library / lib7f1333829757edbe
Learning from Data: A Short Course
Yaser S. Abu-Mostafa, Malik Magdon-Ismail, Hsuan-Tien Lin · 2012
In a sentence
A comprehensive introductory textbook that teaches statistical reasoning as a way of learning about the world from variable, messy data through descriptive and inferential procedures.
Learning From Data: An Introduction to Statistical Reasoning teaches readers a new way of thinking about and learning from data across psychological, social, educational, political, and economic domains. Rather than treating statistics as mere number-crunching, the book emphasizes the logic underlying each procedure: why variability matters, how sampling distributions enable inference, and how the six-step hypothesis-testing schema applies from the simplest z-test to complex factorial ANOVA. Built around two real datasets (a smoking-cessation study and a maternity/marital-satisfaction study), the book devotes a full chapter to each difficult concept, uses extensive repetition, and integrates parametric and nonparametric procedures so students learn to choose the right tool. It uniquely confronts the gap between random sampling (which statistics textbooks preach) and random assignment (which experiments actually use), giving readers the conceptual tools to question and challenge data-based claims and to conduct sound research themselves.
The four lenses
- Science
- Statistics
- Systems
- Strategy
The model
A framework model representing how study design levers (sampling method, sample size, significance level, alternative hypothesis type, measurement quality) and contextual conditions (effect size, assumption satisfaction) influence psychological/behavioral states of the analyst (variability of the sampling distribution, statistical power) and ultimately outcomes (correct inference, generalizability, causal warrant). Inferred from the book's repeated emphasis on factors affecting power, error rates, and valid conclusions.
Sampling Methoddesign lever
The procedure used to obtain observations from the population, ranging from independent (within-sample) random sampling to biased sampling; determines whether probability theory applies and whether generalization is warranted.
Random Assignmentdesign lever
The procedure of assigning subjects to experimental conditions using a random method, applied independently of subject characteristics in order to eliminate confounds and license causal conclusions about the independent variable.
Sample Sizedesign lever
The number of observations (n) collected in a sample, a key design lever that reduces the variability of the sampling distribution (standard error) and thereby increases statistical power and narrows confidence intervals.
Significance Level (alpha)design lever
The probability of a Type I error chosen by the analyst, defining the rejection region; lowering it reduces Type I error probability but increases Type II error probability (reduces power), embodying a deliberate trade-off.
Alternative Hypothesis Typedesign lever
Whether the analyst uses a directional (one-tailed) or nondirectional (two-tailed) alternative hypothesis; appropriate use of a directional alternative increases power but precludes detecting deviations in the other direction.
Measurement Qualitydesign lever
The validity, reliability, and scale-appropriateness of how variables are measured, including careful data-collection procedures that reduce excess variability introduced by measurement error.
Statistical Assumption Satisfactioncontextual condition
The degree to which the data meet the population, sampling, and data-scale assumptions (e.g., normality, homogeneity of variance, independence) required by a chosen parametric procedure, which determines whether results are accurate and whether nonparametric alternatives are needed.
Effect Sizecontextual condition
The degree to which the phenomenon is present in the population, i.e., how false the null hypothesis is (e.g., the standardized difference between population means); larger effect sizes increase power but cannot themselves be manipulated.
Variability of the Sampling Distributionpsychological state
The standard error of the sampling distribution of the test statistic, a state of the analysis that reflects how dispersed sample statistics are around the parameter; smaller variability sharpens inference. Reduced by larger sample sizes, careful measurement, and dependent sampling.
Statistical Powerpsychological state
The probability of correctly rejecting a false null hypothesis (1 minus beta); a central state of an analysis influenced by alpha, effect size, sampling-distribution variability, sample size, and alternative-hypothesis type.
Correct Statistical Inferenceoutcome metric
The outcome of making the right decision about the null hypothesis (correctly rejecting a false null or correctly retaining a true null) while avoiding Type I and Type II errors.
Generalizability of Conclusionsoutcome metric
The extent to which conclusions can legitimately be extended to a broader population, determined by whether random sampling was used and the breadth of the population sampled; overgeneralization occurs when inferences exceed the sampled population.
Causal Warrantoutcome metric
The degree to which the analysis supports a causal conclusion that the independent variable produced the observed effect, licensed by random assignment and absence of confounds rather than by the form of the statistical test.
How they connect
- sample size − influences sampling distribution variability
- measurement quality − influences sampling distribution variability
- sampling distribution variability − influences statistical power
- significance level → influences statistical power
- effect size → influences statistical power
- alternative hypothesis type → moderates statistical power
- statistical power → predicts correct inference
- significance level → influences correct inference
- assumption satisfaction → moderates correct inference
- sampling method → predicts generalizability
- sampling method → influences correct inference
- random assignment → predicts causal warrant
- sample size → mediates statistical power
A candidate measure
Learning from Data: A Short Course — derived measurement candidates
Sampling Method
randomness classification code; independence-of-samples flag; documented sampling-frame coverage
self-report suitability: low
Random Assignment
present/absent flag; baseline equivalence checks
self-report suitability: none
Sample Size
n per group; total N; number of pairs or blocks
self-report suitability: none
Significance Level (alpha)
alpha value; rejection-region boundary
self-report suitability: none
Alternative Hypothesis Type
directional vs. nondirectional code; one- vs. two-tailed flag
self-report suitability: none
Measurement Quality
reliability coefficient; validity coefficient; scale type; standardization checklist
self-report suitability: low
Statistical Assumption Satisfaction
normality diagnostic; homogeneity-of-variance test result; scale-type code
self-report suitability: none
Effect Size
Cohen's d (d-hat); r-squared; delta inputs
self-report suitability: none
Variability of the Sampling Distribution
standard error of the mean; estimated standard error; pooled standard error
self-report suitability: none
Statistical Power
1 minus beta from Table B; delta
self-report suitability: none
Correct Statistical Inference
alpha (Type I error probability); beta (Type II error probability)
self-report suitability: none
Generalizability of Conclusions
overgeneralization flag; match between claim scope and sampling frame
self-report suitability: none
Causal Warrant
random assignment flag; confound checklist
self-report suitability: none
The story
The reader A student or researcher in the behavioral sciences who wants to understand, conduct, and critically evaluate data-based claims about the world.
External problem
Data in the behavioral sciences are messy and variable, making it impossible to see clear facts without proper analysis, and the reader must choose and apply the right statistical procedure.
Internal problem
The reader feels intimidated by statistics, fears the math, and worries they cannot tell good data from misleading claims.
Philosophical problem
It is wrong to accept or reject claims about psychology, society, and policy on faith or intuition when sound statistical reasoning could reveal the truth.
The plan
- Learn to describe data with frequency distributions, central tendency, variability, and z scores.
- Understand probability and sampling distributions as the foundation of inference.
- Master the six-step hypothesis-testing schema and apply it across procedures.
- Use the Statistical Selection Guide to pick the right test for any situation.
- Distinguish random sampling from random assignment to know what you can conclude.
- Report results clearly, including effect sizes, and interpret them in context.
Success
- The reader can choose and correctly apply the appropriate statistical procedure for any situation.
- The reader can interpret data and their limitations, drawing conclusions only about the populations sampled.
- The reader can critically question and challenge data-based claims in everyday life.
- The reader retains the material through the schema, repetition, and detailed explanations.
At stake
- Being fooled by biased samples, misinterpreted p-values, and unwarranted causal claims.
- Making costly Type I or Type II errors due to misunderstanding power and significance.
- Misunderstanding the body of knowledge in the behavioral sciences and its limitations.
Related in the library