You are on page 1of 14

Machine Translated by Google

The University of Electro-Communications Graduate School Second semester 2022

International Politics and Economy

5th
Probabilistic reasoning and its foundation
Machine Translated by Google

last class

ÿ Fundamental issue of
causality Treatment effect is not strictly
measurable Observable only in treated/untreated individuals

ÿ Halo effect

ÿ Preparing a pseudo-same environment/same


individual Considerably possible in natural
sciences Difficult in social sciences (because it limits human behavior)
Machine Translated by Google

Previous test answer example

(1) “I asked 10 Harvard college students, lifestyle habits to become smarter! ]


Points to note when reading

•Pay attention to what you specifically mean by “getting smarter”


•Use it as a reference (half of the story) as something that has a halo effect

(2) How to elucidate "lifestyle habits that make you smarter"

•Comparison of identical twins (with restrictions on freedom)


•Comparison of human clones (with human rights issues)
•Comparison of two groups by randomization (*To be covered in the next class)

Limitations: •There is no proof that you will become really smart with
that lifestyle (*This is the theme of today's and later's episodes)
Machine Translated by Google

from your question

Q. About the example of the halo effect. Isn't collecting similar


things and comparing them a basic procedure for searching for
causes? Shouldn't we compare similar things?

A. The problem with the halo effect is that it collects cases with
the same results (successful companies) and attributes them to
commonalities (customer first, etc.). If you want to know the
effect, you have to control and confirm that the results are
different with or without treatment. (Defining company success,
removing the factors that influence success, and confirming
that companies that did or didn't take customer centricity
succeeded/didn't)
Machine Translated by Google

probabilistic causal inference

ÿCollect a large number of cases (large-N)

ÿ Examine whether there is a


statistically significant relationship between cause and
effect after controlling

ÿ It is enough to say that the relationship is


sufficiently certain.
Machine Translated by Google

Inference procedure

ÿ Hypothesis

presentation Decide what questions you want to investigate and what you want to

clarify ex. In order for elections to function properly, voters need to be

knowledgeable about politics. Question: What kind of people have correct

knowledge? Hypothesis: People who read newspapers have correct

knowledge

ÿ Descriptive reasoning

Prepare shareable information necessary for hypothesis verification

Collect a wide range of explanatory variables

Ex. Conducting a questionnaire

Testing knowledge of politics Newspaper

reading time Attributes (age, gender,

educational background, etc.)


Machine Translated by Google

Inference procedure

ÿ Causal
inference Identify relational expressions based on collected data

Linear is the basis

+ +…

y: dependent variable (political knowledge test score)


x: explanatory variable ( : newspaper subscription
(yes=1), : age, : gender (female=1))

a:
coefficient u: error
Machine Translated by Google

Inference procedure

ÿThe dependent variable is continuous

ex. Economic level of a country

(GDP per capita)

ÿ Dependent variable is discrete value

ex. Medication effect (yes=1, no=0)


Machine Translated by Google

Ordinary least squares (OLS)

ÿ One of the ways to identify the function


that expresses the trend of the sample

ÿ When considering two


variables, calculate a and b
that minimize the sum of (observed
value - predicted value)^2 for y = ax + b

ÿ If we increase the number of

explanatory variables, we should be able to

measure the effect more accurately...


Machine Translated by Google

Results of regression analysis (example)

model I II III IV

newspaper subscription 0.458 (0.052) 0.442 (0.051) 0.274 (0.053) 0.218 (0.053)

college graduate 0.410 (0.051) 0.342 (0.052) 0.284 (0.052)

watching TV news 0.220 (0.093) 0.165 (0.092)

moderate income 0.112 (0.054) 0.099 (0.053)

high income 0.220 (0.063) 0.183 (0.062)

41-64 years old 0.295 (0.058) 0.256 (0.058)

65 years of age or older 0.386 (0.067) 0.295 (0.067)

Woman -0.351 (0.045) -0.318 (0.044)

Has a spouse 0.008 (0.052) -0.011 (0.051)

political interest 0.215 (0.030)

slice 0.981 (0.044) 0.886 (0.045) 0.668 (0.100) 0.230 (0.117)

Respondents 1542 1542 1542 1541

adjusted coefficient of determination 0.048 0.085 0.148 0.175

*Standard error in parentheses


Machine Translated by Google

model I II III IV

How to read the table newspaper subscription 0.458 (0.052) 0.442 (0.051) 0.274 (0.053) 0.218 (0.053)

college graduate 0.410 (0.051) 0.342 (0.052) 0.284 (0.052)

watching TV news 0.220 (0.093) 0.165 (0.092)

moderate income 0.112 (0.054) 0.099 (0.053)

ÿ Models with numbers high income 0.220 (0.063) 0.183 (0.062)

41-64 years old 0.295 (0.058) 0.256 (0.058)

Filled items are on the right side 65 years of age or older 0.386 (0.067) 0.295 (0.067)

Woman -0.351 (0.045) -0.318 (0.044)

Has a spouse 0.008 (0.052) -0.011 (0.051)

political interest 0.215 (0.030)

Model I slice 0.981 (0.044) 0.886 (0.045) 0.668 (0.100) 0.230 (0.117)

Respondents 1542 1542 1542 1541

score = * newspaper subscription + u adjusted coefficient of determination 0.048 0.085 0.148 0.175

Model II

score = *newspaper subscription + *university graduate + u

ÿ Numbers are coefficients (single effect)

Model I score = 0.458*newspaper subscription + u


Model II score = 0.442*newspaper subscription + 0.410*college degree + u
Machine Translated by Google

Limits of control

ÿ Number of
explanatory variables Variables not
handled by any model ex. Leisure time

ÿ No matter how much you increase it, you can't control it


Overlap between explanatory variables occurs by the amount increased

(Example) Newspaper subscription and political interest level

ÿ Unobservable variables cannot be set ex.


Personality
Machine Translated by Google

quiz

The regression results on p.


10 have standard errors next to the coefficients.

The standard error indicates the degree of variability of


the coefficients. The coefficient divided by the standard error is called the t-score.

If the sample size is large


enough, if |t| > 1.96 the probability of the coefficient being 0 is
less than 5%, and if |t| > 2.57 the probability is less than 1%
and the coefficient is not 0. It can be inferred that there is a high
possibility (that is, it is effective).

For Models I to III, what are


the independent variables for which you can infer no
effect at the 5% level (significance level)?
Machine Translated by Google

References

Tetsuya Matsubayashi 2021 "Political Science and Causal Inference: Politics and Society as Seen through Comparison" Iwanami Shoten

You might also like