Question bank

Q1 · Disciplines of data science (Venn diagram)

The lecture's Venn diagram places data science at the overlap of computer science/IT, math and statistics, and domain/business knowledge. A sports physiologist has strong statistics skills and deep domain knowledge but hardly any programming or IT skills. In which overlap of the diagram does this profile fall?

1. AA. Machine learning
2. BB. Software development
3. CC. Traditional research
4. DD. Data science

Q2 · Data, information, knowledge, wisdom

The lecture's data-to-wisdom pyramid (from knowledge discovery in databases, KDD) describes one football moment at four levels. At which level does the statement 'I am pretty close to the goal' sit, and why?

1. AA. Information, because the raw coordinates have been organized into a meaningful description
2. BB. Knowledge, because information has been integrated and understood in context
3. CC. Wisdom, because it is an informed decision about what to do
4. DD. Data, because it is a raw observation of the player's position

Q3 · Typical data science questions

A sports physician has three requests: (1) 'How many weeks will this athlete need to return to play?' (2) 'Is today's resting heart rate of this athlete weird compared with her usual values?' (3) 'Which of three training programmes should this athlete follow next month?' According to the lecture's list of typical data science questions, which approaches match requests 1, 2 and 3?

1. AA. Classification – clustering – regression
2. BB. Regression – classification – clustering
3. CC. Clustering – anomaly detection – classification
4. DD. Regression – anomaly detection – recommendation

Q4 · Structured vs unstructured data

Which statement about structured and unstructured data matches the lecture?

1. AA. Unstructured data (e.g. images, audio, video, e-mails) cannot be displayed in rows and columns, make up an estimated 80% of enterprise data and require more storage
2. BB. Structured data make up an estimated 80% of enterprise data and are more difficult to manage with legacy solutions
3. CC. Less than 1% of structured data is used in decision-making, whereas about half of all unstructured data is analysed
4. DD. Unstructured data have a well-defined data model, which is why their volume has grown faster than that of structured data

Q5 · Data science vs statistics

The lecture contrasts statistics with data science. Which set of characteristics does the lecture attribute to data science?

1. AA. An established field that starts with hypothesis testing and shines with limited data, such as few test subjects
2. BB. A field that starts with hypothesis testing and is usually practised by people with a background in mathematics or statistics
3. CC. An emerging field that starts with data (data mining: hypothesis generation) and shines with lots of data that are long and wide
4. DD. A field that needs no carefulness or rigour, because exploration removes the danger of drawing unfounded conclusions

Q6 · Blei & Smyth (2017)

Blei & Smyth (2017) argue that data science is more than the combination of statistics and computer science. Which of the following is NOT one of the additional requirements named in their quote on the slide?

1. AA. Understanding the context of data
2. BB. Appreciating the responsibilities involved in using private and public data
3. CC. Clearly communicating what a dataset can and cannot tell us about the world
4. DD. Collecting as many variables as possible, so that no prior question is needed

Q7 · Data science lifecycle: order of steps

Which sequence follows the order of the data science lifecycle shown in the lecture?

1. AA. Data acquisition → data exploration → data preparation → feature engineering → data modelling
2. BB. Data acquisition → data preparation → data exploration → feature engineering → data modelling
3. CC. Data preparation → data acquisition → data exploration → data modelling → feature engineering
4. DD. Data acquisition → feature engineering → data preparation → data exploration → data modelling

Q8 · R as a programming language

According to the lecture, which of the following is a disadvantage of R?

1. AA. Computation time with large datasets, and many packages that offer the same functionality
2. BB. It is a licence-based commercial tool with a small community of users
3. CC. It offers limited options for visualizations and statistics
4. DD. It is hardly used outside academia

Q9 · Data science applications in sport

The lecture (Chmait & Westerbeek, 2021) groups examples of data science in sports research into four areas. Which pairing of example and area is correct as listed on the slide?

1. AA. Ticket pricing – game analytics
2. BB. Player injury modelling – training & coaching
3. CC. Event classification – talent identification
4. DD. Player recruitment – fan & business focused

Q10 · RStudio environment

You have imported a data frame in RStudio and want to see which objects you have created so far in this session. Which of the four RStudio panes shown in the lecture is meant for this?

1. AA. Code editor
2. BB. R console
3. CC. Workspace & history
4. DD. Plots & files

Q1 · Vectors and coercion

What does the following code print?

x <- c(12, "15", TRUE)
typeof(x)

1. AA. "double"
2. BB. "character"
3. CC. "logical"
4. DD. "list"

Q2 · Variables, assignment and packages

Which statement about working in R is correct according to the lecture?

1. AA. R is not case-sensitive, so Weight and weight refer to the same object
2. BB. 2nd_test <- 12 creates a valid variable, because the name is assigned with <-
3. CC. ls() shows the data type (e.g. double, character) of every variable in the workspace
4. DD. install.packages("lubridate") downloads and installs the package, whereas library(lubridate) loads it so that its functions (e.g. ymd()) can be used

Q3 · Matrices

What does the last line return?

m <- matrix(1:6, nrow = 2)
m[1, 3]

1. AA. 3
2. BB. 6
3. CC. 5
4. DD. An error, because the matrix has only two columns

Q4 · Factors

What does the following code print?

event <- factor(c("sprint", "road", "track", "road", "sprint"))
as.integer(event)

1. AA. 2 1 3 1 2
2. BB. 1 2 3 2 1
3. CC. 1 2 3 4 5
4. DD. NA NA NA NA NA

Q5 · Lists

Given the list below, which expression returns the single number 61?

athlete <- list(name = "Sanne", hr = c(58, 61, 64), injured = FALSE)

1. AA. athlete["hr"][2]
2. BB. athlete[2, 2]
3. CC. athlete[["hr"]][2]
4. DD. athlete[[2]][3]

Q6 · Data frames and logical operators

What does the last line return?

d <- data.frame(name = c("Ann", "Bo", "Cas", "Dirk"),
                vo2  = c(48, 55, 61, 52),
                mass = c(60, 72, 80, 68))
d[d$vo2 > 50 & d$mass < 75, "name"]

1. AA. "Bo" "Cas" "Dirk"
2. BB. "Bo" "Dirk"
3. CC. "Ann" "Bo" "Cas" "Dirk"
4. DD. "Cas"

Q7 · Conditional statements

A coach labels a session by its rating of perceived exertion (rpe, an effort score from 0 to 10). What is the value of zone after running this code?

rpe <- 8
if (rpe >= 9) {
  zone <- "maximal"
} else if (rpe >= 5) {
  zone <- "hard"
} else if (rpe >= 7) {
  zone <- "very hard"
} else {
  zone <- "easy"
}
zone

1. AA. "maximal"
2. BB. "very hard"
3. CC. "easy"
4. DD. "hard"

Q8 · For-loops

What does the last line print?

hr <- c(150, 172, 185, 168, 190, 176)
count <- 0
for (h in hr) {
  if (h >= 170 & h < 185) {
    count <- count + 1
  }
}
count

1. AA. 2
2. BB. 3
3. CC. 4
4. DD. 6

Q9 · Writing functions

What do the two function calls return?

abs_vo2 <- function(rel, mass = 70) {
  rel * mass / 1000
}
abs_vo2(c(50, NA, 60))
abs_vo2(mass = 80, rel = 50)

1. AA. An error for the first call, because mass is not supplied; 4 for the second call
2. BB. 3.5 4.2 for the first call (the NA is dropped); 4 for the second call
3. CC. 3.5 NA 4.2 for the first call; 4 for the second call
4. DD. 3.5 NA 4.2 for the first call; 3.5 for the second call (the default mass of 70 is used)

Q10 · Missing values and na.rm

What does the last line return?

vo2 <- c(40, 50, NA, 60)
mean(vo2, na.rm = TRUE)

1. AA. NA
2. BB. 37.5
3. CC. 150
4. DD. 50

Q1 · Structured vs unstructured data in sport and health

A rehabilitation centre films patients walking, runs markerless motion-capture software that writes the hip, knee and ankle positions of every video frame into a table, and lets therapists add a free-text note after each session. How should these three data sources be classified, following the lecture?

1. AA. Video: unstructured; table of joint positions: structured; free-text notes: unstructured
2. BB. Video: structured, because every frame has a fixed size; table: structured; notes: structured
3. CC. All three are unstructured, because the table is derived from the video
4. DD. Video: unstructured; table: unstructured; notes: structured, because there is one note per session

Q2 · Interpreting the Fitbit graph

The Fitbit slide plots average resting heart rate (y-axis) against BMI (x-axis) for men and women. Which conclusion is best supported by the graph?

1. AA. Resting heart rate rises in a straight line with BMI over the whole BMI range, for both sexes
2. BB. Women's resting heart rate is higher than men's at every BMI, and both are lowest at a normal BMI (about 20–24), not at the lowest BMI
3. CC. Men with a BMI of 45 have a lower resting heart rate than women with a BMI of 20
4. DD. The graph shows that lowering your BMI makes your resting heart rate drop

Q3 · Value of data: activity-based insurance discount

The lecture discussed an insurer (ASR) that gives discounts and gifts when members share activity-tracker data and reach personal activity goals. What did the lecturer report about the effect of such a scheme?

1. AA. Previously inactive members became much more active, so the scheme solved the inactivity problem
2. BB. Members stopped sharing their data, so no effect could be determined
3. CC. Mainly members who were already active benefited, while previously inactive members hardly changed
4. DD. It proved that tracking physical activity causes lower healthcare costs

Q4 · Restricted range in regression

In the van der Zwaard et al. (2018) cycling study, pennation angle (the angle at which the thigh-muscle fibres pull, measured with ultrasound) showed no relationship with sprint power in the elite cyclists. According to the lecture, how should you interpret this?

1. AA. Pennation angle is irrelevant for producing force and can be dropped from future studies
2. BB. The relationship was hidden because Deming regression ignores measurement error in the predictor
3. CC. Because a real determinant always shows a significant relationship, the pennation-angle measurement must have been faulty
4. DD. Elite cyclists probably all have similar pennation angles; if a variable barely varies in the sample, a relationship cannot show up, even if it exists in a wider group

Q5 · Determinants of combined sprint and endurance performance

In the van der Zwaard et al. (2018) results shown in the lecture, which combination of muscle characteristics together explained 67% of the differences between riders in performance VO₂ (R² = 0.67), a key driver of combined sprint and endurance performance?

1. AA. A higher percentage of fast fibres and a larger muscle volume
2. BB. A larger fibre cross-sectional area (FCSA), a larger PCSA and a lower oxidative capacity
3. CC. A higher oxidative capacity, more capillaries × myoglobin, and a smaller PCSA
4. DD. A larger pennation angle, more muscle fibres and a higher blood lactate after the test

Q6 · Excel and CSV limits

A sports scientist stores heart rate once per second for 25 athletes during a 2-hour training, with one row per athlete per second: 25 × 7,200 = 180,000 rows and 3 columns. Which statement about saving this in a single sheet or file is correct, using the lecture's limits?

1. AA. .xls cannot hold it (row limit 65,536), but .xlsx (row limit 1,048,576) and .csv (no built-in limit) can
2. BB. Neither .xls nor .xlsx can hold it; only .csv can
3. CC. All three formats can hold it, because only the column limit matters
4. DD. .xlsx cannot hold it, because its column limit is 256

Q7 · Application Programming Interface (API)

Which statement about an API matches the lecture?

1. AA. An API is the database on the server in which the data are stored
2. BB. An API is the code that governs the access point to a server's database; programs send it HTTP requests without a frontend and often get JSON or XML back
3. CC. An API is a program that reads the HTML of web pages made for people and copies the information into a table
4. DD. Data obtained through an API are always of high quality, so no quality checks are needed

Q8 · Speed-skating API example

Seasonal bests of speed skaters were retrieved through the SpeedskatingResults.com API and shown in a radar (spider-web) chart per distance, as the seasonal best relative to the world record on the race day. In this chart, what do the direction and the size of the filled shape show?

1. AA. Direction: track type (indoor vs open); size: number of races skated
2. BB. Direction: the athlete's age; size: number of world records
3. CC. Direction: performance level; size: specialisation
4. DD. Direction: specialisation (sprint, middle distance, long distance, all-round); size: performance level (sub-elite vs world class)

Q9 · Extract–transform–load (ETL)

A club combines training files from three GPS-watch brands into its data warehouse. Removing duplicate sessions, converting all speeds to m/s and flagging heart rates above a plausible limit belong to which ETL step, and where does that usually happen?

1. AA. Extract, inside the brands' own systems
2. BB. Transform, with the data usually held in a staging area for testing before loading
3. CC. Load, by the business-intelligence tools that read the warehouse
4. DD. Load, in a data mart for each department

Q10 · Data warehouse vs data lake

Which characteristic belongs to a data lake rather than a data warehouse, according to the lecture's comparison?

1. AA. Schema-on-write; stores processed, structured data that are mainly used by business professionals
2. BB. Mature security and a fixed set-up that is less agile
3. CC. Schema-on-read; stores structured, semi-structured and unstructured data raw, in low-cost storage, mainly used by data scientists
4. DD. A focused slice for one department, containing summarised data for that unit

Q1 · Measurement scales

Which statement is meaningful, given the measurement scale of the variable?

1. AA. A 20 °C training day is twice as warm as a 10 °C training day.
2. BB. If gold, silver and bronze are coded 3, 2 and 1, a gold medal is worth three times a bronze medal.
3. CC. An athlete who trained for 30 minutes trained twice as long as an athlete who trained for 15 minutes.
4. DD. The mean shirt number of a team describes the team's typical playing position.

Q2 · Data quality dimensions

Which pairing of a data problem with a DOMA data-quality dimension is correct?

1. AA. The same training session was imported twice — completeness
2. BB. A height column contains the text "1m80" where a number in cm is required — validity
3. CC. A 1.83 m footballer is recorded at 28 kg — timeliness
4. DD. One record lists sex = male and pregnant = yes — uniqueness

Q3 · FAIR data

You deposit your thesis dataset in a trusted repository. It has a persistent identifier (DOI) and rich, machine-readable metadata, and it can be downloaded after logging in. You did not add a usage licence or any information on how, when and by whom the data were collected. Which FAIR principle is least well met?

1. AA. Findable
2. BB. Accessible
3. CC. Interoperable
4. DD. Reusable

Q4 · dplyr: filter with | and NA

How many rows does this pipeline return?

olympics_small <- data.frame(
  Name  = c("A", "B", "C", "D", "E"),
  Year  = c(2008, 2008, 2012, 2016, 2012),
  Medal = c("Gold", NA, "Gold", "Silver", NA)
)

olympics_small %>%
  filter(Year == 2008 | Medal == "Gold") %>%
  nrow()

1. AA. 3 rows (athletes A, B and C)
2. BB. 1 row (athlete A only)
3. CC. 2 rows (athletes A and C)
4. DD. 4 rows (all athletes except D)

Q5 · Duplicates and distinct()

In the Olympics data each row is one athlete in one event at one Games (271,116 rows; 135,571 different athlete IDs). The lecturer ran the code below. Why does distinct() remove 84,039 rows?

df1 <- data_olympics %>%
  select(ID, Name, Age) %>%
  rename(Athlete_code = ID)

df1 %>% nrow()                 # 271116
df1 %>% distinct() %>% nrow()  # 187077

1. AA. Those 84,039 rows are data-entry errors that should also be deleted from the full dataset.
2. BB. After select(), an athlete who entered several events at the same Games (same ID, name and age) has identical rows, and distinct() keeps only one of them.
3. CC. distinct() removes rows that have a missing value in ID, Name or Age.
4. DD. distinct() keeps exactly one row per athlete, so 187,077 is the number of different Olympians.

Q6 · Joining data frames

You want a table with all four athletes and their countermovement-jump (CMJ) result, with NA where no test exists. Which call gives this?

athletes <- data.frame(ID    = c(1, 2, 3, 4),
                       Sport = c("Rowing", "Judo", "Rowing", "Cycling"))
tests    <- data.frame(Athlete_code = c(2, 4, 5),
                       CMJ_cm       = c(41, 38, 45))

1. AA. inner_join(athletes, tests, by = c("ID" = "Athlete_code"))
2. BB. right_join(athletes, tests, by = c("ID" = "Athlete_code"))
3. CC. left_join(athletes, tests, by = "ID")
4. DD. left_join(athletes, tests, by = c("ID" = "Athlete_code"))

Q7 · EDA: suspicious values

summary() of the Olympics data shows a minimum weight of 25 kg, a maximum height of 226 cm and a maximum age of 97 years. What is the most appropriate next step?

1. AA. Delete these records, because values this extreme must be measurement errors.
2. BB. Replace each value with the column mean so it no longer influences the analysis.
3. CC. Filter the records with these values and check the context (sport, event, year, age) before deciding whether they are errors.
4. DD. Ignore them, because summary statistics such as the median are not affected by extreme values.

Q8 · Boxplot reading

summary() of Olympic athletes' age gives 1st Qu. = 21, Median = 24, 3rd Qu. = 28 and Max. = 97. Under the boxplot rule shown in the lecture, which statement is correct?

1. AA. Ages above 38.5 years are drawn as separate points: potential outliers that need inspection, not automatic deletion.
2. BB. The upper whisker extends to 97, the maximum value in the data.
3. CC. Ages above 35 years (Q3 + IQR) are drawn as outliers.
4. DD. The box spans from the minimum to the median, i.e. the youngest 50% of athletes.

Q9 · Correlation and explained variance

In the lecture's corrplot of Olympic athletes, height and weight correlate at r = 0.8, and age and height at r = 0.14. Which statement is correct?

1. AA. Height explains 80% of the variation in weight.
2. BB. Because r is high, the correlation proves that being taller causes athletes to be heavier.
3. CC. Age and height cannot be related in any way, because r = 0.14 is close to zero.
4. DD. Height accounts for about 64% of the variation in weight (0.8 × 0.8 = 0.64).

Q10 · ggplot2: facets and layers

Using the mpg dataset, which code produces a separate scatterplot panel of displ against hwy for each vehicle class?

1. AA. ggplot(mpg, aes(x = displ, y = hwy, color = class)) + geom_point()
2. BB. ggplot(data = mpg) + geom_point(mapping = aes(x = displ, y = hwy)) + facet_wrap(~ class, nrow = 2)
3. CC. ggplot(mpg, aes(x = displ, y = hwy)) %>% geom_point() %>% facet_wrap(~ class)
4. DD. ggplot(mpg, aes(x = displ, y = hwy, group = class)) + geom_point() + theme_minimal()

Q1 · Supervised task typecatch-up notes

You predict next week's fatigue score (a number) from training duration and workload, using past weeks in which the real fatigue score is known. Which task is this?

1. AA. Classification
2. BB. Regression
3. CC. Clustering
4. DD. Reinforcement learning

Q2 · Feature engineering: choosing an aggregatecatch-up notes

One faulty heart-rate reading can make a seven-day minimum unusually low. Why might a percentile (e.g. the 25th) of the same seven days be a more useful aggregate feature?

1. AA. A percentile depends less on a single extreme value than the minimum does
2. BB. Percentiles are always the best aggregate function
3. CC. A percentile removes all measurement error
4. DD. Using a percentile automatically validates the model

Q3 · Naive baseline comparisoncatch-up notes

On new athletes, a model predicting recovery time performs no better than always predicting the training-set mean. What is the strongest conclusion?

1. AA. Recovery time cannot be predicted by any method
2. BB. The model is necessarily overfitting
3. CC. The model has not shown any added predictive value over that simple benchmark
4. DD. A low p-value would fix the prediction errors

Q4 · AI, machine learning, deep learning and data science

Which statement matches how the lecture positions AI, machine learning (ML), deep learning and data science?

1. AA. Data science is part of machine learning, which is part of AI
2. BB. Machine learning is part of deep learning, which is the broader field
3. CC. AI is part of machine learning, which is part of data science
4. DD. Deep learning is part of ML, which is part of AI; data science overlaps with AI and ML but is neither inside nor around them

Q5 · Dummy variables

A linear regression with an intercept predicts power output from cycling discipline (sprint, pursuit or road). Why use two dummy columns (road coded 0,
0) instead of three one-hot columns?

1. AA. With an intercept, the three one-hot columns always add up to 1 and repeat the intercept's information (the dummy variable trap); the dropped category becomes the reference
2. BB. Two columns let the model treat discipline as an ordered number, from road up to sprint
3. CC. One-hot encoding may only be used for variables with exactly two categories
4. DD. Dropping one column removes the road cyclists from the analysis, which reduces noise

Q6 · MAE and MSE calculation

A model predicts 400-m times (s) for five athletes. What are the MAE and MSE?

actual    <- c(50, 52, 48, 55, 60)
predicted <- c(48, 54, 47, 56, 54)

1. AA. MAE = 1.2 s; MSE = 9.2 s²
2. BB. MAE = 2.4 s; MSE = 5.76 s²
3. CC. MAE = 2.4 s; MSE = 46 s²
4. DD. MAE = 2.4 s; MSE = 9.2 s²

Q7 · MAE and MSE: four runners

A model predicts the 10-km race times of four runners (in minutes). What are the MAE and the MSE?

actual    <- c(40, 45, 50, 55)
predicted <- c(39, 48, 49, 58)

1. AA. MAE = 2 min; MSE = 4 min²
2. BB. MAE = −1 min; MSE = 5 min²
3. CC. MAE = 2 min; MSE = 5 min²
4. DD. MAE = 2 min; MSE = 20 min²

Q8 · Overfitting with a small sample

On the least-squares slides, a new line Size = 0.4 + 1.3 × Weight passes exactly through the two red training points (sum of squared residuals =
0) but has a large sum of squared residuals for the green testing points. What does this show?

1. AA. The line underfits: it is too simple to follow the training points
2. BB. The line is overfitted to the small training sample: perfect training fit, poor predictions for new points
3. CC. Least squares always gives the best test performance, because it minimises the sum of squared residuals
4. DD. The testing points should be added to the training data until the testing error is also zero

Q9 · Ridge regression calculation

Two training points and λ = 1. Line 1 (least squares) has slope 1.5 and residuals 0 and 0. Line 2 has slope 0.5 and residuals 0.4 and 0.2. The ridge total is the sum of squared residuals + λ × slope². Which statement is correct?

1. AA. Line 1: 0; Line 2: 0.20. Line 1 is preferred because its residuals are smallest
2. BB. Line 1: 1.5; Line 2: 0.70. Line 2 is preferred
3. CC. Line 1: 2.25; Line 2: 0.45. Line 2 is preferred: it accepts some training error for a much flatter slope
4. DD. Line 1: 2.25; Line 2: 0.45. Line 1 is still preferred, because the penalty never changes which line wins

Q10 · Ridge: the role of λ

Which statement about λ in ridge regression is correct?

1. AA. λ = 0 gives the least-squares line, a larger λ makes the slope smaller, and λ is chosen with cross-validation
2. BB. λ should be the value that gives the lowest sum of squared residuals on the training data
3. CC. A larger λ makes predictions more sensitive to small changes in the predictor
4. DD. With a large enough λ, ridge sets slopes exactly to zero, which LASSO cannot do

Q11 · Ridge versus LASSO

You have 40 rowers and 25 candidate predictors of their 2000-m rowing-machine time, and you expect only a handful of predictors to matter. Which choice best follows the lecture's summary?

1. AA. Ridge regression, because its penalty sets the slopes of unnecessary predictors exactly to zero
2. BB. Least squares, because with few athletes any penalty increases the test error
3. CC. Ridge regression with λ = 0, the setting that is most robust to overfitting
4. DD. LASSO regression, because its λ × |slope| penalty can shrink the slopes of unnecessary predictors to zero, removing them from the model

Q12 · Holdout and tuning

You fit LASSO models with 20 values of λ, compute each model's error on the test set, pick the λ with the lowest test error, and report that error as your final performance. What is the problem?

1. AA. Nothing: the test set is meant for choosing λ
2. BB. The test set has been used for tuning, so the reported error is too optimistic; λ should be tuned with cross-validation on the training data and the test set used only once at the end
3. CC. λ should be chosen on the training error instead, which is unbiased
4. DD. LASSO cannot be cross-validated, so a single test set is the only option

Q15 · Data leakage: splitting repeated measuresfrom the recording

You have 30 training sessions from each of 12 cyclists (360 rows) and want to predict session RPE (how hard the athlete rated the session) from workload variables. How should you create the test set, following the lecture?

1. AA. Randomly select 20% of the 360 rows, because a random row split is unbiased
2. BB. Use the 20% of sessions with the most extreme RPE values, so the model is tested on difficult cases
3. CC. Put all sessions of a subset of cyclists (about 20%) in the test set, so that no cyclist appears in both the training and the test data
4. DD. Fit the model on all 360 rows first, then pick test rows so that every cyclist is represented

Q16 · Unsupervised learning generates hypothesesfrom the recording

A researcher gives a clustering algorithm only the 100-m times of a group of sprinters and long jumpers and asks for two clusters. Both clusters contain a mix of sprinters and long jumpers. According to the lecture, what is the best interpretation?

1. AA. This generates a hypothesis: discipline does not seem to determine sprint ability, so other factors may; unsupervised methods typically generate hypotheses rather than test them
2. BB. The clustering failed, because a valid clustering must reproduce the discipline labels
3. CC. Because the groups are known, this was a supervised classification that tested whether discipline predicts 100-m time
4. DD. The result proves that sprinters and long jumpers have identical sprint ability

Q17 · Encoding categorical variablesfrom the recording

In a football data set with players from 100 countries, a student codes country as one number column (Italy = 1, …, United States = 100) and uses it to predict salary. What is the main problem the lecturer pointed out?

1. AA. R regression functions cannot use numeric codes as predictors
2. BB. The column will contain too many zeros
3. CC. Categorical variables can never be used as predictors, so country must be removed
4. DD. The model treats the codes as amounts, so a higher code looks like 'more' of something and can create a false (spurious) link with salary; indicator (dummy or one-hot) columns avoid this

Q18 · Shallow ML versus deep learningfrom the recording

Why did the lecturer advise students to use shallow (classical) machine learning rather than deep learning for their course data sets?

1. AA. Deep learning cannot be used for regression problems
2. BB. With small, simple data sets deep learning usually does no better, and you lose control: the network picks its own features without domain knowledge and it is very hard to trace why it predicts what it does
3. CC. Shallow ML keeps improving as more data are added, whereas deep learning levels off
4. DD. Deep learning requires the researcher to build all features by hand

Q1 · Precisioncatch-up notes

Injured is the positive class. A classifier has TP = 12, FP = 3, FN = 8 and TN = 77. What is its precision?

1. AA. 60%
2. BB. 80%
3. CC. 89%
4. DD. 12%

Q2 · Bias–variance diagnosiscatch-up notes

A model that classifies training sessions as interval or endurance has a training error of 3% and a validation error of 15% (the validation error is measured on data kept apart from learning). What is the main problem?

1. AA. High variance (overfitting): the training error is low (3%), but the validation error is 12 percentage points higher
2. BB. High bias (underfitting), because a validation error of 15% is high
3. CC. Good generalisation, because both errors are below 20%
4. DD. Nothing can be judged until the model is also run on a separate test set

Q3 · Decision trees: choosing the first splitcatch-up notes

You want a decision tree that predicts whether a runner gets a stress fracture (yes/no) from previous injury (yes/no), sex and age in years. How does the tree choose its first question (the root)?

1. AA. It computes the Gini impurity (how mixed the classes in a group are) for each predictor and starts with the predictor with the highest Gini, because that predictor carries the most information.
2. BB. It leaves age out, because a decision tree can only ask about yes/no predictors, and starts with the yes/no predictor with the lowest Gini.
3. CC. It sorts the runners by age and uses the mean of each pair of neighbouring ages as a candidate cut-off; it computes the weighted Gini for every yes/no predictor and every age cut-off; the split with the lowest weighted Gini becomes the root.
4. DD. It splits age once at the overall mean age, computes the weighted Gini for each predictor, and keeps the split whose child groups are most mixed.

Q4 · Reading R output: the positive classcatch-up notes

A model predicts whether a youth player gets injured this season (Yes/No). Below is R's confusionMatrix() output for the test set; check which class R treats as positive. Which value is the share of the players who really got injured (Yes) that the model also predicted as injured?

Confusion Matrix and Statistics

          Reference
Prediction   No  Yes
        No   54    8
       Yes    6   32

               Accuracy : 0.86
                 95% CI : (0.7763, 0.9213)
    No Information Rate : 0.6
    P-Value [Acc > NIR] : 1.293e-08

                  Kappa : 0.7059

 Mcnemar's Test P-Value : 0.7893

            Sensitivity : 0.9000
            Specificity : 0.8000
         Pos Pred Value : 0.8710
         Neg Pred Value : 0.8421
             Prevalence : 0.6000
         Detection Rate : 0.5400
   Detection Prevalence : 0.6200
      Balanced Accuracy : 0.8500

       'Positive' Class : No

1. AA. 0.90 (the Sensitivity line)
2. BB. 0.80 (the Specificity line)
3. CC. 0.86 (the Accuracy line)
4. DD. 0.85 (the Balanced Accuracy line)

Q5 · Reading R output: class imbalance

A model predicts overtraining (Yes/No) in 200 cyclists; overtraining is rare. Below is R's confusionMatrix() output for the test set. Which conclusion is most appropriate?

Confusion Matrix and Statistics

          Reference
Prediction  Yes   No
       Yes    6    4
        No   14  176

               Accuracy : 0.91
                 95% CI : (0.8615, 0.9458)
    No Information Rate : 0.9
    P-Value [Acc > NIR] : 0.3724

                  Kappa : 0.3571

 Mcnemar's Test P-Value : 0.03389

            Sensitivity : 0.3000
            Specificity : 0.9778
         Pos Pred Value : 0.6000
         Neg Pred Value : 0.9263
             Prevalence : 0.1000
         Detection Rate : 0.0300
   Detection Prevalence : 0.0500
      Balanced Accuracy : 0.6389

       'Positive' Class : Yes

1. AA. Excellent performance, given its accuracy of 0.91: 91% of overtrained cyclists are detected.
2. BB. Excellent performance, given its specificity of 0.98: it rarely misses an overtrained cyclist.
3. CC. Good performance, given the narrow 95% confidence interval of its accuracy.
4. DD. Rather poor performance, given its balanced accuracy of 0.64: only 30% of overtrained cyclists are detected (sensitivity 0.30), and accuracy (0.91) is barely above the no-information rate (0.90).

Q6 · Choosing a metric for a rare outcome

In a running club, about 4% of the young runners get a stress fracture in a season. You build a model that predicts stress fracture (yes/no). Which metric should you use to evaluate it?

1. AA. Accuracy, because it uses all four cells of the confusion matrix
2. BB. Specificity, because almost all runners stay fracture-free
3. CC. F1 score or balanced accuracy, because a model that never predicts a fracture would already reach about 96% accuracy
4. DD. Mean absolute error (MAE), because it shows the average size of the prediction errors

Q7 · Fixing high bias

Your model has a training error of 14% and a validation error of 15%. Following the lecture's bias–variance flowchart, which action is most likely to help?

1. AA. Collect more training data
2. BB. Train a more complex model or add more informative features
3. CC. Increase regularization (a built-in penalty that keeps the model simpler)
4. DD. Reduce the number of features

Q8 · Validation design in caret

This lecture code classifies training sessions as interval or endurance. Which statement is correct?

train_id   <- createDataPartition(model_data$training_type, p = 0.7, list = F, times = 1)
train_data <- model_data[ train_id, ]
test_data  <- model_data[-train_id, ]

fitControl <- trainControl(method = 'cv', number = 10, classProbs = T)

model <- train(training_type ~ ., data = train_data,
               method    = 'rf',
               trControl = fitControl,
               verbose   = F,
               metric    = 'ROC')

1. AA. 70% of the sessions form test_data and 30% form train_data.
2. BB. The 10-fold cross-validation uses test_data as one of its ten folds.
3. CC. Within train_data, each session is used for validation once and for training nine times; test_data is kept aside for the final evaluation.
4. DD. With number = 10 this is leave-one-out cross-validation, because k = 10.

Q9 · Decision trees: splitting a numerical feature

To split on the numerical feature age, you follow the lecture's procedure: sort by age, use the mean of each pair of consecutive ages as a candidate threshold, and compute the weighted Gini impurity for each. Which threshold is chosen, and what is its weighted Gini?

Age:      15   17   20   22   26
Injured:  No   No   Yes  Yes  No

1. AA. Age < 18.5, weighted Gini ≈ 0.27
2. BB. Age < 16, weighted Gini = 0.40
3. CC. Age < 21, weighted Gini ≈ 0.47
4. DD. Age < 20, weighted Gini ≈ 0.27

Q10 · Elbow and silhouettecatch-up notes

You run k-means on standardised (z-scored) fitness data of 60 athletes for k = 1–8. Below are the values of the elbow plot (total within-cluster sum of squares) and of the silhouette plot (average silhouette score). What does each method suggest?

k                       1     2     3     4     5     6     7     8
Within-cluster SS     900   560   330   150   132   118   107    99
Average silhouette      –  0.38  0.44  0.51  0.53  0.60  0.49  0.42

1. AA. Elbow: k = 8, because the within-cluster sum of squares is lowest there. Silhouette: k = 6.
2. BB. Elbow: k = 2, because the sum of squares drops most between k = 1 and k = 2. Silhouette: k = 6.
3. CC. Elbow: k = 4. Silhouette: k = 2, because a lower silhouette score means tighter clusters.
4. DD. Elbow: k = 4. Silhouette: k = 6. The methods disagree, so weigh both (e.g. with NbClust and knowledge of the sport) instead of claiming one proven k.

Q11 · PCA: explained variancecatch-up notes

A PCA on four standardised fitness tests gives eigenvalues (the amount of variance along each component) of 12, 5, 2 and 1. Which statement is correct?

1. AA. PC1 explains 60% of the variance; PC1 and PC2 together explain 85%.
2. BB. PC1 explains 12% of the variance and PC2 explains 5%.
3. CC. PC1 explains 71% of the variance (12/17), so PC1 and PC2 together explain 100%.
4. DD. PC1 classifies 60% of the athletes correctly.

Q12 · PCA: purposecatch-up notes

A sports scientist has 40 fitness and GPS variables for each of 300 players. What is principal component analysis (PCA) mainly used for in this situation?

1. AA. To predict which players will get injured, with injury as the labelled target, which makes PCA supervised learning
2. BB. To assign each player to a cluster, so PCA replaces k-means
3. CC. To fit the ordinary regression line of one variable on another, minimising vertical distances
4. DD. To summarise the 40 variables with a few new perpendicular axes (principal components) that keep as much of the variance as possible, e.g. to plot the players in 2-D or to cluster on fewer variables

Q13 · Case study: clustering cyclists

In van der Zwaard et al. (2019), elite track sprinters, team-pursuit cyclists and road cyclists were clustered with k-means on anthropometry (body size, composition and shape). Which result did the lecture report?

1. AA. Three clusters that exactly matched the three disciplines: sprint, team pursuit and road.
2. BB. A mesomorphic cluster containing the sprinters, and two meso-ectomorphic clusters (tall and short) that each contained both pursuit and road cyclists.
3. CC. Two clusters, sprinters versus all endurance cyclists, chosen because the silhouette peaked at k = 2.
4. DD. The clusters were formed using cycling discipline as the target label, so the method was supervised.

Q18 · Underfittingfrom the recording

Which model did the lecturer describe as the most extreme ('ultimate') example of underfitting?

1. AA. A model so flexible that it passes through every training data point
2. BB. A model without predictors that predicts the training-set mean for every new case
3. CC. A random forest tuned with 10-fold cross-validation
4. DD. A model with 1% training error and 11% validation error

Q19 · Splitting on the targetfrom the recording

You want to classify sessions as extensive or intensive interval training. A random 70/30 holdout split happens to put almost all extensive sessions in the test set. What was the lecturer's advice?

1. AA. Nothing needs to change: a random split is always representative.
2. BB. Switch to leave-one-out cross-validation, which avoids splitting altogether.
3. CC. Make the test set larger (e.g. 50%) so that it contains more sessions of each class.
4. DD. Base the split on the target (e.g. createDataPartition on the training type) so that every class appears in both the training and the test set.

Q20 · Standardising inputs for k-meansfrom the recording

Before k-means clustering of cyclists on height (cm), body mass (kg) and sum of skinfolds (mm), why did the lecturer prefer z-scores over dividing each variable by its maximum?

1. AA. Dividing by the maximum removes the units but not the differences in spread, so variables can still contribute unequally to the distances; z-scores give every variable mean 0 and SD 1, so all contribute equally.
2. BB. Dividing by the maximum changes the order of the athletes on each variable, whereas z-scores keep it.
3. CC. Z-scores reduce the number of variables, like PCA, which speeds up the algorithm.
4. DD. Z-scores give more weight to variables measured in larger units, because those carry more information.

Q21 · Case study: Rommers et al. (2020)from the recording

In Rommers et al. (2020), 50% of players got injured. Test precision, recall and F1 were all 85%, and training values were 84%, 83% and 83%. Which interpretation matches the lecture?

1. AA. The model overfits, because the training and test scores are not identical.
2. BB. Accuracy would have been misleading, because the injury data were strongly imbalanced.
3. CC. The classes were balanced, so the choice of metric mattered little, and the nearly identical training and test scores indicate neither over- nor underfitting.
4. DD. The model underfits, because the test scores are higher than the training scores.

Q1 · Centering vs standardizationcatch-up notes

Before modelling, you subtract the mean from each predictor (centering) but you do not divide by its standard deviation (SD). What does this achieve?

1. AA. Each variable gets mean 0, but the differences in spread between variables stay the same
2. BB. Every variable gets a standard deviation of 1
3. CC. Each variable becomes normally distributed
4. DD. Missing values are replaced with observed values

Q2 · Missing data: interpolationcatch-up notes

A heart-rate sensor records one value per second. One reading is missing between two valid readings. Which approach estimates the missing value from the readings just before and after it?

1. AA. Dummy encoding
2. BB. Numerical binning
3. CC. Interpolation
4. DD. Z-scoring

Q3 · Scaling and KNN

In the lecture example, KNN (k =
2) classified students as male or female from height (training range 1.50–1.96 m) and age (training range 21–38 years). Without any scaling the test accuracy was only 0.46. What is the most likely explanation?

1. AA. KNN requires normally distributed features, and raw height and age are not normal
2. BB. KNN cannot use a numeric feature measured in metres
3. CC. Raw height and age contain missing values that KNN silently drops
4. DD. Distances are dominated by age, whose differences (years) are numerically far larger than height differences (metres), so height hardly affects which neighbours are chosen

Q4 · Z-score and min–max calculation

In a training set, height has mean 170 cm, SD 10 cm, minimum 150 cm and maximum 200 cm. A new player is 165 cm tall. Using these training values, what are the player's z-score and min–max value?

1. AA. z = 0.5; min–max = 0.3
2. BB. z = −0.5; min–max = 0.3
3. CC. z = −5; min–max = 15
4. DD. z = −0.5; min–max = 0.83

Q5 · Standardization vs normalization

A sprint-speed feature contains a few extreme values caused by GPS errors. According to the lecture's cheat sheet (its summary slide), why is min–max normalization more affected by these outliers than z-score standardization?

1. AA. Min–max removes the outliers, whereas z-scoring keeps them
2. BB. Z-scoring converts the feature to a normal distribution, which deletes the outliers
3. CC. Min–max uses the observed minimum and maximum, so one extreme value stretches the range and squeezes all other values into a narrow part of 0–1
4. DD. Min–max can only be applied to features that follow a normal distribution

Q6 · Log transformation

Training-load values 1, 2, 4 and 8 (each double the previous one) are log₂-transformed. Which statement describes the result?

1. AA. They become 0, 0.14, 0.43 and 1: the values are mapped to a fixed 0–1 range
2. BB. They get mean 0 and SD 1, like z-scores
3. CC. The doubling information is lost, because the logarithm removes ratios
4. DD. They become 0, 1, 2 and 3: equal fold changes become equal steps

Q7 · Imputation with mice

The lecture filled in the missing values of the nhanes data (columns age, bmi, hyp, chl) with the code below. Which statement is correct?

my_imp = mice(input_data, m = 5,
              method = c("", "pmm", "logreg", "pmm"),
              maxit = 20)
final_clean = complete(my_imp, 5)

1. AA. maxit = 20 creates 20 imputed data sets, and complete() averages them
2. BB. m = 5 creates five completed versions of the data, and complete(my_imp, 5) returns version number 5
3. CC. "logreg" means that bmi is imputed with logistic regression
4. DD. The empty string "" means that age is imputed with its mean

Q8 · Neural networks vs simple models

A physiotherapist has data from 40 patients and wants a model whose link between the predictors and recovery time she can explain to patients. Based on the lecture's comparison, why is a simple model (e.g. linear regression) a better choice here than a neural network?

1. AA. A neural network fits complex functions but is hard to interpret, needs much more data and needs more attention to bias and variance
2. BB. Neural networks cannot take relationships between inputs into account
3. CC. Simple models automatically capture all interactions between inputs
4. DD. Neural networks can only be used for images

Q9 · Odds ratio calculation

In one season, 20 of 100 footballers with a high training load and 10 of 100 footballers with a low training load got injured. What is the odds ratio (OR) of injury for high versus low load?

1. AA. 0.44
2. BB. 2.0
3. CC. 0.10
4. DD. 2.25

Q10 · Probability, odds and odds ratio

Which statement about probability, odds and the odds ratio (OR) is correct?

1. AA. Probability and odds both range from 0 to 1, so an OR can never exceed 1
2. BB. An event with probability 0.75 has odds of 3, and an OR of 1 means the odds are the same in the exposed and the unexposed group
3. CC. An OR of 0.5 means that the exposure halves the probability of the outcome
4. DD. Likelihood is the probability that a future event will occur in an individual

Q11 · Absolute vs relative risk

A headline says a supplement 'doubles the risk' of a rare side effect. In the study, 2 of 10,000 users and 1 of 10,000 non-users had the side effect. Which statement reports the result the way the lecture recommends?

1. AA. Risk rose from 0.01% to 0.02% (RR = 2): about 1 extra case per 10,000 users
2. BB. The risk increased by 100 percentage points
3. CC. Because RR = 2, about half of all users will get the side effect
4. DD. The odds ratio is 2, so each user has a 2% chance of the side effect

Q14 · Scaling in the lifecyclefrom the recording

According to the lecturer, in which phase of the data science lifecycle does a log transformation of a right-skewed feature belong?

1. AA. Data preparation, because every change to a feature's values counts as cleaning the data
2. BB. Feature engineering, because it changes not only the scale but also the shape (variability) of the feature's distribution
3. CC. Data exploration, because it is mainly done to make histograms easier to read
4. DD. Data modelling, because the transformation is part of fitting the KNN model

Q15 · Neural network nodefrom the recording

A single node of a neural network decides whether an athlete trains today. Inputs: slept well x1 = 1, no muscle soreness x2 = 1, rain x3 = 0. Weights: w1 = 2, w2 = 3, w3 = 4. Threshold = 6. As in the lecture's surfing example, the node outputs 1 if the weighted sum minus the threshold is greater than 0, otherwise 0. What does the node compute?

1. AA. 2 + 3 + 4 − 6 = 3, so the output is 1
2. BB. 2 + 3 + 6 = 11, so the output is 1
3. CC. 5 ÷ 6 ≈ 0.83, an 83% probability that the athlete trains
4. DD. 2 + 3 + 0 − 6 = −1, so the output is 0

Q16 · Missing data optionsfrom the recording

In a match-statistics data set, the column 'goals scored' shows NaN for every player who did not score in a match. Which approach matches the lecturer's advice?

1. AA. Correct the wrong values: replace these NaN entries with 0, because the true value is known to be zero
2. BB. Delete all rows that contain NaN, so that only complete matches remain
3. CC. Impute them with mice (pmm), because NaN entries must always be treated as unknown
4. DD. Replace them with the mean number of goals, the simplest form of imputation

Q17 · Multiple imputationfrom the recording

You filled in missing values with mice using m = 5. According to the lecturer, what could you learn by running your model on each of the five completed data sets?

1. AA. Which version contains the true values, namely the one that gives the highest model accuracy
2. BB. Nothing: the five data sets are identical and m only sets the number of fill-in rounds
3. CC. How uncertain the imputation is: very different model results mean the filled-in values are a big guess, similar results mean they are fairly stable
4. DD. Whether the imputed variables are normally distributed, which pmm requires

Q1 · Exposure data choicecatch-up notes

You study the link between long-term (10-year) NO₂ exposure and death from any cause in adults across the Netherlands. NO₂ comes mostly from traffic and changes strongly over short distances. Which exposure data fit this question best?

1. AA. One national annual mean NO₂ concentration assigned to every participant
2. BB. Hourly NO₂ from a single monitoring station in the centre of the country
3. CC. Long-term NO₂ concentrations predicted by a land-use regression (LUR) model at each participant's home address
4. DD. One week of personal NO₂ monitoring in a single participant

Q2 · Hazard, exposure and risk

A cyclist commutes for 2 hours a day along a busy road with high NO₂ levels. Using the lecture's definitions, which combination is correct?

1. AA. Hazard = the 2 hours per day on the road; exposure = NO₂; risk = the cyclist's lung function
2. BB. Hazard = NO₂ (something with the potential to harm); exposure = how much NO₂ the cyclist meets and for how long; risk = the chance that this causes harm
3. CC. Hazard = the chance of getting ill; exposure = NO₂; risk = the duration of cycling
4. DD. Hazard, exposure and risk are synonyms that all describe the NO₂ level

Q3 · Risk assessment paradigm

In the risk assessment paradigm, which step examines the probability and severity of a health outcome as a function of the dose (exposure)?

1. AA. Hazard identification
2. BB. Exposure assessment
3. CC. Risk characterization
4. DD. Hazard characterization

Q4 · Exposure assessment: residential proxy

Why does the lecture's study use the average NO₂ level at each participant's home address instead of true personal exposure?

1. AA. Measuring true exposure in all micro-environments (home, work, commuting) is not feasible for a large population, so the home-address average is used as a proxy
2. BB. Home-address levels equal personal exposure, because people spend all their time at home
3. CC. Personal exposure is irrelevant for long-term health effects
4. DD. There is a monitoring station at every participant's address

Q5 · Exposure measurement sources

A researcher needs measurements of a pollutant that the national network does not measure, at specific sites near schools. Which approach fits, and what is its main drawback according to the lecture?

1. AA. Routine monitoring network; drawback: the data are not continuous
2. BB. Remote sensing (satellites); drawback: it can only measure indoor air
3. CC. Designed campaign (own fixed sites or mobile monitors); drawback: high cost, and it cannot cover large populations
4. DD. Routine monitoring network; drawback: it is too expensive

Q6 · Land-use regression (LUR)

In the case study's land-use regression (LUR) model, what is the outcome variable and what are the predictors?

1. AA. Outcome: mortality; predictor: NO₂ concentration
2. BB. Outcome: NO₂ measured at 69 monitoring stations; predictors: traffic and land-use features around those stations
3. CC. Outcome: traffic intensity; predictor: NO₂ concentration at participants' homes
4. DD. Outcome: NO₂ computed from emission sources with physical laws; no regression coefficients are estimated

Q7 · Health data sources

Which health data source does the lecture describe with these pros and cons: 'earlier in the chain; possibly less bias than clinical outcomes' versus 'can be invasive; expensive/time-consuming; instrument/observer effects'?

1. AA. Physiological measurements (e.g. biomarkers)
2. BB. Routine databases (e.g. a mortality registry)
3. CC. Questionnaires
4. DD. Remote sensing

Q8 · Correlation and confounding

Why does the lecture call the correlation coefficient between NO₂ and mortality 'not an interesting measure of association'?

1. AA. Because a correlation cannot be computed when the outcome is yes/no
2. BB. Because the correlation would always be close to zero in a large cohort
3. CC. Because NO₂ is measured in µg/m³ and a correlation requires unitless data
4. DD. Because a third factor linked to both exposure and outcome (a confounder, e.g. smoking or income) can create or distort the association, and a correlation does not adjust for it

Q9 · Hazard ratio interpretation

The adjusted Cox model reports HR = 1.04 (95% CI 1.03–1.04), p < 0.001, for a fixed step (increment) in NO₂ exposure. Which interpretation is correct?

1. AA. Each step higher in NO₂ is associated with a 4% higher hazard (instantaneous death rate), after adjustment for confounders
2. BB. Each step higher in NO₂ raises the 10-year chance of dying by 4 percentage points
3. CC. NO₂ causes 4% of all deaths in the Netherlands
4. DD. There is no association, because 1.04 is close to 1

Q10 · Cox model in R

The lecture's Cox model formula is shown below (shortened). Which statement is correct?

Surv(age_b, age_end, mort) ~ NO2_exposure + strata(sex) +
  as.numeric(smoke_dur) + as.factor(smoking) + as.factor(bmi_cat) +
  as.factor(edulev) + as.numeric(meaninc_neighbor_2011) + ...

1. AA. mort is the NO₂ exposure, and age_b and age_end are confounders
2. BB. NO2_exposure is the outcome that the model predicts
3. CC. age_b and age_end give each participant's age at the start and end of follow-up, mort shows whether the participant died, NO2_exposure is the exposure and the remaining terms adjust for confounders
4. DD. strata(sex) removes all women from the analysis

Q13 · LUR model validationfrom the recording

After fitting a land-use regression (LUR) model for NO₂ on the 69 Dutch monitoring stations, the guest lecturer described the usual way to check that the model predicts well. Which procedure did she describe?

1. AA. Correlate predicted NO₂ at participants' homes with their mortality; a significant hazard ratio confirms the exposure model
2. BB. Keep adding land-use predictors until R² on all 69 stations is close to 1
3. CC. Fit the model on part of the stations (e.g. 40), predict NO₂ at the remaining stations, and compare predicted with measured values
4. DD. Replace the LUR model with a dispersion model and check whether both produce the same map

Q14 · Deterministic vs statistical exposure modelsfrom the recording

Which statement correctly distinguishes the two approaches the lecture described for predicting NO₂ at addresses without a monitoring station?

1. AA. Dispersion (chemical transport) models start from pollution sources and their emissions and apply physical and chemical laws; land-use regression ignores the sources and relates measured concentrations to the surrounding land use
2. BB. Dispersion models start from concentrations measured at monitoring stations; land-use regression starts from how much each pollution source emits
3. CC. Both are deterministic models, but land-use regression also uses satellite data
4. DD. Land-use regression is based on physical laws of air flow, whereas dispersion models are statistical regressions on traffic and green space

Q15 · Routine health databasesfrom the recording

In the Dutch mortality registry, more deaths are recorded on Mondays than on Sundays. According to the guest lecture, what does this illustrate?

1. AA. Confounding, because the day of the week is related to both NO₂ exposure and mortality
2. BB. People really die more often on Mondays, e.g. because of weekday NO₂ peaks from commuter traffic
3. CC. Instrument or observer effects, a typical problem of physiological measurements
4. DD. Routine health data depend on administrative factors (registration practice, medical shifts), not only on when deaths really occur

Q16 · Health data categoriesfrom the recording

A researcher cannot access diagnoses because medical records are too sensitive, but can obtain how much each person spent on specific therapies per year. In the lecture's grouping of health data (Galetsi et al., 2020), what type of data is this, and how can it be used?

1. AA. Pharmaceutical data; it measures the toxic effects of the therapies
2. BB. Administrative data; it can serve as a proxy (stand-in) for clinical data on who had which condition
3. CC. Real-time patient data; it continuously tracks each person's health
4. DD. Clinical data; it directly records clinical endpoints such as new cases of disease

Q1 · Challenges of user-generated datacatch-up notes

A wearable company has millions of heart-rate measurements and all the analysis software it needs. But the heart-rate data are noisy, and there is no verified record of which users actually got ill. What is the main obstacle to building a useful illness model?

1. AA. Trustworthy measurements and a trustworthy reference outcome (the verified 'answer' a model learns from)
2. BB. Choosing between R and Python for the analysis
3. CC. Recruiting more users, because sample size is the main limitation of user-generated data
4. DD. Computing time, because millions of measurements are too many for a model

Q2 · APIcatch-up notes

What is an API (application programming interface) in the context of sport and health data?

1. AA. A quality score that a wearable reports with every heart-rate reading
2. BB. A storage system in which all raw wearable data are kept for later analysis
3. CC. A defined connection through which one program automatically asks another program for data, e.g. an app fetching a user's runs from Strava
4. DD. A sensor type that measures heart-rate variability

Q3 · From data to knowledge: three steps

A company wants to use heart-rate variability (HRV) measured with the phone camera to study stress in thousands of users. In which order should the work proceed, according to the lecture?

1. AA. Discover new relations → use the app at scale → validate the measurement
2. BB. Use the app at scale → discover new relations → validate the measurement only if results look odd
3. CC. Validate the measurement → discover new relations → confirm a result already known from lab studies
4. DD. Validate the measurement (phone camera vs ECG) → use the app at scale and confirm a result already known from lab studies → discover new relations or build products

Q4 · HRV and context

A user's morning HRV (rMSSD) is well below their normal range, and their resting heart rate is higher than usual. Based on the lecture's HRV4Training data, what is the best interpretation?

1. AA. The user is ill, because only sickness lowers HRV
2. BB. Several stressors (hard training, sickness, heavy alcohol intake) cause similar changes, so context such as the user's tags is needed to know the cause
3. CC. The measurement must be faulty, because HRV and heart rate always change in the same direction
4. DD. The user most likely did an easy training session the day before

Q5 · Limitations of typical lab studies

A lab study measured how HRV responds to training intensity in N = 10 male students. Which limitation does the lecture emphasize?

1. AA. The results may hold only for that sample (not necessarily for women, other ages, health conditions or sports), and checking other groups requires another costly study
2. BB. The data were of low quality because they were collected in the lab
3. CC. The study had no dependent variable
4. DD. Ten participants are too many to analyse without machine learning

Q6 · Noisy data: estimating max heart rate

Without lab tests, an app estimates each user's maximum heart rate from the highest heart rates in their everyday workouts, so it can express training intensity as a percentage of that maximum. Which assumption does this need?

1. AA. That the heart-rate sensor never produces values above 300 bpm
2. BB. That every user's maximum heart rate equals the value estimated from age
3. CC. That the monitored period contains some genuinely hard sessions, so it must be long enough
4. DD. That users report their maximum heart rate after every workout

Q7 · Missing data and selection bias

To simplify the analysis, you remove all users with missing workouts from a large app data set. What is the main risk the lecture points out?

1. AA. None, because user-generated data sets are so large that removing users never matters
2. BB. Bias: the users who remain may differ systematically from those removed, so you no longer have the full picture
3. CC. The remaining data become noisier, because removing users adds measurement error
4. DD. The model can no longer be tested with cross-validation

Q8 · Reference data

Which question is NOT on the lecture's list of things to ask about reference data (the outcomes a model learns from) before collecting data in an app?

1. AA. What are the outcomes, and can we track them?
2. BB. Are we asking too much of the user, given that it is not a clinical study?
3. CC. Is it ethical to collect them?
4. DD. Can missing outcomes be replaced by simply collecting more sensor data?

Q9 · Running performance example

In the lecture's running study (about 2,100 runners), groups of features were added step by step to a regression model that predicts each runner's best 10 km time. Which statement matches the results?

1. AA. Adding previous performance gave the highest R² (0.87), well above the resting-physiology feature set (R² = 0.33)
2. BB. Resting physiology (morning heart rate and HRV) gave the best prediction (R² = 0.87)
3. CC. The model is a classifier that is correct for 87% of the runners
4. DD. Adding more workouts kept lowering the error, so all available training history should be used

Q10 · Pandemic example

During the first COVID-19 lockdown (March–May 2020), most people in the lecturer's poll (67.6%) expected resting heart rate to rise. What did data from about 5,500 app users show?

1. AA. Resting heart rate rose, as most people expected
2. BB. Resting heart rate did not change compared with 2019
3. CC. Resting heart rate fell compared with 2019, while travel dropped and sleep time rose
4. DD. Resting heart rate fell because people slept less

Q11 · Why user-generated data matters

HRV4Training (managing stressors), Bloomlife (start of labour) and Oura (spotting infections) were the lecture's examples. What do these uses have in common?

1. AA. Each was the original goal of the app or wearable
2. BB. Each was developed in a small lab study with N = 2–10
3. CC. Each relies on the hardware alone, without data science
4. DD. None was the original goal; they became possible through context recorded over time and reference points (key events from APIs or user reports)

Q15 · Sickness: detection vs predictionfrom the recording

A wearable company says its ring 'predicts' infections, because users' resting heart rate rose before they reported their first symptoms. How did the guest lecturer judge such a claim?

1. AA. It is real prediction, because the heart-rate change came before the reported symptoms
2. BB. It is invalid, because resting heart rate and HRV do not change during an infection
3. CC. At best it is detection: once heart rate changes, the body is already sick, and the reference point (infection day, symptom day, test day) is itself uncertain
4. DD. It becomes prediction once the sample contains millions of users, because large samples remove timing errors

Q16 · False positives at scalefrom the recording

An app's sickness alert has a false-positive rate of 5% (5 in every 100 healthy users get a wrong alarm). Why is this a bigger problem for an app with 1,000,000 healthy users than for a study with 20 healthy participants?

1. AA. The rate stays 5%, but the number of false alarms grows: about 1 person in the study versus 50,000 app users (enough to flood family doctors)
2. BB. The false-positive rate itself rises as the number of users grows
3. CC. False alarms average out in large samples, so they are only a problem in the small study
4. DD. With a million users, false positives turn into false negatives, so real cases are missed instead

Q17 · Lifestyle stressors vs trainingfrom the recording

In the HRV4Training data, how did the morning changes in HRV and resting heart rate after sickness or heavy alcohol intake compare with those after hard training, and what did the lecturer conclude?

1. AA. They were smaller than after hard training, so a training app can safely ignore the user's 'sick' or 'alcohol' tags
2. BB. They were about the same size as after hard training, so these tags add nothing for a training app
3. CC. They went in the opposite direction (HRV up, heart rate down), so they cancel out the effect of hard training
4. DD. They were several times larger than after training, so a training tool that ignores these stressors will mostly be reading them instead of training

Q18 · Max heart rate from free-living datafrom the recording

You estimate users' maximum heart rate from months of everyday workouts. According to the guest lecturer, how can the data reveal users who probably never trained hard, and when is this a serious problem?

1. AA. Their highest recorded heart rate is above 300 bpm; it is serious whenever even one such user exists
2. BB. Their session heart rates have a very narrow spread (always about the same intensity); it is serious when such users are a large share of the sample, much less so when they are a handful among a million users
3. CC. Their age-based maximum is higher than their recorded maximum; it is never a problem, because the age-based value can simply replace the recorded one
4. DD. They have missing workouts; it is only a problem in small samples, because missing data in large samples are always random

Q1 · Scaling + validation for KNN

You build a KNN (k-nearest neighbours) classifier that predicts injury (yes/no) from weekly running distance (0–120 km) and sleep duration (5–10 h). Which workflow is most appropriate?

1. AA. Train KNN on the raw features and evaluate it on the training set
2. BB. Split the data, scale the features (z-scores or min–max) using numbers from the training set, fit KNN on the training set, and evaluate it on the unseen test set or with k-fold cross-validation
3. CC. Scale the whole data set, choose k by the highest test-set accuracy, and report that accuracy as the expected performance on new athletes
4. DD. Skip scaling, because KNN is a tree-based model that is not affected by units

Q2 · Imbalanced outcome from wearable data

An app predicts whether a user will tag a day as 'sick'; sickness is tagged on about 3% of days. A model that always predicts 'not sick' reaches 97% accuracy. What should you do?

1. AA. Accept the model, because 97% accuracy is high
2. BB. Evaluate it with R², because the outcome is yes/no
3. CC. Use a metric suited to imbalanced classes, such as the F1 score or balanced accuracy, and check that the 'sick' tags (the reference data) are reliable and correctly timed
4. DD. Use the mean absolute error, because sickness is a number

Q3 · Choosing the association measure

Match each design to the most suitable measure of association. (1) 100 injured and 100 uninjured runners are asked which shoe type they use. (2) 5,000 adults are followed for up to 10 years, with different follow-up times, and dates of death are recorded.

1. AA. (1) Odds ratio; (2) hazard ratio from a Cox model
2. BB. (1) Relative risk; (2) odds ratio
3. CC. (1) Hazard ratio; (2) correlation coefficient
4. DD. (1) Absolute risk; (2) odds ratio

Q4 · Preprocessing free-living data

In a free-living running data set, 25% of users have missing resting-heart-rate values, and the heart-rate data contain impossible readings above 300 bpm. Which preprocessing plan is most defensible?

1. AA. Delete every user with any missing value; the data set is large, so this cannot bias the results
2. BB. Replace the impossible readings and the missing values with 0, so no rows are lost
3. CC. Keep the impossible readings, because z-scoring will later remove their influence
4. DD. Change the impossible readings to missing, then consider imputation (e.g. with mice) and check whether users with missing data differ from the others

Q7 · Relative vs absolute risk

In app data from 4,000 runners, 6% of months in which morning HRV was more than 10% below the runner's normal range ended in an injury, against 2% of other months. What are the relative risk (low-HRV months compared with other months) and the absolute difference in risk?

1. AA. Relative risk 4; absolute difference 3 percentage points
2. BB. Relative risk 0.33; absolute difference 4 percentage points
3. CC. Relative risk 3; absolute difference 4 percentage points (4 extra injured months per 100)
4. DD. Relative risk 3; absolute difference 6 percentage points