Artikel 1 – Carey et al. (2020) “Ten simple rules for reading a
scientific paper”
1. Carey et al. emphasize that readers should “pick your reading goal” and “understand the authors’
goal.” Explain these two rules and describe, with a concrete example related to a
big-data/consumer-analytics article, how they would change which sections you prioritize when
reading.
In Rule 1, “pick your reading goal,” Carey et al. argue that readers should first decide what
they want to get out of an article, because that intention should determine which parts they read
closely and which they skim. In Rule 2, “understand the authors’ goal,” they emphasize that readers
should also consider why the authors wrote the paper and what kind of article it is (for example, a
methods paper, a resource paper, or a research article), because this shapes how results and
interpretations are presented. For example, in a big‑data consumer‑analytics paper that introduces a
new machine‑learning model to predict online purchases, if my goal is to learn the technique, I would
prioritize the Methods and any supplemental sections on model architecture, feature engineering,
and evaluation, while skimming the Introduction and Discussion for high‑level motivation and
implications; if the authors’ goal is mainly to release a large dataset as a resource, I would instead
focus on the data description, sampling and preprocessing details, and any limitations of the dataset,
using the modeling Results only as illustrative examples rather than as definitive substantive
findings.
2. According to the “six questions” rule, readers should systematically interrogate each paper and each
figure. List the six questions and illustrate how you would apply them to a specific figure reporting
prediction performance of a machine-learning model in a social media study. [5 pts]
According to Rule 3, readers should ask six questions of each paper and each figure: (1)
What do the authors want to know (motivation)? (2) What did they do (approach/methods)? (3) Why
was it done that way (context within the field)? (4) What do the results show (figures and data
tables)? (5) How did the authors interpret the results (interpretation/discussion)? and (6) What
should be done next? Applied to a specific figure showing the prediction performance of a
machine‑learning model in a social media study, I would first identify the motivation (for example,
predicting users’ purchase intent from their likes and posts), then examine what was done (which
model, which training/validation split, which performance metric such as AUC or F1), and why those
modeling choices and baselines were appropriate in light of prior work on social media analytics. I
would then carefully read the axes, error bars, and comparisons in the figure to see what the data
actually show, compare this to how the authors describe the performance in the Results or
Discussion, and finally think about what should come next, such as testing generalizability to another
platform, assessing fairness across demographic groups, or comparing with simpler interpretable
models.
3. Carey et al. discuss the different functions of article sections such as Introduction, Methods, Results,
and Discussion. Choose two of these sections and describe both (a) the intended role of the section
and (b) one common pitfall a novice reader might fall into when interpreting that section in a big data
paper. [5 pts]
Carey et al. describe the Introduction as the section that presents the research question and
explains why it is important, giving newcomers to a field a brief “crash course” in the relevant
background and motivating the study in relation to existing literature. A common pitfall for novice
readers of big‑data papers is to take the Introduction’s story about novelty and importance at face
1
, value, for example assuming that analyzing billions of social media posts automatically makes the
contribution groundbreaking, without noticing that key prior work or critical debates about ethics and
bias might be missing or underplayed. The Methods section is described as the place where the
authors explain what was done and how it was done, ideally in enough detail that someone could
replicate the study, although in practice many details are condensed or moved to supplemental
materials, especially in big‑data or modeling papers. A typical novice pitfall here is to skim or ignore
the Methods because it seems technical, and then assume the results are robust just because the
dataset is large, overlooking crucial issues like sampling, data cleaning, feature selection, and
evaluation procedures that may limit generalizability or introduce systematic bias in big‑data
consumer‑analytics studies.
4. The authors argue that scientific papers are not “truths etched in stone” and stress the importance of
critical but kind reading. Describe one cognitive bias that can affect readers’ judgment of a paper
(either confirming or disconfirming their own views) and propose a concrete strategy, based on the
rules in the article, to reduce that bias when evaluating empirical results. [5 pts]
Carey et al. emphasize that scientific papers, even those in high‑impact journals or written by
prominent researchers, are not “truths etched in stone,” and they warn that readers bring their own
expectations and biases to the interpretation of results. One cognitive bias that can affect judgment
is expectancy or self‑fulfilling prophecy: readers may find a paper particularly compelling because it
supports what they already believe, or may discount it too quickly when it challenges their prior
assumptions, much like the expectancy effects described in psychological research. A concrete
strategy to reduce this bias, consistent with their rules, is to first work through the six questions for
the central figures—especially those reporting key empirical results—before reading the authors’
Discussion, so that the reader forms an independent view of what the data show. Then, following
Rule 6 (“Be critical”) and Rule 7 (“Be kind”), the reader can explicitly ask whether there are other
plausible explanations, note both strengths and limitations, and compare their own initial
interpretation with the authors’ claims in a deliberate, respectful way, rather than simply accepting or
rejecting the conclusions because they align or conflict with pre‑existing beliefs.
5. Carey et al. use the metaphor of building a Lego wall to describe how individual papers contribute to
cumulative science. Using this metaphor, explain how you would integrate findings from Youyou et
al. (computer vs. human judgment) and Kosinski et al. (Likes predicting traits) to motivate a new
research question in consumer analytics using big data. [5 pts]
In Rule 10, Carey et al. use the Lego wall metaphor to explain that each paper is like a brick
in a larger structure of cumulative science, and that building strong “walls” requires clicking together
multiple studies rather than relying on isolated findings. Applying this metaphor to
consumer‑analytics research using big data, the findings from Youyou et al. on computer versus
human judgment and from Kosinski et al. on predicting traits from Facebook Likes can be treated as
two adjacent Lego bricks: one brick shows that algorithmic models trained on digital traces can, in
some cases, outperform human judges, and the other brick shows that social‑media behavior
encodes stable psychological traits that are predictable from large‑scale like data. Using these
bricks, a new research question becomes the empty brick placed on top, such as: “Under what
conditions do big‑data machine‑learning models that infer psychological traits from consumers’
social‑media footprints provide added value over human marketing experts’ judgments when
predicting consumer preferences or responses to personalized advertising, and how should these
models be designed and communicated to ensure fair and trustworthy use?” This question explicitly
builds upward from the existing Lego wall by integrating the insights about algorithm–human
performance differences and the predictive power of Likes into a next step that is specific to
consumer analytics.
2
, Artikel 2 – Oswald & Putka (2017) “Big data methods in the
social sciences”
1. Oswald and Putka distinguish between supervised and unsupervised big data methods. Define each
type, give one concrete example technique, and illustrate with a consumer-analytics example (e.g.,
predicting customer churn vs. segmenting customers) why a researcher would choose one over the
other. [5 pts]
Oswald and Putka define supervised big data methods as approaches where there is a
specific criterion to be predicted, such as job performance or well-being, using many predictor
variables; examples include random forests, gradient-boosted trees, neural networks, and support
vector machines. Unsupervised methods, by contrast, have no outcome to predict and instead aim
to discover structure or clusters in the data, using techniques such as principal components analysis,
k‑means clustering, or hierarchical partitioning. In a consumer‑analytics context, predicting customer
churn from rich CRM and clickstream data would typically use supervised methods (for example, a
gradient‑boosted tree predicting churn as a binary criterion), because the research goal is to
maximize predictive accuracy for a known outcome, whereas segmenting customers into
psychologically or behaviorally similar groups for targeted marketing would use unsupervised
methods such as k‑means clustering, because the aim is to uncover meaningful segments without
predefined labels.
2. Explain what is meant by the “curse of dimensionality” in big data contexts. Describe why this is a
problem when the number of predictors is very large relative to the number of cases, and name one
methodological strategy (with an example method) that helps address it. [5 pts]
The “curse of dimensionality,” as highlighted in Oswald and Putka’s summary of Domingos,
refers to the idea that as the number of predictor dimensions grows, the data become increasingly
sparse in the high‑dimensional space, so even “big” datasets may not provide enough information to
reliably populate all combinations of predictors. This sparsity makes it easy for complex models to
overfit noise, especially when the number of predictors is very large relative to the number of cases,
because the model can find spurious patterns that do not generalize to new data. One
methodological strategy to address this problem is to use regularization methods such as lasso or
elastic net regression, which shrink or set many coefficients to zero and effectively perform variable
selection; for example, a lasso model predicting customer lifetime value from thousands of
behavioral and demographic features can reduce the effective dimensionality while maintaining or
improving out‑of‑sample predictive performance
3. Domingos (2012), as summarized by Oswald and Putka, argues that “simpler algorithms based on
more data tend to beat cleverer algorithms based on less data.” Describe the logic behind this claim
and give a concrete example in a marketing or HR analytics context where this principle would
influence your model choice. [5 pts]
Domingos’ claim, as summarized by Oswald and Putka, that “simpler algorithms based on
more data tend to beat cleverer algorithms based on less data” captures the idea that, in many
practical settings, the key to good prediction is having large, representative samples rather than
using highly complex models on small or noisy datasets. The logic is that with more data, even
relatively simple models such as regularized linear or logistic regression can capture stable patterns
and avoid overfitting, whereas complex algorithms with many tuning parameters can easily fit
idiosyncrasies when data are limited. In a marketing analytics context, this principle might lead a
researcher to prefer a regularized logistic regression or random forest model trained on millions of
ad‑impression and purchase records to predict click‑through, rather than a highly intricate
3
scientific paper”
1. Carey et al. emphasize that readers should “pick your reading goal” and “understand the authors’
goal.” Explain these two rules and describe, with a concrete example related to a
big-data/consumer-analytics article, how they would change which sections you prioritize when
reading.
In Rule 1, “pick your reading goal,” Carey et al. argue that readers should first decide what
they want to get out of an article, because that intention should determine which parts they read
closely and which they skim. In Rule 2, “understand the authors’ goal,” they emphasize that readers
should also consider why the authors wrote the paper and what kind of article it is (for example, a
methods paper, a resource paper, or a research article), because this shapes how results and
interpretations are presented. For example, in a big‑data consumer‑analytics paper that introduces a
new machine‑learning model to predict online purchases, if my goal is to learn the technique, I would
prioritize the Methods and any supplemental sections on model architecture, feature engineering,
and evaluation, while skimming the Introduction and Discussion for high‑level motivation and
implications; if the authors’ goal is mainly to release a large dataset as a resource, I would instead
focus on the data description, sampling and preprocessing details, and any limitations of the dataset,
using the modeling Results only as illustrative examples rather than as definitive substantive
findings.
2. According to the “six questions” rule, readers should systematically interrogate each paper and each
figure. List the six questions and illustrate how you would apply them to a specific figure reporting
prediction performance of a machine-learning model in a social media study. [5 pts]
According to Rule 3, readers should ask six questions of each paper and each figure: (1)
What do the authors want to know (motivation)? (2) What did they do (approach/methods)? (3) Why
was it done that way (context within the field)? (4) What do the results show (figures and data
tables)? (5) How did the authors interpret the results (interpretation/discussion)? and (6) What
should be done next? Applied to a specific figure showing the prediction performance of a
machine‑learning model in a social media study, I would first identify the motivation (for example,
predicting users’ purchase intent from their likes and posts), then examine what was done (which
model, which training/validation split, which performance metric such as AUC or F1), and why those
modeling choices and baselines were appropriate in light of prior work on social media analytics. I
would then carefully read the axes, error bars, and comparisons in the figure to see what the data
actually show, compare this to how the authors describe the performance in the Results or
Discussion, and finally think about what should come next, such as testing generalizability to another
platform, assessing fairness across demographic groups, or comparing with simpler interpretable
models.
3. Carey et al. discuss the different functions of article sections such as Introduction, Methods, Results,
and Discussion. Choose two of these sections and describe both (a) the intended role of the section
and (b) one common pitfall a novice reader might fall into when interpreting that section in a big data
paper. [5 pts]
Carey et al. describe the Introduction as the section that presents the research question and
explains why it is important, giving newcomers to a field a brief “crash course” in the relevant
background and motivating the study in relation to existing literature. A common pitfall for novice
readers of big‑data papers is to take the Introduction’s story about novelty and importance at face
1
, value, for example assuming that analyzing billions of social media posts automatically makes the
contribution groundbreaking, without noticing that key prior work or critical debates about ethics and
bias might be missing or underplayed. The Methods section is described as the place where the
authors explain what was done and how it was done, ideally in enough detail that someone could
replicate the study, although in practice many details are condensed or moved to supplemental
materials, especially in big‑data or modeling papers. A typical novice pitfall here is to skim or ignore
the Methods because it seems technical, and then assume the results are robust just because the
dataset is large, overlooking crucial issues like sampling, data cleaning, feature selection, and
evaluation procedures that may limit generalizability or introduce systematic bias in big‑data
consumer‑analytics studies.
4. The authors argue that scientific papers are not “truths etched in stone” and stress the importance of
critical but kind reading. Describe one cognitive bias that can affect readers’ judgment of a paper
(either confirming or disconfirming their own views) and propose a concrete strategy, based on the
rules in the article, to reduce that bias when evaluating empirical results. [5 pts]
Carey et al. emphasize that scientific papers, even those in high‑impact journals or written by
prominent researchers, are not “truths etched in stone,” and they warn that readers bring their own
expectations and biases to the interpretation of results. One cognitive bias that can affect judgment
is expectancy or self‑fulfilling prophecy: readers may find a paper particularly compelling because it
supports what they already believe, or may discount it too quickly when it challenges their prior
assumptions, much like the expectancy effects described in psychological research. A concrete
strategy to reduce this bias, consistent with their rules, is to first work through the six questions for
the central figures—especially those reporting key empirical results—before reading the authors’
Discussion, so that the reader forms an independent view of what the data show. Then, following
Rule 6 (“Be critical”) and Rule 7 (“Be kind”), the reader can explicitly ask whether there are other
plausible explanations, note both strengths and limitations, and compare their own initial
interpretation with the authors’ claims in a deliberate, respectful way, rather than simply accepting or
rejecting the conclusions because they align or conflict with pre‑existing beliefs.
5. Carey et al. use the metaphor of building a Lego wall to describe how individual papers contribute to
cumulative science. Using this metaphor, explain how you would integrate findings from Youyou et
al. (computer vs. human judgment) and Kosinski et al. (Likes predicting traits) to motivate a new
research question in consumer analytics using big data. [5 pts]
In Rule 10, Carey et al. use the Lego wall metaphor to explain that each paper is like a brick
in a larger structure of cumulative science, and that building strong “walls” requires clicking together
multiple studies rather than relying on isolated findings. Applying this metaphor to
consumer‑analytics research using big data, the findings from Youyou et al. on computer versus
human judgment and from Kosinski et al. on predicting traits from Facebook Likes can be treated as
two adjacent Lego bricks: one brick shows that algorithmic models trained on digital traces can, in
some cases, outperform human judges, and the other brick shows that social‑media behavior
encodes stable psychological traits that are predictable from large‑scale like data. Using these
bricks, a new research question becomes the empty brick placed on top, such as: “Under what
conditions do big‑data machine‑learning models that infer psychological traits from consumers’
social‑media footprints provide added value over human marketing experts’ judgments when
predicting consumer preferences or responses to personalized advertising, and how should these
models be designed and communicated to ensure fair and trustworthy use?” This question explicitly
builds upward from the existing Lego wall by integrating the insights about algorithm–human
performance differences and the predictive power of Likes into a next step that is specific to
consumer analytics.
2
, Artikel 2 – Oswald & Putka (2017) “Big data methods in the
social sciences”
1. Oswald and Putka distinguish between supervised and unsupervised big data methods. Define each
type, give one concrete example technique, and illustrate with a consumer-analytics example (e.g.,
predicting customer churn vs. segmenting customers) why a researcher would choose one over the
other. [5 pts]
Oswald and Putka define supervised big data methods as approaches where there is a
specific criterion to be predicted, such as job performance or well-being, using many predictor
variables; examples include random forests, gradient-boosted trees, neural networks, and support
vector machines. Unsupervised methods, by contrast, have no outcome to predict and instead aim
to discover structure or clusters in the data, using techniques such as principal components analysis,
k‑means clustering, or hierarchical partitioning. In a consumer‑analytics context, predicting customer
churn from rich CRM and clickstream data would typically use supervised methods (for example, a
gradient‑boosted tree predicting churn as a binary criterion), because the research goal is to
maximize predictive accuracy for a known outcome, whereas segmenting customers into
psychologically or behaviorally similar groups for targeted marketing would use unsupervised
methods such as k‑means clustering, because the aim is to uncover meaningful segments without
predefined labels.
2. Explain what is meant by the “curse of dimensionality” in big data contexts. Describe why this is a
problem when the number of predictors is very large relative to the number of cases, and name one
methodological strategy (with an example method) that helps address it. [5 pts]
The “curse of dimensionality,” as highlighted in Oswald and Putka’s summary of Domingos,
refers to the idea that as the number of predictor dimensions grows, the data become increasingly
sparse in the high‑dimensional space, so even “big” datasets may not provide enough information to
reliably populate all combinations of predictors. This sparsity makes it easy for complex models to
overfit noise, especially when the number of predictors is very large relative to the number of cases,
because the model can find spurious patterns that do not generalize to new data. One
methodological strategy to address this problem is to use regularization methods such as lasso or
elastic net regression, which shrink or set many coefficients to zero and effectively perform variable
selection; for example, a lasso model predicting customer lifetime value from thousands of
behavioral and demographic features can reduce the effective dimensionality while maintaining or
improving out‑of‑sample predictive performance
3. Domingos (2012), as summarized by Oswald and Putka, argues that “simpler algorithms based on
more data tend to beat cleverer algorithms based on less data.” Describe the logic behind this claim
and give a concrete example in a marketing or HR analytics context where this principle would
influence your model choice. [5 pts]
Domingos’ claim, as summarized by Oswald and Putka, that “simpler algorithms based on
more data tend to beat cleverer algorithms based on less data” captures the idea that, in many
practical settings, the key to good prediction is having large, representative samples rather than
using highly complex models on small or noisy datasets. The logic is that with more data, even
relatively simple models such as regularized linear or logistic regression can capture stable patterns
and avoid overfitting, whereas complex algorithms with many tuning parameters can easily fit
idiosyncrasies when data are limited. In a marketing analytics context, this principle might lead a
researcher to prefer a regularized logistic regression or random forest model trained on millions of
ad‑impression and purchase records to predict click‑through, rather than a highly intricate
3
