no difference

Difference: 3g

Comparison of means

Hypothesis 0: single group, no difference

Difference: -5g

Comparison of means

p(difference | H0)

=

False Positive Rate

=

P-value

 

P-value

Is it good?

P=0.015

Significant!

The observed decrease in PV levels and synaptic contacts might indicate impaired maturation of PV+ interneurons.

Data

Statistics

P-value

Conclusion

History of p-values

Two conflicting theories

K. Pearson

R. Fisher

Two conflicting theories

K. Pearson

R. Fisher

Long-term strategy

Power of single evidence

*Lehmann E. The Fisher, Neyman-Pearson theories of testing hypotheses: one theory or two? Journal of the American Statistical Association. 1993;88: 1242-9.

What's wrong with

p-values?

Probability to meet a wizard

95%

5%

100%

Actual probability of an error

95%

5%

100%

H_0
H_1

Difference

Difference

Difference

No

Difference

P-value

Actual probability of an error

*Regina Nuzzo. Scientific method: Statistical errors. doi:10.1038/506150a

Hypotheses don't represent underlying models

Cells [...] were markedly less bright than [...]. Thus, the MEF2-binding site might set steady-state levels of Cox6a2 expression and E-box fine tunes the specificity.

P-value is not a 100% proof of the conclusion

P-value:

0.015

Significant!

The observed decrease in PV levels and synaptic contacts might indicate impaired maturation of PV+ interneurons.

Online daters do better in the marriage...

PNAS, 2013

Study on more than 19,000 people:

those who meet their spouses online are less likely to divorce (p < 0.002) and more likely to have high marital satisfaction (p < 0.001) than those who meet offline

*J. Cacioppo, et. al. Marital satisfaction and break-ups differ across on-line and off-line meeting vensdues.  https://doi.org/10.1073/pnas.1222447110

Regina Nuzzo. Scientific method: Statistical errors. doi:10.1038/506150a

Divorce rate:

7.67% vs 5.96%

Happiness:

5.48 vs 5.64

Can we compare p-values?

P-value: 0.02

Previous study:

P-value: 0.001

Your study:

Significant p-value VS Common Sense

Sources:

  • Outliers
  • Broken test assumptions
  • Modification of data, which results in significant difference (p-hacking)
  • Bad luck

Summary

  • No information about size of effect
  • No way to aggregate p-values across several studies
  • No way to integrate prior knowledge
  • No way to estimate real probability of an error
  • Mistakenly considered as a 100% proof
  • Hypotheses, which we test, are weakly connected to real models
  • Easy to fool yourself

Solution:

Bayesian models

Two types of analysis

Evidence

Evidence

Evidence

...

Evidence

Model

Evidence

Evidence

Evidence

...

Evidence

Exploration

_



 

_

 


 

Confirmation

Experiment 1

Experiment 2

Two types of analysis

Evidence

Evidence

Evidence

...

Evidence

Model

Evidence

Evidence

Evidence

...

Evidence

Exploration

_



 

_

 


 

Confirmation

Experiment 1

Experiment 2

Don't care about significance

Don't use p-values

Model vs hypothesis

Hypothesis:

Linear regression has non-zero slope

Model:

Dendrite length ~ S * Age + Noise

Noise ~ Normal(mean=0, std=1)

Prior knowledge:

S ~ Normal(mean=0.2, std=0.1)

Model vs hypothesis

Model 1:

Length ~ S * Age + Noise

Noise ~ Normal(0, 1)

Prior probability: p0=0.1

Model 2:

Length ~ m + Noise

Noise ~ Normal(0, 1)

Prior probability: p0=0.9

Validation of a model

*https://www.wavemetrics.com/products/igorpro/dataanalysis/curvefitting

Residuals and confidence band

Validation of a model

Predictive power

Predictions

Validation of a model

Train-test split / Cross-Validation

Two types of analysis

P-values Bayesian
No information about effect size Effect size is fitted by a model
No way to aggregate p-values across several studies Hierarchical models
No way to integrate prior knowledge Prior probabilities
No way to estimate error probability Prior probabilities
Mistakenly considered as a 100% proof Gives "goodness of fit", but not a binary answer
Hypotheses, which we test, are weakly connected to real models Can use very complex models
Easy to fool yourself More transparent  system with priors (but still you can do it)

How to live in the p-value world

Logo of wrong statistics

For normal distribution:

  • Std show effect size
  • SE allows to validate significance of p-values

For non-normal distribution:

  • Std means nothing
  • SE means nothing

Proper visualization

Confidence intervals

Small data

Big data

Multiple comparison

adjustment

Multiple comparison

adjustment

Avoid selective reporting

  1. Predetermine rule for publishing of the data and results
     
  2. Publish this rule
     
  3. Publish all data according to this rule
     
  4. Publish all manipulations and all measures in the study

Avoid selective reporting

Validation: P-curve

Avoid selective reporting

Validation: P-curve

Summary

Problem Solution
No information about effect size Better reporting (e.g. swarmplots with confidence intervals)
No way to aggregate p-values across several studies Adjustment for multiple comparisons
No way to integrate prior knowledge -
No way to estimate error probability -
Mistakenly considered as a 100% proof Keep in mind: p-value is just an evidence. Rely on common sense.
Hypotheses, which we test, are weakly connected to real models Use better hypotheses, learn underlying assumptions
Easy to fool yourself Predetermine rule for publishing and follow it

References

  • Nuzzo, R. (2014), “Scientific Method: Statistical Errors,” Nature, 506, 150–152. doi:10.1038/506150a
  • Goodman, S. N. (1999). Toward evidence-based medical statistics: I. The p value fallacy. Annals of Internal Medicine, 130, 995–1004.  doi:130(12):995-1004
  • Greenland, S., Senn, S.J., Rothman, K.J., Carlin, J.B., Poole, C., Goodman, S.N. and Altman, D.G.: “Statistical Tests, P-values, Confidence Intervals, and Power: A Guide to Misinterpretations.
     
  • Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014). P-curve: A key to the file-drawer. Journal of Experimental Psychology: General, 143(2), 534–547. doi:10.1037/a0033242
  • Ronald L. Wasserstein & Nicole A. Lazar (2016) The ASA's Statement on p-Values: Context, Process, and Purpose, The American Statistician, 70:2, 129-133, DOI: 10.1080/00031305.2016.1154108

Thank you!

Viktor Petukhov

University of Copenhagen

Khodosevich lab

viktor.s.petuhov@ya.ru

P-value Hell

By Viktor Petukhov

P-value Hell

  • 490
Loading comments...

More from Viktor Petukhov