Punjabi University Students: Statistics Refresher for Data Analytics Internship Tasks

The statistics you need for a data analytics internship is a small, practical set: describing data (mean, median, spread), sampling, confidence intervals, hypothesis tests for A/B comparisons, and the discipline to separate correlation from causation. This article maps each concept to the task where it appears, works one A/B test by hand, and ends with a 15-question quiz with explanations so you can find your weak spots before the internship begins.

One-page statistics map: which concept shows up in which internship task

Interns are rarely asked to derive formulas. They are asked to clean a dataset, compare two groups, summarise a survey, or explain whether a change “worked”. The table below connects the task you will be handed to the statistical idea behind it and to the question you should ask yourself.

Describing data: mean, median, spread and when each misleads

Almost every internship task starts here. Know these four points cold:

• Mean is sensitive to extreme values. A handful of very large orders can drag it far above what a typical customer spends.
• Median is the middle value and resists outliers, so it is usually safer for incomes, order values, delivery times and anything with a long right tail.
• Standard deviation describes how spread out individual observations are. Standard error describes how uncertain your sample mean is. Interns often confuse the two.
• Percentiles and IQR (the range between the 25th and 75th percentiles) give a robust view of spread and a standard rule for flagging outliers: values beyond 1.5 times the IQR from the quartiles.

When you report a number, say which summary you used and why. A line such as “median delivery time was 3 days; the mean was 4.6 because of a few delayed shipments” shows judgment that a bare average does not.

Sampling for analysts: getting a subset you can trust

Real internship datasets are often too large, too messy, or only partly accessible, so you work with samples. The method of selection matters more than the size.

• Simple random sampling: every record has an equal chance. Good default when the population is fairly uniform.
• Stratified sampling: split the population into groups (for example, region or device type) and sample within each. Use it when a small group matters and a random draw might miss it.
• Systematic sampling: take every k-th record. Be careful if the data has a repeating pattern that lines up with k.
• Convenience sampling: whoever is easiest to reach. It is fast and often biased.

Two biases to watch for are self-selection (people choose whether to take part, as in an opt-in poll) and non-response (the people who do not reply differ from those who do). Increasing sample size does not fix either. A million biased responses are still biased.

A/B testing basics: hypotheses, p-values, errors and sample size

An A/B test compares a control (A) with a variant (B) on a metric, usually a conversion rate or an average. The logic runs in five steps:

1. State the null hypothesis: there is no real difference between A and B.
2. Choose a significance level in advance, commonly 0.05.
3. Decide the sample size and run length before starting, so you are not tempted to stop when the numbers look good.
4. Compute a test statistic and a p-value: the probability of seeing a difference at least this large if the null were true.
5. Report the effect size and confidence interval, not just “significant” or “not significant”.

A Type I error is a false positive: declaring a difference that is not real. A Type II error is a false negative: missing a real difference, usually because the sample was too small. Power is the chance of detecting a real effect of a given size; small effects need large samples.

Worked A/B test by hand: a two-proportion comparison with illustrative numbers

The figures below are invented for illustration and were calculated by hand, not taken from a real study and not run in software. Suppose a sign-up page is tested for one week.

Step 1: the observed difference. 13.5% minus 12.0% is 1.5 percentage points.

Step 2: the pooled rate under the null. Total sign-ups 1,020 out of 8,000 visitors gives 0.1275.

Step 3: the standard error of the difference. The square root of 0.1275 × 0.8725 × (1/4000 + 1/4000) is about 0.00746.

Step 4: the z statistic. 0.015 divided by 0.00746 is about 2.01. A two-sided p-value for z = 2.01 is roughly 0.044.

Step 5: a 95% confidence interval for the difference. Using the separate group rates, the standard error is about 0.00746, so the interval is 0.015 plus or minus about 0.0146, or roughly 0.04 to 2.96 percentage points.

How to write it up: “B converted at 13.5% against 12.0% for A. The difference is statistically significant at the 0.05 level (p ≈ 0.04), but the interval is wide: the true uplift could be anywhere from almost nothing to about three points. A longer test or a follow-up would give a firmer estimate.” That last sentence is what separates a careful analyst from one who just prints “significant”.

Confidence intervals, effect size and statistical versus practical significance

A 95% confidence interval is produced by a method that, repeated over many samples, captures the true value about 95% of the time. It is not a 95% probability that this particular interval contains the truth, though in practice analysts use it to show the plausible range of an effect. If an interval for a difference includes zero, the data are compatible with no effect.

With a very large sample, a tiny difference can be statistically significant yet commercially meaningless. Always pair the p-value with the size of the change and what it means in rupees, users or hours.

Correlation, causation and the traps that catch new analysts

• Confounding: ice cream sales and heatstroke both rise in summer; neither causes the other.
• Peeking: checking an A/B test daily and stopping the moment p dips below 0.05 inflates false positives.
• Multiple comparisons: testing 20 segments at 0.05 gives roughly a 64% chance of at least one false “win” even if nothing is going on (1 minus 0.95 to the power 20).
• Imbalanced classes: if 2% of transactions are fraud, a model that labels everything legitimate scores 98% accuracy and catches nothing.
• Survivorship bias: analysing only the customers who stayed tells you nothing about why others left.

A causal claim needs a randomised experiment or a very careful design. When you only have observational data, use words like “is associated with”.

Fourteen-day plan to rebuild statistics alongside an internship

The free two-week online internship from CareerSwami runs for 14 days, so a short, steady routine fits well. Use this plan with any analytics tasks you receive.

1. Days 1-2: Descriptive statistics. Summarise a dataset with mean, median, standard deviation and percentiles; plot a histogram and note the skew.
2. Days 3-4: Sampling. Draw simple random and stratified samples from the same data and compare how well each reproduces the full-data summary.
3. Days 5-6: Distributions and standard error. Understand why the sample mean varies less than individual values.
4. Days 7-8: Confidence intervals. Compute and interpret one for a mean and one for a proportion.
5. Days 9-10: A/B testing. Redo the worked example above with different numbers and write the summary paragraph.
6. Days 11-12: Pitfalls. Practise spotting peeking, confounding and multiple testing in sample case descriptions.
7. Days 13-14: Take the quiz below under time pressure, then write a one-page findings note in plain language.

Fifteen-question statistics quiz for analytics interns, with explanations

Answer each question before reading the explanation. Score yourself at the end.

1. Order values in a shop range mostly from ₹200 to ₹800, but a few bulk orders exceed ₹50,000. Which summary best represents a typical order?
   Answer: the median. The few huge orders pull the mean upward, while the median stays near the typical value.
2. You survey customers by posting a link in a WhatsApp group and analyse those who reply. What is the main problem?
   Answer: self-selection and non-response bias. Only active, interested members respond, so results may not describe all customers.
3. A test returns p = 0.03. What does this mean?
   Answer: if there were truly no difference, data this extreme or more would occur about 3% of the time. It is not the probability that the null hypothesis is true.
4. If you quadruple the sample size, what happens to the standard error of the mean?
   Answer: it halves, because it scales with 1 divided by the square root of n.
5. A 95% confidence interval for a conversion uplift is 0.5 to 3.0 points. What is a sound reading?
   Answer: the data are consistent with an uplift anywhere in that range, and zero is not inside it, so the result is significant at the 5% level. The size of the real uplift remains uncertain.
6. What is a Type I error?
   Answer: rejecting a null hypothesis that is actually true, a false positive.
7. You test 20 independent segments at the 0.05 level and none truly differs. Roughly how likely is at least one “significant” result?
   Answer: about 64%. This is why multiple comparisons need correction or a pre-stated primary metric.
8. Users who open the app more often also spend more. Can you say opening the app causes spending?
   Answer: no. Motivation or product fit could drive both. Causation needs an experiment or a careful causal design.
9. A colleague stops an A/B test the first day p falls below 0.05. Why is this risky?
   Answer: repeated peeking inflates the false-positive rate. Fix sample size or stopping rules in advance, or use a method built for sequential looks.
10. You find a row with an age of 214. What should you do first?
    Answer: investigate. It is probably an entry error, but check the source and document whether you correct, exclude or flag it. Do not delete outliers silently.
11. What is the difference between standard deviation and standard error?
    Answer: standard deviation measures spread among individual observations; standard error measures how much a sample statistic, such as the mean, would vary across repeated samples.
12. When is stratified sampling preferable to simple random sampling?
    Answer: when important subgroups are small or uneven in size and you need each represented, for example a minority region or a rare product category.
13. Which test suits comparing conversion rates (yes/no outcomes) between two page versions?
    Answer: a two-proportion z-test (or an equivalent chi-square test on the 2×2 table).
14. With millions of visitors, a 0.05-point uplift is highly significant. Should the company act automatically?
    Answer: not necessarily. Statistical significance is not practical significance. Compare the gain with the cost of the change and the business value.
15. A fraud model reports 98% accuracy, but only 2% of cases are fraud. Why is that not impressive?
    Answer: predicting “not fraud” for everything also gives 98%. Check recall and precision on the fraud class.

Scoring: 13-15 means you are ready to focus on tools and communication. 9-12 means revisit the sections on testing and sampling. Below 9, spend the first week on descriptive statistics and standard error before attempting A/B work.

Punjabi University, Patiala: placement-cell rules that shape a data analytics internship

Punjabi University in Patiala runs its placement work through the Career and Counselling cum Placement Cell, and the cell publishes a placement information sheet that sets out how students are expected to behave. For anyone taking on data analytics tasks, the practical value lies in knowing which of those rules touch an internship and which do not.

The sheet says that if the Placement Cell arranges internship training for a student who has given consent, joining becomes binding, and skipping it can lead to exclusion from later placement activities unless the authorities accept the reason. In practice, a student should only say yes to a cell-arranged internship once the project scope, dates and exam calendar have been checked, because a casual yes is treated as a commitment.

The same document asks students not to join pool or joint campus drives outside the university before the first round of placements is complete without prior written permission from the Placement Cell. A student who finds an analytics internship through an outside platform or company should therefore ask the cell whether that offer falls under this clause before accepting anything.

Communication runs through department faculty co-ordinators, and the sheet notes that information is posted on the university website or sent by email. The cell’s notice board relays internship offers from outside companies, each carrying its own eligibility conditions, and many of those posts carry older dates, so check the date on every notice before relying on it. The policy also warns that misrepresenting facts during placement activities can lead to exclusion and disciplinary action, so any statistics project or dashboard listed on a resume should be work the student actually did.

These pages describe placement conduct rather than a statistics syllabus, so course-level rules for internships or projects have to come from the student’s own department.

Checks to make before starting an analytics internship

• Ask the faculty co-ordinator of the department whether the internship needs written permission from the Placement Cell.
• Confirm in writing whether the internship was arranged by the cell, since that makes joining binding.
• Check the date and eligibility conditions on any notice-board listing before applying.
• Ask the department whether the internship counts toward any project or practical component of the course.
• Keep a dated record of tasks, datasets used and outputs so that resume claims can be backed up.

Turning refresher knowledge into internship output: what to submit

Whatever tasks you are given, produce a short findings note that shows statistical thinking:

• One sentence on the question and the metric.
• How the data was sampled or collected, and known biases.
• The summary statistic you chose and why.
• The result with an effect size and a confidence interval.
• Limitations, and what you would test next.

Keep your analysis steps reproducible: record filters, exclusions and the exact formulas. Reviewers trust a modest result with clear reasoning more than an impressive one that cannot be traced.

Frequently Asked Questions

How much statistics do I need for a data analytics internship?

Enough to summarise data honestly, understand sampling, run and interpret a basic A/B comparison, and explain uncertainty in plain words. Advanced modelling is usually not required at the start.

Is the p-value the probability that my result is wrong?

No. It is the probability of observing data at least this extreme if the null hypothesis were true. It does not give the probability that either hypothesis is correct.

What sample size do I need for an A/B test as a Punjabi University student?

It depends on the baseline rate, the smallest uplift worth detecting, the significance level and the desired power. Smaller effects need much larger samples. Use a sample-size calculator before the test starts and record the assumptions.

Should I remove outliers from my dataset?

Not automatically. Check whether they are errors or genuine rare events. Report what you did, and consider showing results both with and without them.

Can I say a change caused an increase if the dashboard shows a rise after launch?

Be cautious. A before-and-after comparison can be affected by seasonality, promotions or other changes. A randomised A/B test gives much stronger evidence of cause.

Which Python or Excel functions help with these statistics tasks?

Excel offers AVERAGE, MEDIAN, STDEV.S and CONFIDENCE.T. In Python, pandas handles summaries and sampling, while SciPy and statsmodels provide tests and intervals. Understand the logic first so you can check any tool’s output.

Does the 4-credit internship in the Punjabi University NEP templates cover any online internship?

The template text only states the credit requirement; it does not say which providers or formats qualify. Ask your department in writing whether online internships are accepted and what documents they need.

How can the free two-week internship help me practise statistics for analytics?

Its 14-day length suits a focused routine: apply one concept every day or two to the tasks you receive, then finish with a short findings note. Confirm separately how any internship fits your university requirements.

Related Data Analytics guides

• G.D. Goenka University Students: Asking Good Questions When Your Analytics Brief Is Vague
• Graphic Era Hill University Students: Common Mistakes in Data Analytics Internships and Fixes
• Gujarat Biotechnology University Students: Analysing Biological Datasets in a Short Analytics Internship
• All Data Analytics internship guides