Term 2 · Module 2 of 4

Sampling and Sampling Distributions

Business Statistics for Entrepreneurs II

Practise this module in Claude, ChatGPT or Gemini

Set it up free
Mind map

Sampling: From a Population to Evidence

Sampling selects a subset of observations from a population so that its data can support inferences about population characteristics. A sample statistic (such as a sample mean) describes the selected observations; a population parameter (such as the population mean) describes the entire population and is often unknown. A statistic is an estimate, not a guarantee of equality: a sample of 400 voters with 160 supporting a candidate gives a sample proportion of 160/400=0.40160/400=0.40, but the population proportion may differ. Likewise, an observed mean annual earning of ₹36,500 among 120 temporary workers is not automatically the mean for all workers of interest.

Reliable inference depends on how observations enter the sample, not merely on the arithmetic performed afterward. A useful sampling plan proceeds through four decisions:

  1. Define the target population precisely, including relevant boundaries such as age, gender, and work type when studying migrant temporary workers.
  2. Choose a sampling frame: the source or procedure used to identify eligible elements. Voter registration data may supply a frame for registered voters; a ready-made list may not exist for temporary migrant workers.
  3. Set a sample size that balances precision and confidence needs against the time and cost of collecting data.
  4. Choose a sampling method for selecting cases from the frame.
  1. 1Define target population
  2. 2Choose sampling frame
  3. 3Set sample size
  4. 4Select sampling method
  5. 5Collect sample data
  6. 6Estimate population parameters

Key takeaways

  • A sample is used to learn about a population; sample results need not equal population values.
  • Ambiguous population definitions and incomplete frames can undermine inference before analysis begins.
  • Larger samples can improve precision, but require more time and money.

Sampling Methods: Probability and Non-Probability Designs

Probability sampling selects observations according to a probability mechanism. It supports probability-based assessment of sampling variability, provided the design and assumptions are appropriate. Non-probability sampling selects units without known selection probabilities, limiting how securely results can be generalized.

Probability sampling

Method How selection works Best-fit logic and cautions
Simple random sampling Every population member has equal selection probability; label cases and select using random numbers (for example, spreadsheet RAND() or RANDBETWEEN()). Suitable when the population is reasonably homogeneous.
Stratified sampling Divide the population into mutually exclusive, collectively exhaustive strata, then randomly sample within each. Often allocate sample counts in proportion to stratum sizes. Use when the population is heterogeneous overall but groups are internally more homogeneous. Helps ensure representation of groups such as gender or age bands; relatively small samples can estimate stratum characteristics precisely when strata are homogeneous.
Cluster sampling Divide the population into non-overlapping clusters, randomly select clusters, then include all elements in chosen clusters. Area sampling uses districts, municipalities, or wards. Clusters should each resemble a small-scale population, with diverse members within each. It can save time and cost by concentrating fieldwork, though it generally needs a larger sample than simple random or stratified sampling.

Strata and clusters solve different problems: stratification seeks similarity within each group and samples from every group; clustering seeks representative, internally diverse groups and samples only selected groups.

Non-probability sampling

Method Selection basis Main limitation
Systematic sampling Randomly choose a start in the first interval, then take every kkth listed element. For 400 from 4 crore, the interval is 4,00,00,000/400=100,0004{,}00{,}00{,}000/400=100{,}000. Easier to execute than locating many random list entries; it is often treated as having some simple-random-sampling properties because of the random start. The sequence of the list still matters.
Convenience sampling Include readily accessible units, such as workers encountered at a plant or motorists arriving at a checkpoint. Selection probabilities are unknown; representativeness and robust population inference cannot be assessed reliably.
Judgment sampling A knowledgeable selector chooses cases believed to represent the population, such as a small set of residents. Quality depends on the selector’s judgment.
Voluntary sampling People opt in, as in customer surveys or clinical-trial recruitment. Self-selection can bias results; people with strong or bad experiences may be more likely to respond, and volunteers need not represent the population.

Key takeaways

  • Stratify when every important subgroup should be represented; cluster when fieldwork can be concentrated in representative groups.
  • Equal-probability random selection is the core of simple random sampling.
  • Convenience, judgment, and voluntary samples can be useful for access or exploration, but require caution when generalized.
  • Systematic selection uses a random start and a fixed interval; inspect the frame’s ordering for patterns.

Analysis of Sample Data: Statistics and Point Estimates

The usual targets include the population mean μμ, population standard deviation σσ, and population proportion pp. Their sample counterparts are the sample mean xˉ\bar{x}, sample standard deviation ss, and sample proportion pˉ\bar{p}.

For a customer satisfaction survey with 36 responses, the transcript reports a mean near 57.31, median 59, mode 61, sample standard deviation 6.38, and 17 scores of at least 60. Thus:

pˉ=xn=1736≈0.47\bar{p}=\frac{x}{n}=\frac{17}{36}\approx0.47

Each statistic is a point estimator of its corresponding parameter, and its observed numerical value is a point estimate. For example, xˉ\bar{x} estimates μ\mu and pˉ\bar{p} estimates pp. A different sample from the same population would generally produce different values. A point estimate alone does not show how close it is to the unknown parameter; sampling distributions provide a way to quantify that uncertainty.

Key takeaways

  • xˉ\bar{x}, ss, and pˉ\bar{p} estimate μ\mu, σ\sigma, and pp, respectively.
  • Point estimates vary from sample to sample even when the population is unchanged.
  • The sample proportion is count of qualifying observations divided by sample size.

Sampling Distribution: The Distribution Across Possible Samples

A sampling distribution is the probability distribution of a statistic over all possible samples of a specified design and size. Imagine repeatedly drawing simple random samples of 36 customers and calculating xˉ\bar{x} each time. The resulting means vary; a histogram from 500 such samples would approximate, rather than equal, the sampling distribution of xˉ\bar{x}. In actual studies, usually only one sample is observed. Repeated sampling is a conceptual tool that explains the randomness of the statistic.

For independent, identically distributed observations with population mean μ\mu and standard deviation σ\sigma (and a population large relative to the sample):

E(xˉ)=μ,SE⁡(xˉ)=σxˉ=σnE(\bar{x})=\mu,\qquad \operatorname{SE}(\bar{x})=\sigma_{\bar{x}}=\frac{\sigma}{\sqrt{n}}

The standard error (SE) is the standard deviation of a statistic’s sampling distribution. Since E(xˉ)=μE(\bar{x})=\mu, the sample mean is an unbiased estimator of the population mean: its sampling distribution is centered at the parameter. An individual xˉ\bar{x} need not equal μ\mu; its likely distance from μ\mu depends on the standard error. Increasing nn reduces this spread at the rate 1/n1/\sqrt n. For example, if σ=6\sigma=6 and n=36n=36, then SE⁡(xˉ)=6/6=1\operatorname{SE}(\bar{x})=6/6=1.

If the population itself is normally distributed, xˉ\bar{x} is normally distributed for any sample size, with mean μ\mu and standard error σ/n\sigma/\sqrt n. This exact normality is distinct from the large-sample approximation supplied by the central limit theorem.

Key takeaways

  • A statistic is random because the selected sample is random; its sampling distribution describes that variation.
  • The sample mean is unbiased: its expected value equals μ\mu.
  • Larger nn narrows the sampling distribution; it does not move its center.
  • For a normal population, the sample mean is exactly normal; otherwise, large-sample normality is an approximation.

Central Limit Theorem and Sample-Mean Applications

The central limit theorem (CLT) says that for sufficiently large random samples, the sampling distribution of xˉ\bar{x} can be approximated by a normal distribution, even when the population distribution is not normal. Its mean is μ\mu and, for a population large relative to nn, its standard error is σ/n\sigma/\sqrt n.

The rule of thumb n≥30n\ge30 is empirically useful in many applications, not a universal guarantee. Highly skewed populations may need much larger samples. The CLT concerns the distribution of sample means, not the population: collecting a large sample does not make the population itself normal.

When the approximation applies, standardize a sample mean using

Z=xˉ−μσ/nZ=\frac{\bar{x}-\mu}{\sigma/\sqrt n}

Then translate the desired event into a standard-normal probability. Examples reported in the lecture:

Scenario Sampling calculation Result
Unemployment duration: σ=4\sigma=4, n=50n=50 SE=4/50≈0.57SE=4/\sqrt{50}\approx0.57 weeks. Within 1 week: P(−1.77≤Z≤1.77)≈0.923P(-1.77\le Z\le1.77)\approx0.923; within 0.5 week: P(−0.88≤Z≤0.88)≈0.62P(-0.88\le Z\le0.88)\approx0.62. A representative sample can support a detailed study even when population mean and SD are already known; the known mean can serve as a check on whether the sample reflects the population.
Annual rainfall: σ=4\sigma=4 cm; Jaisalmer n=30n=30, Jhalawar n=44n=44 SEs are 4/30≈0.734/\sqrt{30}\approx0.73 and 4/44≈0.604/\sqrt{44}\approx0.60. Within 1 cm gives $P( Z
Tax preparation fee: σ=Rs. 50\sigma=\text{Rs. }50; n=40n=40 then 7575 SEs are 50/40≈Rs. 7.9150/\sqrt{40}\approx\text{Rs. }7.91 and 50/75≈Rs. 5.850/\sqrt{75}\approx\text{Rs. }5.8. Within ₹10 gives $P( Z
Disease duration: μ=8\mu=8, σ=4\sigma=4, n=30n=30 SE=4/30≈0.73SE=4/\sqrt{30}\approx0.73. Approximate probabilities: P(xˉ<7)≈0.08P(\bar{x}<7)\approx0.08, P(xˉ>7)≈0.92P(\bar{x}>7)\approx0.92, and P(7≤xˉ≤9)≈0.83P(7\le\bar{x}\le9)\approx0.83. The population is described as likely right-skewed, yet the sample mean is treated as approximately normal under the CLT.

For the satisfaction-score example, taking the prior estimate σ=6\sigma=6 and n=36n=36 gives SE=1SE=1. The lecture’s standard-normal calculation gives about a 95% probability that a sample mean falls within 2 units of μ\mu. At n=100n=100, SE=0.6SE=0.6 and P(∣Z∣≤2/0.6)>0.999P(|Z|\le2/0.6)>0.999. The larger sample leaves the center at μ\mu but concentrates possible means more tightly around it.

Key takeaways

  • Apply the CLT to the statistic xˉ\bar{x}, not to the underlying population values.
  • Standardize by the standard error, not by the population standard deviation alone.
  • The n≥30n\ge30 guideline is approximate; severe skew can call for a larger sample.
  • Larger samples increase the chance that xˉ\bar{x} falls within a fixed distance of μ\mu.

Sampling Distribution of a Proportion

For a binary characteristic, let XX count the sample’s successes and define the sample proportion as

pˉ=Xn\bar{p}=\frac{X}{n}

For a large population and a simple random sample, XX is binomial with mean npnp and variance np(1−p)np(1-p). Dividing by nn gives

E(pˉ)=p,SE⁡(pˉ)=p(1−p)nE(\bar{p})=p,\qquad \operatorname{SE}(\bar{p})=\sqrt{\frac{p(1-p)}{n}}

Thus pˉ\bar{p} is unbiased. When both np≥5np\ge5 and n(1−p)≥5n(1-p)\ge5, the binomial count can be approximated by a normal distribution, so pˉ\bar{p} is approximately normal with mean pp and the SE above. For standardization, use Z=(pˉ−p)/SE⁡(pˉ)Z=(\bar{p}-p)/\operatorname{SE}(\bar{p}).

Proportion examples

Scenario Calculation and result
Households spending over ₹2,500 monthly on snacks: p=0.17p=0.17 At n=80n=80, SE=0.17(0.83)/80≈0.042SE=\sqrt{0.17(0.83)/80}\approx0.042; probability within 0.050.05 (from 0.120.12 to 0.220.22) is about 0.7890.789. At n=160n=160, SE≈0.028SE\approx0.028 and probability rises to 0.9260.926.
Book adoption: p=0.25p=0.25, reported SE=0.0625SE=0.0625 Solve 0.0625=0.25(0.75)/n0.0625=\sqrt{0.25(0.75)/n} to get n=48n=48. The lecture checks both normal-approximation counts exceed 5. The approximate chance of a sample adoption proportion above 0.300.30 is 0.2120.212.
Paneer ghevar preference: p=0.75p=0.75, n=200n=200 SE=0.75(0.25)/200≈0.03SE=\sqrt{0.75(0.25)/200}\approx0.03. Approximate chance pˉ>0.80\bar{p}>0.80 is 0.0480.048; chance pˉ<0.60\bar{p}<0.60 is described as close to zero; the central 95% range is about 0.690.69 to 0.810.81.
Doctors believing patients receive unnecessary care: p=0.42p=0.42, n=150n=150 SE=0.42(0.58)/150≈0.04SE=\sqrt{0.42(0.58)/150}\approx0.04. The lecture gives P(0.39≤pˉ≤0.45)≈0.547P(0.39\le\bar{p}\le0.45)\approx0.547. It reports a tail probability of 0.0230.023 for a sample proportion above a threshold, but its stated threshold conflicts (0.40 in the question versus 0.50 in the calculation); see uncertainty note in the final report.

For the customer survey with n=36n=36, 17 of 36 scores at least 60 gives pˉ≈0.47\bar{p}\approx0.47. Under the illustrative assumption p=0.5p=0.5, SE=0.25/36≈0.0833SE=\sqrt{0.25/36}\approx0.0833. The approximate probability of a sample proportion between 0.450.45 and 0.550.55 is reported as 0.4520.452. At n=100n=100, SE=0.05SE=0.05 and the probability becomes 0.6830.683.

Key takeaways

  • The proportion’s center is pp; its standard error depends on both pp and nn.
  • Check npnp and n(1−p)n(1-p) before using the normal approximation.
  • Increasing nn reduces proportion SE and raises the chance of falling within a fixed margin.
  • Keep the population proportion and observed sample proportion separate.

Sample Variance and Standard Deviation

The sample variance measures squared spread around the sample mean. Because the deviations from xˉ\bar{x} must sum to zero, once n−1n-1 observations are known the last is constrained. This lost degree of freedom (DOF) motivates dividing the sum of squared deviations by n−1n-1:

s2=1n−1∑i=1n(xi−xˉ)2s^2=\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{x})^2

Unlike the sample mean, sample variance has no general CLT-style result here that supplies an approximate distribution for arbitrary populations. If independent observations come from a normal population with variance σ2\sigma^2, then

(n−1)s2σ2∼χn−12\frac{(n-1)s^2}{\sigma^2}\sim\chi^2_{n-1}

The chi-square distribution takes only positive values, is generally asymmetric, and becomes more bell-shaped as its degrees of freedom increase. This result requires population normality; conclusions for non-normal populations should be treated cautiously.

Using the chi-square relationship

To find a probability about ss or s2s^2, translate the event into one about χn−12\chi^2_{n-1} using (n−1)s2/σ2(n-1)s^2/\sigma^2, then read the cumulative, tail, or interval probability. Examples in the lecture:

Setting (normal-population assumption) Conversion and reported result
Unemployment: n=50n=50, σ=4\sigma=4 P(s2>20)=P(χ492>49⋅20/16=61.25)≈0.113P(s^2>20)=P(\chi^2_{49}>49\cdot20/16=61.25)\approx0.113. P(s<3)=P(χ492<49⋅9/16=27.56)≈0.006P(s<3)=P(\chi^2_{49}<49\cdot9/16=27.56)\approx0.006.
Tax fees: n=40n=40, σ=50\sigma=50 P(s>60)=P(χ392>39⋅602/502=56.16)≈0.037P(s>60)=P(\chi^2_{39}>39\cdot60^2/50^2=56.16)\approx0.037. P(s<55)=P(χ392<39⋅552/502=47.19)≈0.827P(s<55)=P(\chi^2_{39}<39\cdot55^2/50^2=47.19)\approx0.827, below the staff’s claimed 90%.
Disease duration: n=30n=30, σ=4\sigma=4 P(s<3)=P(χ292<29⋅9/16=16.31)≈0.028P(s<3)=P(\chi^2_{29}<29\cdot9/16=16.31)\approx0.028. P(3.5<s<4.5)=P(22.20<χ292<36.70)≈0.658P(3.5<s<4.5)=P(22.20<\chi^2_{29}<36.70)\approx0.658. The population is described as skewed, so the normality assumption is not credible and these answers warrant caution.

For a satisfaction sample, n=36n=36, s=6.38s=6.38, and s2=40.73s^2=40.73 are reported. Under the normal-population assumption and using 35 degrees of freedom, the lecture’s one-sided chi-square cutoffs yield illustrative bounds: σ2≳30.92\sigma^2\gtrsim30.92 from the 90th percentile 46.0646.06, and σ2≲57.44\sigma^2\lesssim57.44 from the lower-tail cutoff 24.824.8. Taking square roots gives approximately 5.56<σ<7.65.56<\sigma<7.6 under the respective one-sided 90% statements. This should not be mistaken for one two-sided 90% interval.

Key takeaways

  • Center deviations on xˉ\bar{x} and divide by n−1n-1 because one degree of freedom is constrained.
  • The chi-square result for sample variance depends on a normally distributed population.
  • Convert variance or SD thresholds to chi-square thresholds before computing probabilities.
  • For a visibly skewed population, variance conclusions from this model need caution; a larger sample may help, but the stated chi-square result is not guaranteed by the CLT.
Revise before your examAll 1 exam tip from this module in one list
  1. Keep the statistic and parameter distinct: xˉ\bar{x} is calculated from a sample; μ\mu describes the population.

Test yourself on this module

Questions from these notes, one at a time, with feedback. Your weak topics are saved.

Set it up free Takes about a minute. Free plans work.