Gains:
- Ability to report empirical findings to the target audience (academic, policy, management) with the support of artificial intelligence, in an honest and unambiguous language
- Ability to fully write conclusions, limitations and ethical statements (data confidentiality, conflict of interest, disclosure of artificial intelligence use) and avoid misleading presentation
- Ability of artificial intelligence to recognize drafts of reports that produce exaggerated, one-sided or causal language and maintain that responsibility for the final text and claims rests with the researcher.
Module Exam
1. A researcher asks 'What is the correlation coefficient between unemployment and inflation in Türkiye between 2015 and 2020?' without giving any data to artificial intelligence. he asks and writes the number 0.62 directly into his article. What is the fundamental mistake in this approach?
- A) Expecting concrete statistics from artificial intelligence without giving verified data; ✔ Using a number that may have been made up without verifying it
- B) The correlation coefficient should never be used in an article.
- C) Correlation cannot be calculated between unemployment and inflation
- D) Artificial intelligence can only access data after 2020
Explanation: The AI language model cannot produce statistics without accessing the dataset; The number he gives is most likely a hallucination (fabricated). Coefficients, ratios and test results should be calculated from the researcher's own verified data, not taken from the model's memory.
2. What does it mean for missing data to be 'not missing at random' (MNAR) and why is it important?
- A) The deficiency is due to the unobserved value itself; That's why simple assignment can create bias ✔
- B) Data disappears completely randomly and does not create any bias.
- C) Missing data should always be resolved by deleting the row.
- D) Missing data does not affect the analysis in any way.
Explanation: In the case of MNAR (Missing Not At Random), the missingness itself depends on the magnitude of the unobserved value (e.g. high earners do not report their income). In this case, simple deletion or average assignment gives biased results; The mechanism of the deficiency must be modelled. The assignment method suggested by artificial intelligence should not be applied blindly without knowing this mechanism.
3. An analyst has the AI make a bar chart, and the AI starts the y-axis at 40 instead of zero. The actual difference of 2% between the two groups looks huge on the graph. What is the right approach?
- A) Starting the axis from scratch or changing the chart type; ✔ to show the true size of the difference without misleading
- B) Using the chart as is because AI makes the best choice
- C) Starting the axis at 45 to make the difference even larger
- D) Never use visuals instead of bar charts
Explanation: In a bar chart, the dashed axis (the y-axis that does not start from zero) misleads by exaggerating small differences. In honest visualization, the axis of bar charts should start from zero or the difference should be consciously stated clearly. It is the analyst's responsibility to verify the image and correct it if necessary.
4. How is the fact that a coefficient is 'significant' (p<0.05) in linear regression often misrepresented in the interpretation of artificial intelligence?
- A) Exaggerated presentation of significance as the effect is large and important or causality is established ✔
- B) Significance always indicates that the effect is small
- C) p<0.05 proves that the coefficient is absolutely correct
- D) Significant coefficients should never be reported.
Explanation: Statistical significance relates to the strength of evidence that the effect is different from zero; It does not mean that the effect is large, important or causal. AI often presents the word 'significant' as 'significant/strong impact'. The researcher should also evaluate the effect size, confidence interval, and context.
5. What does the term 'p-hacking' refer to and why can AI make it easier?
- A) Trying multiple analyzes until it makes sense; AI does this very quickly, increasing the risk ✔
- B) Analyzing data once and with a predetermined plan
- C) expanding the sample to increase the p-value
- D) Reporting only descriptive statistics
Explanation: p-hacking is trying different variables, subsamples, transformations, or models until a meaningful result emerges; this produces spurious findings based on chance. Since AI can try dozens of model variants in seconds, it can speed up p-hacking without an honest plan. Defense; preregistration, fixed analysis plan and multiple comparison correction.
6. Why might finding a high R-squared regression between two non-stationary (unit root) time series be misleading?
- A) Spurious regression may occur due to common trend; ✔ High R-squared output without real relationship
- B) A high R-squared always proves a strong causal relationship.
- C) Non-stationary series cannot be entered into regression at all.
- D) R-squared cannot be calculated in time series
Explanation: If there is no common cointegration relationship between non-stationary series, spurious regression may produce high R-squared and significant coefficient only due to common trend; whereas there is no real relationship. Therefore, stationarity and unit root tests should be done first, and if necessary, differences should be taken or cointegration should be tested.
7. What basic assumption is based on the validity of the Difference-in-Differences design?
- A) Parallel trends assumption: groups would follow the same direction if there were no intervention ✔
- B) Treatment and control groups are exactly the same at the beginning
- C) Normal distribution of the sample
- D) Variables do not contain any missing data
Explanation: DiD assumes that in the absence of the intervention, the outcome variable for the treatment and control groups would have coursed parallel over time (parallel trends assumption). If this assumption is not met, the estimate is biased. AI can produce the DiD code, but it is the researcher's job to defend and test parallel trends.
8. Which of the following is a mandatory practice for a reproducible analysis?
- A) Fixing the seed in random steps, recording package versions and code ✔
- B) Manually copy the results to the spreadsheet and delete the code
- C) It is normal to see different results in each run.
- D) Keeping only final graphics, not sharing data and code
Description: Reproducibility; The same result can be obtained again with the same data and code. Fixing the seed in random steps, saving package versions, and executing the code within the document (literate programming) are the cornerstones of this. Copying results manually or click-based, non-logging steps impairs reproducibility.
9. What does heteroskedasticity distort a linear regression and what is its common solution?
- A) It distorts standard errors; solution is robust standard errors or appropriate transformation ✔
- B) It completely invalidates the coefficient estimates and has no solution
- C) Always reduces R-squared to zero
- D) Occurs only if the sample is very small and can be ignored
Explanation: Heteroscedasticity leaves coefficient estimates unbiased but miscalculates standard errors; This makes p-values and confidence intervals unreliable. A common solution is to use robust/HC standard errors or appropriate conversion. It must be verified that the solution suggested by the AI actually fixes the problem rather than hiding it.
10. A researcher uploads microdata containing patient-level identification information and income into a public AI chat tool and says 'do regression'. What is the main problem with this behavior?
- A) Uploading personal/licensed data to an external tool constitutes a breach of confidentiality and contract ✔
- B) Regression cannot be done by artificial intelligence at all
- C) Micro data is not suitable for regression at all
- D) AI always miscalculates the regression
Explanation: Personal/special data such as identity and income are protected under KVKK/GDPR and most data usage agreements; Uploading them to an external device that can store data is a violation. The right way; Anonymizing data, using local/institutional secure tools, or simply requesting code from artificial intelligence and running the code in its own environment.
11. What is observed when multicollinearity is high?
- A) Coefficients become unstable and standard errors become inflated; Individual coefficients may turn out to be meaningless ✔
- B) R-squared always decreases to zero
- C) Coefficient estimates become more precise all the time
- D) The scale of the dependent variable changes
Explanation: In high multicollinearity, independent variables are strongly correlated with each other; Coefficient estimates become unstable, standard errors become inflated, and individual coefficients may turn out to be insignificant, while the overall model may be significant. It is diagnosed by criteria such as VIF. Artificial intelligence may suggest variable elimination, but it is up to the researcher to decide which variable should remain theoretical.
12. What is statistical power and what is the risk of a low-power study?
- A) Possibility of detecting the existing effect; low power misses true effects and makes findings unreliable ✔
- B) probability of reducing the p-value; lower power always gives more reliable results
- C) A constant value independent of the sample size
- D) A concept that is valid only when artificial intelligence is used
Explanation: Power is the probability of detecting an effect that actually exists (1 minus type II error). Low-powered studies may miss true effects; Additionally, the few significant results may carry an exaggerated effect size ('winner's curse') and the false positive rate increases. Therefore, power/sample calculation is important before analysis.
13. Why is it ethically necessary to disclose the use of artificial intelligence in an article?
- A) For transparency and accountability; ✔ readers and referees can evaluate the process and the responsibility remains with the author
- B) The use of artificial intelligence automatically invalidates the results
- C) Requested by magazines for advertising purposes only
- D) Transferring the copyright of the explanation to artificial intelligence
Disclosure: Transparency is necessary so that the reader and reviewers can evaluate how the results were produced; Many journals and institutions now require disclosure of AI use. Additionally, since artificial intelligence can produce errors/hallucinations, it is a matter of honesty to state that the responsibility remains with the author and which steps were audited.
14. What does the 'exclusion restriction' require in the Instrument Variable (IV) method?
- A) The tool affects the result only through the internal variable, there is no other direct way ✔
- B) The instrument has no relationship with the outcome variable.
- C) The instrument is unrelated to the endogenous variable.
- D) Means is always a binary (0/1) variable
Explanation: The exclusion constraint requires that the instrument affects the outcome variable only through the endogenous explanatory variable and not through any other direct means or unobserved factors. This assumption cannot be fully tested from data; It is defended on theoretical and institutional grounds. The AI may write the IV code, but it is the researcher's job to defend the validity of the tool.
15. What does the 'Garden of forking paths' problem mean in flexible analyzes without pre-registration?
- A) Too many reasonable analysis choices made based on the data increase the false positive rate, even unintentionally ✔
- B) Analysis methods must be single and mandatory
- C) A problem that only occurs when intentional cheating occurs.
- D) A situation that disappears spontaneously as the sample grows
Explanation: After viewing the data, the researcher can make many reasonable choices about which variable, subgroup, transformation, or threshold to use. Even if there is no single 'intentional p-hacking', this flexibility increases the false positive rate because the significant one is effectively chosen from a large number of analyses. Pre-registration and a fixed analysis plan limit this flexibility.