Philosophy of science Flashcards
Experts
In a specified domain have a greater quantity of accurate information than most people do
Laypeople (novices):
Little information in the specified domain
The novice-experts problem
How should novices choose one putative expert as more credible or trustworthy than another
Possible strategies to adress the novice experts problem (arguments presented)
Advantage: Information from putative experts is widespread and easily available
Problem: How can a novice make an accurate assessment of the putative experts arguments and technical language
Possible strategies to adress the novice experts problem (agreement with other experts
Advantage: For any domain, there is typically more than one expert, and the great majority of experts agree on a certain view
Problem: There are many possible reasons why people in a field might agree, and such agreement doesn’t always signal that they are all correct
Possible strategies to adress the novice experts problem (Appraisal by meta experts)
Advantage: Degree, prizes, work experience etc. Reflect publicly available certifications by other experts of ones expertise
Problem: Novices are not always in a position to assess the significance of ones credentials
Possible strategies to adress the novice experts problem (conflicts of interest)
Advantage: sometimes, conflicts of interest are clear
Problem: In many contexts, novices cannot easily detect more subtle conflicts of interest
Possible strategies to address the novice experts problem (past track-record)
Advantage: It seems easy to check how many times and in what situations a putative epert got it right
Problem: For complex phenomena, it may be beyond the novices capacity to check whether a putative expert got It right
Illusion of understanding
People feel they understand complex phenomena with far greater precision, coherence and depth than they really do; they are subject to an illusion of explanatory depth
Non-scientific practices
Do not aim at generating knowledge in the same way science does; their proponents try to create the false impression that they generate genuine trustworth knowledge
Pseudo-scientific practices
Are not scientific, but their proponents try to create the false impression they generate genuine trustworthy knowledge
Science is a practice
Socially and institutionally organized
Aimed at producing knowledge about natural phenomena
Reproducible studies
Can be performed again
Produces the same or sufficiently similar ersults as the original study
Why replicate a study
Limits the role of luck and error
E.g. False positives (type 1 error)
False negatives (type 2 error)
INcreases confidence a hypothesis is true (or false)
e.g. more evidence from different sources/labs
Helps science to self-correct
Why do many results fail to replicate
Fraud
Questionable research practices (hacking - checking statistical significance of results before deciding whether to collect more data
Incentive structure and organization of science institutions
Examples of social-institutional conditions that influence self-correcting
Open datasets
Replace null hypothesis significane testing
Reward replication work
Publish negative results
Diversity science
-
-
Three common features of scientific practice
1) publicly shared (oft mathematical) representations and techniques (hypothesis)
2)openness to criticism (Grounded in hypothesis)
3) empirical evidence
Rationale for experimental control
Any measured change in the dependent variable is due only to the intervention
By dividng participants randomly
You (supposedly) distribute participants with particular characteristics “equally” among the two groups. This would minimize the differences between the two groups with all known and unknown extraneous variables
Does randomization really solve the problem of variable control?
Not really: Random group assignment does not guarantee that researchers selection of experimental groups does not distort experimental results
BUT
Random group assignment does not guarantee that extraneous variables do in fact vary equally across the two groups in any single experiment
It is only over an indefinite series of repetitions of the random division that the variables Z will be equally distributed between the two groups
Repeat random division a lot and a lot of times and the frequency of education in one group will be about the same as the frequency of education in the other group
But
Researchers do not make random division of experimental participants indefinitely often, they do it once
Whats data science
Use of computational, algorithmic, statistical and mathematical techniques to analyse and gain knowledge from the big data
Any tool for data analysis does:
Makes assumptions (e.g. about the statistical structure of the data, about how to weigh different data etc.)
Based on algorithms
“trained” or “labelled” sample data to extract patterns or to make predictions
The bay model
A model oft he san franciso bay: a downsized reconstruction of the bay in san Francisco
1000 times smaller than the actual bay
Mimics the tides and currents of the actual bay
Scientific model: used to learn not just about the model but about the actual bay
John reber wanted to fill parts of the bay by building a dam
The model showed this was not a good idea
By manipulating and studying the model we can learn about the actual bay
Models
Models represent a target system: models are about something else, namely their target
How do models represent their target? By being relevantly similar, not by being identical
The bay model replicated tides and currents but not the number of sailboats or the houses on the coast
Models are incomplete and simpler versions of their target, theyre idealiizations
The bay model has much faster tidal cycles
Models are thus abstractions of their target
Representations
These involve triadic relation between an agent or human representation and a represented world
Models that exemplify
For a model to exemplify a (group of) target systems it must be a group member
Fruit fly (Drosophila melanogaster) s a model organism in genetics and developmental biology
Mice are model organisms in biology and medicine to model human diseases
Fruit flies and mice are relevantly similar to their target system, in this case humans, but there are of course many differences
Why model at all
When its impossible to perform experiments on the target (solar system)
When its impractical to perform experiments on the actual target (The Bay model)
When its immoral to perform experiments on the actual target (Using mice as a model organism to test vaccines for covid)
Models of data
A model of data, or data model, is a regimented representation of some data set, often with the aim of highlighting whether or not the data count as evidence for a given hypothesis
Data are any public records produced by observation, measurement or experiment
Raw data
Video recordings of capupchin monkey behavior, observations of teh position of planets in the night sky, readings of a thermometer, participants answers on a questionnaire in a psychology experiment
Models of data (steps involved)
1) Eliminating errors
2) Displaying measurements in a meaningful way
3) Extrapolating from those measurements to the expected data for measurements that weren’t actually taken
Scale models
Concrete physical objects that serves a down-sized or enlarged representation of their target system
Building a model
1) specification of the target system(s)
2) Construction of the model
3) analysis of the model
The solar system
How is the model build in the documentary similar to its target system and how is it different
Similar: in terms of the size of the planets and the distance between them
Different: the composition of the planets, no atmosphere on earth, no satellites, no comets or debris
Mechanistic models
Mechanistic models are representations of mechanisms
Mechanisms are organized systems consisting of component parts and component operations that are organized spatially and temporally, so as to causally produce a phenomenon
Mechanistic models represent the causal activities of organized component parts that produce some such phenomenon
This illuminates how the target phenomenon works and how it depends on the orchestrated functioning of the mechanism that produces it
Computer models
Computer models or simulations are programs run on a computer using algorithms to explore aspects or changes to a target system
Thomas schellings checkerboard model of segregation (3 assumptions)
Assumption 1:
Two sorts of agents
Agents live in a two-dimensional grid
Agents initially randomly distributed on the grid#
Assumption 2:
Agents have preferences for their neighbourhood
Agent satisfied only if surrounded by at least t% (e.g. 30%) of agents like its self
Assupmtion 3:
Agents interact accordingly to a behavioural rule
When an agent is not satisfied the agent moves to any vacant location on the grid
Idealized models
Deliberately simplified or distorted representations
Omitting, abstracting from certain known features of a target system/phenomenon
Why?
Make the model easy to construct, manipulate, analyse and run on a computer.
Individual choices
Can lead (under specific conditions) to significant unintended consequences for larger groups
Model
Idealized representation of something compicated with the goal of making it more simple/tracable or understandable
Building a model
Figuring out what should be included in the model and how given certain aims is an opportunity to learn about the model
Manipulating a model
Figuring out how the model changes if you intervene on it in some way is an opportunity to learn about the odel
Representation (model)
Meant to stand in for their target systems
Different features of model more or less similar to certain features of target system
Robustness analysis
Build slightly different models of the same target
Manipulate the models in comparable ways
Compare models results
Whats the point of robustness
Assess sensitivity of a model to changes in its basic structure
identify model features responsible for certain results
Evaluate with similarities and idealizations matter to learning about the world
Rational
To be rational is to reason in accordance with principles of reasoning that are grounded in logic
Logic as consequence relation
Let X a set of sentences and c any sentence, then c is a logical consequence of X just in case there is no situation in which everything in X is true but c is untrue
Counterexamples:
Is an exeption to a proposed general rule or hypothesis
Do actual decision makers reason as economic models predict?
Economic models are idealized
There is more than one system of logic
Deviations from economic mmodels may indicate the logic assumed by economists is inadequate to represent human rationality
Are humans irrational?
Human deviations from economic models and from deductive logic need not be symptoms of irrationality
Generalization vs prediction
Generalization: o1 o2 and on have each been observed with property P. Threfore all Os have property P
Prediction: O1,O2,…On each have been observed with property P. Therefore the next observed On+1 will have propety P
Inductive reasoning
A kind of risky reasoning
Conclusions from non-deductive arguments only follow probabilistically from premises
How can we acquire (non lucky, trustworthy) knowledge from a risky form of reasoning
justifying induction deductively:
If the conclusion of an inductive argument followed deductively from its premises, the falsehood of the conclusion would contradict the truth of the premises
But the falsehood of the conclusion of an inductive argument does not contradict the truth of its premises
therefore, the conclusion of an inductive argument cannot follow deductively
WHen is reasoning trustworthy
A form of reasoning is trustworthy or reliable if it yields true conclusions most of the time
Justifying induction non-deductively
Inductive argument 1; inductive argument 2; inductive argument n have all been reliable in the past
Therefore, inductive argument n+1 will be reliable in the future
This argument is itself reliable only if we assume that nature is uniform
why cant it be proved non-deductively
In trying to show non-deductively that its true that nature is uniform, you presuppose the reliability of inductive reasoning
But its exactly the reliability of inductive reasoning we want to establish
Induction cannot be shown to be reliable via a deductive argument
This violates the risky character of induction
Induction cannot be shown to be reliable via a non-deductive argument
This involves circular reasoning
The best way to make generalizations:
Nature is uniform: use induction; other method = success or failure
Nature is not uniform:
Use induction –> failure
Use some other method –> failure
Abduction
inference to the best explanation
How can we acquire (non-lucky) knowledge from (a risky form of reasoning, which is guided by) explanatory considerations?
What candidate hypotheses should be considered in relation to agny given set of observations
If abduction is to be reliable, the, at least typically, the set of candidate hypotheses should contain true hypotheses
What is best explanation
If abduction is to be reliable, then it should be clear and agreed upon what explanatory virtues like simplicity, fruitfulness and unifying power mean
What is probability (subjectivist interpretation)
The probability of an outcome is an individuals subjective, rational degree of belief that the outcome will obtain
Degrees of belief
Level of confidence in the truth of a given hypothesis
Revealed by possible bets you would accept and reject
Andrey kolmogorovs axioms of probability
Axiom 1: All probabilities are numbers between 0 and 1
Axiom 2: if a proposition is certainly true, then it has a probability of 1. If certainly false, then it has prob. 0
Axiom 3: If h and h* are exclusive alternatives (they cannot both be true at the same time), then P(h or h) = P(h) + P(h)
Dutch book arguments: Basic idea
If your degrees of belief do not conform to the rules of probability, there are possible betting situations where you are guaranteed to lose money (you fall prey of a dutch book)
You do not want to lose money
Therefore, your degrees of belief should respect the rules of probability
Problems with subjectivism
If the only constraint on your degrees of belief is they cohere with the axioms of probability, then you may have very odd “rational” degrees of belief
Betting behaviour doesnt seem to generally be a good guide about what one (should) believe
Frequency interpretation
The probability of an outcome is the frequency with which the outcome occurs in a long sequence of trials
What is probability
Three possible answers:
- subjective
- Frequency
- Propensity
What is a long sequence of trials
A long sequence of similar trials:
Similar to an experiment repeated over and over again to produce an infinite series of observations about the value of a variable of interest
Problem of single-case probabilities
Cannot assign probaabilities to one-off events
Propensity interpretation
The probability of an outcome is a propensity inherent in the physical conditions producing the outcome
Propensities can be understood as
Causal dispositions of a situation to produce certain outcomes
Whats a causal disposition of a system
The systems tendency to behave in a certain way under certain circumstances
What is statistics good for
Description
Estimation
Generalization
Hypothesis testing
Null hypothesis significance testing
1) formulate a null hypothesis
E.g. this treatment is not effective
There is no correlation between these variables
This person has NO special ability etc
2) develop expectations in the form of probability distributions for possible outcomes given the truth of hypothesis
3) gather data/observations and evaluate to what degree observed data violate expectations
4) draw an inference from this comparison
Significance level
Decision about how improbable, given the truth of the null, an observed result must be to warrant rejecting null
How surprising/improbable should an outcome be to be considered significant
No uncontroversial answer
Largely convention and background knowledge about phenomenon
p-value:
Probablity of obtaining test results at least as extreme as the results actually observed under the assumption that the null hypothesis is true
P value: Basic idea
An index of how incompatible observed data are with a statistical hypothesis
The smaller the p-value, the more surprising data are given the null
Higher significance level (i.e. a lower probability for statistical significance)
reduces chance of type I, but increases chance of type II error
lower significance level (i.e. a higher probability for statistical significance)
reduces chance of type II, but raises chance of type I
Choice of significance level determines the degree to which one should be willing to accept diffirent kinds of errors
Type I error (false positive)
Erroneously rejecting the null hypothesis
Type II error (false negative)
Erroneously failing to reject null hypothesis
NHST goals
Develop expectations assuming H0 is true
Check the likelihood of data given H0
Decide whether to reject H0
Problems with NHST
1) silent on whats true (we want to know what hypotheses are true… but p values and denying H0 does not say how likely a hypothesis is
2) Silent on priors
We have a f? (remember, NHST only says what we cannot deny) airly good idea of how the world works before testing NHST cannot account for that
3) Null good, what now?
If we fail to deny H0 what should we believe
Bayes theorem
1) formulate competing hypothesis
2) assign a prior probability to each one
3) Gather data
4) Evaluate the degree to which the data (dis)contfirms the hypothesis
5) Update probabilities of the hypotheses (inductively)
Advantage of Bayes Theorem
1) Allows us to account for our previous knowledge of the world
2) Allows us to check how much the data confirms or disconfirms a hypothesis
Bayes Factor
The typical approach is to calculate the posterior probability for all hypotheses
But we can instead measure the ratio to which observation (dis)confirms each hypothesis
Problems for Bayesianism
1) How do we define priors (there are often no objective criteria to define priors in hypotheses
Solution: We do have a lot of background knowledge that constrains our common priors
Variability in priors isnt always bad
Makes it transparent how disagreements arise and how to settle disputes
2) Bayesianism not always the right approach
Should science be free from all social values, if it is to delive objective, trustworthy knowledge
Science should be free from the influence of any social value
Good science involves only logical reasoning and evidence
Social values
Things, relationships or states that are (believed by some community to be) good
Value free ideal
Suggests that science can produce objective knowledge to the extent it is free from values
illegitimate roles of values in science
Endorsing a scientific theory not because of evidence but because we want it to be true
Manipulating results to support a particular hypothesis and get published
Exclusion of members of certain groups from scientific societies and institutions
Legitimate roles of values in science
Choice of research questions
Choice of how much evidence is needed before accepting or rejecting a hypothesis
If science is to deliver objective trustworthy knowledge then
Scientists judgements and methods should be critically and openly assessed in light of diverse bodies of data, competing interpretations and alternative hypotheses
–> an intersubjective process
How can we make reliable causal inferences
Understanding causes matters, causality allows us:
To intervene and predict, so that we can stop bad things happening and make more good things happen
To explain why or how things happen
Whats causation
Regular association
Difference-making and manipulability
Energy transference
Dispositions/tendencies
Temporal succession
C regularly comes before E
Contiguity
C and E happen nearby in space
Problems with regular association view
1) some variables are causally related, but not (spatially or temporally) contiguous
2) Causal relationships are asymmetric (C causes E but E does not cause C)
3) Some variables are conditionally dependent/associated but not causally related
Manipulability
Two variables C and E are causally related when, if the value of C changed, the value of E would change too
Basic idea: If C and E are merely associated or correlated, then intervening on C will not change the value of E
Inferences about causal relationships
You always need assumptions that connect what can be observed to the underlying causal structure that generates the observations
The principle of common cause
An observed dependece between two variables X and Y is indicative of either X causing Y, Y causing X, or the existence of a common cause Z
The principle of common cuase (What does it bridge)
Observed patterns of conditional dependence and independence in a set of variables with causal relationships between the variables
Causal Markov condition (CMC) says:
that each variable in a graph is independent of every other variable (exept its effects) conditional on all of its direct causes
Scientific Revolution
A radical change of a reigning scientific paradigm being overturned in favor of a new paradigm
Change in: out image of reality
How data is collected analyzed interpreted
Which logic/methods are accepted
Thomas kuhn argues on paradigms:
It is impossible to compare different paradigms
Khun suggests paradigm shifts are like gestalt switches where someones perspective changes from one thing to another
If hes right: scientific revolutions prevent science from proceeding in a straight line
The no miracles argument
1) our best scientific theories are massively predictively successful, and facilitate incredible technological innovation
2) the best explanation forr these successes is that our best theories are true
What should we conclude about scientific progress, if we pay more attention to the history of science
A pessimistic induction:
1) There have been many empirically successful theories in the history of science that have subsequently been rejected as false
2) our current best theories are no different in kind from those theories that were rejected
By induction we have reason to bleieve our current best theoreis will be rejected as false
-> current theories not true
What should we conclude about scientific progress, if we pay more attention to the history ofscience (Conclusions)
History and sociology of science indicate its simplistic and inaccurate to say tha tscientific knowledge accumulates linearly over time science proceds more erratically
Controversial to say that because of its practical successes science makes progess delivering us an increasingly accurate image of the world
