Tilburg University
Study Program: Master Data Science and Society
Academic Year 2021/2022, Semester 2, Block 3 (January to March 2022)
Course: Statistics and Methodology (880259-M-6)
Lecturers: L.V.D.E. Vogelsmeier
,Lecture 1: Statistical Inference, Modeling and Prediction
Introduction to statistical inference
Statistical Reasoning
• consideration of uncertainty
• systematize the way we account for uncertainty when making data-based decisions
→ avid bias by ourselves: “get the result I wish to find”
Probability Distributions
• Probability distributions quantify how likely it is to observe each possible value of some
probabilistic entity “re-scaled frequency distributions”
• they show the proportion of observations that are in a certain bin, not the absolute number /
frequency of observations
• probability distributions with higher standard deviation are broader and less high
Statistical Testing
• When we conduct statistical tests, we weight the estimated effect by the precision of the
estimate.
𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒 − 𝑁𝑢𝑙𝑙 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑧𝑒𝑑 𝑉𝑎𝑙𝑢𝑒
• Wald Test (type of T test) 𝑇 =
𝑉𝑎𝑟𝑖𝑎𝑏𝑖𝑙𝑖𝑡𝑦
o if there is no effect hypothesized, we assume “0”
o in general, the larger the test statistic, the better
Sampling Distribution of the test statistic
• probability distribution of a statistic
• The sampling distribution quantifies the possible values of the test statistic over infinite
repeated sampling.
• The area of a region under the curve represents the probability of observing a test statistic
within the corresponding interval.
• To quantify how exceptional our estimated test statistic is, we compare the estimated value
to a sampling distribution of t-statistics assuming no effect (null hypothesis)
o null hypothesis = no effect → “nil-null”
• If our estimated statistic would be very unusual in a population where the null hypothesis is
true, we reject the null and claim a “statistically significant” effect
Interpreting P-Values
• All that we can say is that there is a 0.032 probability (p value) of observing a test statistic at
least as large as 𝑡̂, if the null hypothesis is true.
Introduction to statistical modeling
• For simple questions we can use statistical testing to control for uncertainty. In most real-
world cases, we want to employ a modeling perspective to control for confounding variables.
• When modeling, we can make inferences about the model parameters, or we can predict
outcomes for new cases.
, Lecture 2: Research Cycle, Research Design and Exploratory Data Analysis
Discuss research/data science cycle
• CRISP-DM: The Cross-industry
Standard Process for Data
Mining was developed to
standardize the process of data
mining in industry applications
• The Data Science Cycle combines
the classical Research Cycle and
the CRISP-DM. The grey colored
activities are mandatory.
Discuss research design in data science
• In data science, we rarely design experiments/empirical studies
• Research design is still crucial to data science to design an appropriate analysis.
o You must know how to operationalize the question in a statistically rigorous way.
▪ Make sure you understand exactly what is being asked
▪ Convert each aspect of the question into something quantifiable
▪ If possible, code the research question into a set of hypotheses.
o You must be able to choose/build a statistical model, statistical test, or machine
learning algorithm that can answer your well-operationalized research question.
▪ Once you have a well-operationalized research question, you need to
convert that question into some type of model or test.
o You must understand what types of data/data sources you’ll need.
Introduce EDA (Exploratory Data Analysis)
• interactively analyze/explore your data
• More of a mindset than a specific set of techniques or steps: data driven approach to explore
something, not to test hypothesis
• diverse selection of tools to use
o Statistical graphics: Histograms, Boxplots, Scatterplots, Traceplots
o Summary graphics: measures of tendency & dispersion, order statistics
o Data Screening/Cleaning: missing data, outliers, invalid values
Interfacing EDA & CDA (Confirmatory Data Analysis)
• CDA: there is usually a clear hypothesis to test, we have some prior knowledge which we
want to test, e.g., by using hypothesis testing
• unsupervised learning models are usually more EDA because we want to find pattern
• Either can stand alone, but they play together better
o When the data are well-understood, we can proceed directly to CDA.
o If we don’t care about testing hypotheses, we can focus on EDA.
• EDA can be used to generate hypotheses for CDA.
• EDA can be used to sanity check (Plausibilitätsprüfung) hypotheses