Central Limit Theorem (CLT) and Normal Distribution Questions

Understand the CLT: when you take multiple random samples and calculate their means, those sample means are normally distributed (bell-shaped) even if the underlying data isn't. Know that normal distribution is parameterized by mean and standard deviation. Appreciate why this matters: it allows you to estimate population characteristics from samples and construct confidence intervals.

MediumTechnical

28 practiced

Describe how skewness in the underlying observation distribution affects the sampling distribution of the sample mean as sample size increases. Include an explanation of the role of sample size in mitigating skewness and a short plan for a small empirical simulation to show the effect.

EasyTechnical

30 practiced

When should a data scientist use the t-distribution instead of the normal distribution to construct a confidence interval for a mean? Describe the rule-of-thumb regarding sample size and unknown population variance, and explain the practical consequences of using a t-based CI versus a z-based CI on interval width for small samples.

MediumTechnical

36 practiced

You are forecasting expected daily revenue but the raw revenue per user is extremely heavy-tailed with occasional huge values. Discuss whether the CLT justifies using sample mean-based CIs and propose robust alternatives you would use as a data scientist to provide reliable uncertainty quantification.

EasyTechnical

30 practiced

You collected 30 independent sample means, where each sample mean was calculated from n = 5 raw observations. Discuss whether the CLT justifies treating the 30 sample means as approximately normally distributed. Provide considerations about underlying distribution shape, the role of n versus number of sample-means, and practical alternatives if normality is doubtful.

HardTechnical

35 practiced

Survey data uses stratified sampling with weights. Explain how the CLT applies to weighted sample means and describe methods to estimate the variance of a weighted mean (e.g., Taylor linearization, replicate weights such as bootstrap or jackknife). Briefly outline how you would implement variance estimation for a weighted mean in Python or R.

Sample Answer

The central limit theorem (CLT) extends to weighted sample means: under regularity (independent or appropriately weakly dependent sampling within strata, finite variance, no single unit dominates weight), the weighted meanμ̂ = (∑ wi yi) / (∑ wi)is asymptotically normal around the true population mean with variance driven by the design and weights. Intuitively, weights re-scale contributions and change the effective sample size (neff ≈ (∑wi)^2 / ∑wi^2), which appears in the variance.

Variance estimation methods

1. Taylor (linearization)- Treat μ̂ as a smooth function of totals. Linearize μ̂ by first-order Taylor expansion to get an approximate influence function: ui ≈ (wi/∑wj)(yi − μ̂).- Then apply standard design-based variance formulas (strata/PSU variance) to the linearized variable ui.- Works well for large samples and functions that are smooth.

2. Replicate-weight methods- Create R replicate weight vectors w(r) that reflect the sampling design (bootstrap resamples within PSUs/strata, jackknife delete-one-PSU, BRR for two-PSU strata).- Compute μ̂(r) for each replicate; estimate variance as variance of replicates times appropriate scale factor.- Flexible and often more accurate for complex designs and nonlinear estimators.

Implementation outline

R (survey package):

library(survey)
# data: y, weight, strata, psu
design <- svydesign(ids=~psu, strata=~strata, weights=~weight, data=df, nest=TRUE)
# Taylor linearization variance
svymean(~y, design)
# Replicate weights (bootstrap)
rdes <- as.svrepdesign(design, type="bootstrap", replicates=1000)
svymean(~y, rdes)

Python (statsmodels / custom):

python

import numpy as np
# weighted mean
mu = (w * y).sum() / w.sum()
# approximate linearized residuals
u = (w / w.sum()) * (y - mu)
# if you have PSUs grouped by 'psu' and strata: compute between-PSU variance
# replicate bootstrap example (simplified)
R = 1000
reps = np.empty(R)
for r in range(R):
    idx = resample_psu_indices(...)  # resample PSUs respecting strata
    wr = build_replicate_weights(idx, original_w)
    reps[r] = (wr * y).sum() / wr.sum()
var_rep = reps.var(ddof=1) * scale_factor

Practical notes: check weight trimming, finite-population corrections in small-pop designs, and use existing libraries (survey in R, srvyr wrapper or statsmodels.robust for some features) when possible.

Unlock Full Question Bank

Get access to hundreds of Central Limit Theorem (CLT) and Normal Distribution interview questions and detailed answers.

Join thousands of developers preparing for their dream job.