Point Estimators and Sources of Error
A point estimator is a sample statistic used to estimate a population parameter. The sample mean estimates the population mean , the sample proportion estimates the population proportion , and the sample variance estimates the population variance . A point estimate is useful, but one random sample cannot be expected to reproduce a population parameter exactly. Its quality depends both on the estimator and on how the sample was collected.
Let denote any population parameter and its estimator. Four desirable properties help compare estimators:
| Property | Meaning | Implication |
|---|---|---|
| Unbiasedness | The estimator’s average over repeated samples equals the parameter: . | Errors balance around the true value. A quantified bias may sometimes be corrected. |
| Efficiency | Among unbiased estimators of the same parameter, one has a smaller standard error. | Its estimates tend to cluster more tightly around the parameter. |
| Consistency | Estimates tend to approach the parameter as sample size grows. | Smaller sampling variability with larger supports more precise estimates. |
| Sufficiency | The estimator uses all the sample information relevant to the parameter. | The sample mean and proportion use the full sample; the sample median is not sufficient in the sense described here. |
The sample mean and sample proportion are unbiased for their corresponding population parameters. The sample variance divides the sum of squared deviations by , making unbiased for . For a normal population, the sample mean is more efficient than the sample median: the median is more robust to outliers, but has a larger standard error. For the mean, ; for the proportion, . Both shrink as increases, illustrating consistency.
Sampling and non-sampling error
Sampling error is the difference between a sample and its population caused by random selection. Repeated probability samples of the same size are representative on average, but any one sample can differ from the population. Sampling error is unavoidable; standard errors quantify its expected scale.
Non-sampling error arises for reasons other than random sampling and may bias estimates. A large sample does not automatically remove it.
| Error | What goes wrong | Example |
|---|---|---|
| Coverage error | The sampled population does not match the population relevant to the decision. | A retailer surveying only its current customers may miss dissatisfied competitor customers or untapped buyers when the goal is sales growth. |
| Non-response error | Some population segments are less likely to be represented or respond. | An in-store-only survey misses online buyers; an online-only survey may miss store shoppers. Different response rates across channels can also distort representation. |
| Measurement error | The collection method or question produces inaccurate answers. | Non-anonymous forms, staff watching respondents, confusing or lengthy questions, poor recall, hurried completion, or deliberate dishonesty. |
Reduce non-sampling error by defining the target population carefully, designing and training for the collection process, piloting it to find problems, and choosing a sampling method (such as stratified, cluster, or systematic sampling) that reflects relevant groups. Statistical inference combines sample data with sampling distributions to make statements about populations; sound sampling and measurement are both necessary for useful decisions.
Key takeaways
- , , and estimate , , and respectively.
- Unbiasedness concerns the estimator’s center; efficiency concerns its spread; consistency concerns what happens as grows; sufficiency concerns information used.
- Sampling error is random variation; non-sampling error can systematically mislead.
- Larger samples reduce sampling variability, but do not cure a poorly defined population or flawed measurement.
Confidence Intervals for a Population Mean
A confidence interval turns a point estimate into a range that communicates its precision. Its general form is
The confidence coefficient is . A confidence level describes the long-run proportion of intervals from repeated samples that would contain the fixed population parameter. For a particular computed interval, the parameter either lies inside it or does not; the confidence level is not the probability that a fixed parameter moves around, nor does a confidence interval for a mean contain that percentage of individual observations.
The parameter being estimated determines the sampling distribution and interval method:
Known population standard deviation: interval
When is known, the standard error is . If the population is normal or the sample is sufficiently large for the sampling distribution of to be approximately normal, then
\bar{x}\pm z_{\alpha/2}\frac{\sigma}{\sqrt{n}}.$$ Here $z_{\alpha/2}$ is the positive standard-normal critical value with upper-tail area $\alpha/2$; the margin of error is $E=z_{\alpha/2}\sigma/\sqrt{n}$. **Customer satisfaction illustration.** A random sample of $n=36$ has $\bar{x}=57.31$, and historical $\sigma=5$. Thus $SE=5/\sqrt{36}=0.833$. For 95% confidence, $z_{\alpha/2}=1.96$, so $E=1.96(0.833)=1.63$ and the interval is $57.31\pm1.63=[55.68,58.94]$. Across repeated samples, 95% of intervals formed this way would contain $\mu$. Changing confidence while keeping the estimate and standard error fixed changes the critical value and margin of error: | Confidence | $z_{\alpha/2}$ | Margin of error | Interval for $\mu$ | |---:|---:|---:|---:| | 80% | 1.28 | $1.28(0.833)=1.07$ | $57.31\pm1.07=[56.24,58.38]$ | | 90% | 1.64 | $1.64(0.833)=1.37$ | $57.31\pm1.37=[55.94,58.68]$ | | 95% | 1.96 | $1.96(0.833)=1.63$ | $57.31\pm1.63=[55.68,58.94]$ | | 99% | 2.58 | $2.58(0.833)=2.15$ | $57.31\pm2.15=[55.16,59.46]$ | Higher confidence requires a wider interval. A 100% confidence level would require an unbounded interval under this construction, which is not useful. The appropriate confidence level depends on the cost or risk of being wrong; use the same level when comparing alternatives. **Worked example: average customer bill.** A random sample of $n=50$ customers has $\bar{x}=₹2{,}200$ and historical population standard deviation $\sigma=₹84$. The large sample supports an approximately normal sampling distribution. 1. $SE=84/\sqrt{50}=₹11.88$. 2. At 80% confidence, $z_{\alpha/2}=1.28$: $E=1.28(11.88)=₹15.22$; interval $=2{,}200\pm15.22=[₹2{,}184.78,₹2{,}215.22]$. 3. At 95% confidence, $z_{\alpha/2}=1.96$: $E=1.96(11.88)=₹23.28$; interval $=2{,}200\pm23.28=[₹2{,}176.72,₹2{,}223.28]$. **Worked example: massage-session bill.** A sample of $n=10$ sessions has $\bar{x}=₹3{,}600$ and known $\sigma=₹500$. Because $n<30$, the calculation assumes the population spending distribution is normal. Then $SE=500/\sqrt{10}=₹158.11$. 1. At 95%, $z_{\alpha/2}=1.96$ and $E=1.96(158.11)=₹309.90$. The interval is $3{,}600\pm309.90=[₹3{,}290.10,₹3{,}909.90]$. 2. At 99%, $z_{\alpha/2}=2.58$ and $E=2.58(158.11)=₹407.92$ (the supplied calculation reports approximately ₹407.27). Using the supplied margin gives $[₹3{,}192.73,₹4{,}007.27]$; using the displayed standard error and critical value gives approximately $[₹3{,}192.08,₹4{,}007.92]$. The source arithmetic is slightly inconsistent on this last margin and interval; the method is unchanged. With a small sample, the normal-population assumption matters and should be checked against the data when possible. ### Unknown population standard deviation: $t$ interval In most applications $\sigma$ is unknown, so the sample standard deviation $s$ replaces it. Because $s$ is itself random, the standardized statistic follows a **Student’s $t$ distribution** with $n-1$ degrees of freedom when the sample comes from a normal population: $$t=\frac{\bar{x}-\mu}{s/\sqrt{n}},\qquad \bar{x}\pm t_{\alpha/2,n-1}\frac{s}{\sqrt{n}}.$$ The $t$ distribution is bell-shaped, symmetric around zero, and has heavier tails than the standard normal. As its degrees of freedom grow it approaches the standard normal. The interval is exact for a normal population. With non-normal data it is approximate: samples of at least 30 are usually adequate; with strong skewness or outliers, a larger sample (around 50 or more) is preferable; roughly symmetric, bell-shaped populations may work with samples as small as 15. For small samples, use the method only when approximate population normality is reasonable. **Worked example: alumni salaries.** For $n=103$, $\bar{x}=25.92$ lakhs and $s=3.66$ lakhs, so $s/\sqrt{n}=0.361$ lakhs and $df=102$. 1. At 90%, $t_{0.05,102}=1.66$; $E=1.66(0.361)=0.60$ lakhs; interval $=25.92\pm0.60=[25.32,26.52]$ lakhs. 2. At 95%, $t_{0.025,102}=1.98$; $E=1.98(0.361)=0.72$ lakhs; interval $=25.92\pm0.72=[25.20,26.64]$ lakhs. **Worked example: monthly credit-card spend.** For $n=55$, $\bar{x}=₹3{,}852.11$, $s=₹1{,}454.23$, $s/\sqrt{55}=₹196.09$, and $df=54$. 1. At 80%, $t_{0.10,54}=1.30$; $E=1.30(196.09)=₹254.92$ (reported as ₹254.41); the supplied interval is $₹3{,}597.70$ to $₹4{,}106.52$. 2. At 95%, $t_{0.025,54}=2.00$; $E=2.00(196.09)=₹392.18$ (the transcript reports ₹393.13); its reported interval is $₹3{,}458.98$ to $₹4{,}245.24$. The supplied interval endpoints do not exactly match the stated center and margin of error; retain the stated figures as approximate table-based results. Both intervals become wider at higher confidence. ### Further worked examples: unknown-$\sigma$ means **Corporate bond maturity and yield.** Fifteen bonds give average years to maturity $\bar{x}=7.567$ and $s=3.146$. With $df=14$, $t_{0.025,14}=2.145$, and $s/\sqrt{15}=0.813$: $$E=2.145(0.813)=1.743,\qquad 7.567\pm1.743=[5.824,9.310]\text{ years}.$$ For yield, the transcript gives $\bar{x}=3.552$, $s=0.931$, and $s/\sqrt{15}=0.240$. Using the same $t$ value, $E=2.145(0.240)=0.515$, giving approximately $3.552\pm0.516=[3.036,4.068]$. The source also states a center of $3.553$ and upper endpoint $4.067$; these differ by rounding. Both estimates use $n=15$ and 14 degrees of freedom. If a $t$ table lacks the exact degrees of freedom, a nearby row can be used. Choosing a lower degrees-of-freedom row gives a flatter distribution and a larger critical value, producing a conservative, wider interval. **Car-insurance premium in Jhansi.** For $n=20$, sample mean $₹19{,}215.70$ and $s=₹3{,}588.15$, the coefficient of variation is about 19%. The standard error is $3{,}588.15/\sqrt{20}=₹802.30$. With $df=19$, $t_{0.025,19}=2.093$, so $E=2.093(802.30)=₹1{,}679.30$. Thus the 95% interval for the population mean is $₹19{,}215.70\pm₹1{,}679.30=[₹17{,}536.40,₹20{,}895.00]$. The interval is for the population **mean**, not for 95% of individual premiums. The national average of ₹21,115 is outside this interval, suggesting Jhansi’s population mean may differ from the national average; this interval alone does not establish a formal hypothesis test or compare individual premiums. ### Choosing a sample size for a mean For known $\sigma$, choose $n$ before data collection by setting the desired margin $E$ equal to $z_{\alpha/2}\sigma/\sqrt{n}$ and solving: $$n=\left(\frac{z_{\alpha/2}\sigma}{E}\right)^2.$$ Use a planning value for $\sigma$ from previous studies, a pilot sample, or judgment. A rough estimate is the population range divided by 4. Always round the required sample size up. **Worked example: customer satisfaction.** For 99% confidence, $z_{\alpha/2}=2.58$, historical $\sigma=5$, and desired $E=1$: $$n=\left(\frac{2.58(5)}{1}\right)^2=166.41\text{ (approximately; the source reports }165.8\text{)},$$ The transcript reports 165.8 and rounds up to 166; however, the displayed inputs give 166.41, which must be rounded up to 167 to ensure the target margin. The small discrepancy is due to rounded or inconsistent arithmetic. **Worked example: sales-call duration.** A pilot gives planning $\sigma=4$ minutes $=240$ seconds. At 95%, $z_{\alpha/2}=1.96$. 1. For $E=75$ seconds: $n=[1.96(240)/75]^2=39.37$, so use $n=40$. 2. For $E=1$ minute $=60$ seconds: $n=[1.96(240)/60]^2=61.47$, so use $n=62$. A tighter margin requires more observations, which can raise collection cost. **Key takeaways** - Known $\sigma$ uses $z$; unknown $\sigma$ uses $t$ with $n-1$ degrees of freedom. - Higher confidence widens the interval; larger samples narrow it. - A mean interval describes uncertainty about the population mean, not the spread of individual data. - Plan sample size from the desired margin, confidence level, and a defensible estimate of variability; round up. ## Confidence Intervals for a Population Proportion A **population proportion** $p$ is estimated by the sample proportion $\hat{p}=x/n$, where $x$ counts observations meeting the chosen success criterion. The sampling distribution is approximately normal when both expected counts are adequate: $np\ge5$ and $n(1-p)\ge5$. Since $p$ is unknown, estimate its standard error by substituting $\hat{p}$: $$\hat{p}\pm z_{\alpha/2}\sqrt{\frac{\hat{p}(1-\hat{p})}{n}}.$$ **Worked example: satisfaction score at least 60.** Of 36 customers, 17 scored 60 or more. Thus $\hat{p}=17/36\approx0.47$. For 95%, $z_{\alpha/2}=1.96$ and the estimated standard error is $\sqrt{0.47(0.53)/36}\approx0.08$. The margin is $1.96(0.08)=0.16$, so the interval is $0.47\pm0.16=[0.31,0.64]$. **Worked example: spend below ₹3,000.** Of 55 customers, 16 spent below ₹3,000, so $\hat{p}=16/55\approx0.29$ and the estimated standard error is $\sqrt{0.29(0.71)/55}\approx0.06$. - At 80%, $z_{\alpha/2}=1.28$, $E=1.28(0.06)=0.08$, and the interval is $0.29\pm0.08=[0.21,0.37]$. - At 99%, $z_{\alpha/2}=2.58$, $E=2.58(0.06)=0.15$ using the displayed rounded standard error (the transcript reports 0.16), giving about $[0.14,0.44]$; using the reported margin gives $[0.13,0.45]$. The width makes the estimate less useful for precise decisions, motivating sample-size planning. ### Planning a sample size for a proportion For desired margin $E$, solve the margin equation for $n$: $$n=\frac{z_{\alpha/2}^{2}p^*(1-p^*)}{E^2},$$ where $p^*$ is a planning value chosen from a previous study, a pilot, or judgment. If none is available, use $p^*=0.5$; this gives a conservative large sample size. **Worked example: desired 99% margin of 0.05.** With $z_{\alpha/2}=2.58$ and $p^*=0.5$: $$n=\frac{2.58^2(0.5)(0.5)}{0.05^2}=665.64,$$ so plan for 666 respondents. ### Further worked examples: proportion **Uninsured vehicles.** Of 200 vehicles, 46 are uninsured, so $\hat{p}=46/200=0.23$. For 90% confidence, $z_{\alpha/2}=1.64$ and $SE=\sqrt{0.23(0.77)/200}\approx0.030$. Hence $E=1.64(0.030)\approx0.05$ and the interval is $0.23\pm0.05=[0.18,0.28]$. **Non-zero credit-card balance.** At 98% confidence, desired $E=0.03$, and planning $p^*=0.70$, $z_{\alpha/2}=2.33$: $$n=\frac{2.33^2(0.70)(0.30)}{0.03^2}=1{,}266.7,$$ so use 1,267. If no planning proportion is trusted, $p^*=0.5$: $$n=\frac{2.33^2(0.5)(0.5)}{0.03^2}=1{,}508.03,$$ so use 1,509. **Credit-card ownership by age group.** With $n=455$ in each group, 95% confidence, and $z_{\alpha/2}=1.96$, the reported intervals are: | Age group | $\hat{p}$ | Approx. margin | 95% interval | |---|---:|---:|---:| | 18–24 | 0.67 | 0.04 | $[0.63,0.71]$ | | 25–34 | 0.83 | 0.03 | $[0.80,0.86]$ | | 35–49 | 0.76 | about 0.04 | $[0.72,0.80]$ | | 50+ | 0.78 | about 0.04 | $[0.74,0.82]$ | For a common sample size that keeps every margin at or below 0.03, the $p(1-p)$ term is largest for the lowest listed planning proportion, $p^*=0.67$. Thus: $$n=\frac{1.96^2(0.67)(0.33)}{0.03^2}=943.75,$$ so use $n=944$ per group. The transcript contains a mislabeled age range once; the four age groups are 18–24, 25–34, 35–49, and 50+. **Key takeaways** - Define “success” before calculating $\hat{p}$; then $\hat{p}=x/n$. - Substitute $\hat{p}$ for unknown $p$ in the standard error. - A 0.5 planning value is a safe fallback when no prior proportion is available. - When multiple groups share a margin and confidence target, size the sample using the group with the largest $p^*(1-p^*)$. ## Confidence Intervals for Population Variance and Standard Deviation Variance measures spread, and its confidence interval uses a different sampling distribution. If the population is normal, $$\frac{(n-1)s^2}{\sigma^2}\sim\chi^2_{n-1}.$$ The chi-square distribution is asymmetric, so the interval is not written as an estimate plus or minus a symmetric margin. Let $\chi^2_{q,\nu}$ denote the chi-square value with right-tail probability $q$ and $\nu$ degrees of freedom. For confidence coefficient $1-\alpha$: $$\frac{(n-1)s^2}{\chi^2_{\alpha/2,n-1}}\le\sigma^2\le \frac{(n-1)s^2}{\chi^2_{1-\alpha/2,n-1}}.$$ Take square roots of the two endpoints to obtain a confidence interval for $\sigma$. The normal-population assumption is essential to this method. **Worked example: customer-satisfaction variance.** With $n=36$, $s^2=40.7$, $df=35$, and 95% confidence, the right-tail critical values are $\chi^2_{0.025,35}=53.20$ and $\chi^2_{0.975,35}=20.57$. Therefore: $$\frac{35(40.7)}{53.20}=26.79\le\sigma^2\le\frac{35(40.7)}{20.57}=69.30.$$ Taking roots gives $5.18\le\sigma\le8.33$. At 80%, the critical values are 46.06 (tail 0.10) and 24.80 (tail 0.90): $$\frac{35(40.7)}{46.06}=30.95\le\sigma^2\le\frac{35(40.7)}{24.80}=57.49,$$ so $5.56\le\sigma\le7.58$. At 90%, the critical values are 49.80 (tail 0.05) and 22.47 (tail 0.95): $$\frac{35(40.7)}{49.80}=28.62\le\sigma^2\le\frac{35(40.7)}{22.47}=63.46,$$ so $5.35\le\sigma\le7.97$. **Worked example: alumni salaries.** For $n=103$, $s=3.66$ lakhs and $s^2=13.39$ lakhs squared, with $df=102$. At 95%, the right-tail values are 131.84 for tail 0.025 and 75.95 for tail 0.975: $$\frac{102(13.39)}{131.84}=10.36\le\sigma^2\le\frac{102(13.39)}{75.95}=17.98.$$ Thus $3.22\le\sigma\le4.24$ lakhs. ### Further worked examples: volatility For returns, the sample standard deviation expresses volatility in the same units as returns; its square is variance. **Quarterly stock returns.** Eight quarterly percentage returns have sum 25.6, so $\bar{x}=25.6/8=3.2\%$. The transcript’s stated sum of squared deviations (7,873.6) conflicts with its reported variance; the reported $s^2=253.37$ implies a sum of $1,773.6$. Dividing that by $n-1=7$ gives $s^2=253.37$, and $s=\sqrt{253.37}=15.92$. The individual eight return values are not stated, so the deviation sum is the available step in the calculation. For a 95% interval, $df=7$ and the right-tail critical values are $\chi^2_{0.025,7}=16.01$ and $\chi^2_{0.975,7}=1.69$: $$\frac{7(253.37)}{16.01}=110.76\le\sigma^2\le\frac{7(253.37)}{1.69}=1{,}049.54.$$ Taking square roots gives $10.52\le\sigma\le32.40$ percentage points. **Daily stock-price changes.** For 12 companies, the reported sample mean is 0.78, the sum of squared deviations is 5.223, and $df=11$. Thus $s^2=5.223/11=0.4748$ and $s=\sqrt{0.4748}=0.6891$. At 95%, the right-tail critical values are 21.92 (tail 0.025) and 3.82 (tail 0.975): $$\frac{11(0.4748)}{21.92}=0.24\le\sigma^2\le\frac{11(0.4748)}{3.82}=1.37.$$ Therefore $0.49\le\sigma\le1.17$. The underlying 12 price changes are not stated, so the given sum of squared deviations is the available calculation detail. > Exam tip: For variance, the larger chi-square critical value is the denominator of the lower endpoint; the smaller critical value is the denominator of the upper endpoint. Taking reciprocals reverses the inequality. **Key takeaways** - A chi-square interval for variance requires a normal population and uses $n-1$ degrees of freedom. - The interval is asymmetric; divide $(n-1)s^2$ by the two tail critical values in reverse order for lower and upper bounds. - Square-root the variance endpoints to get the standard-deviation interval. - Wider confidence levels produce wider intervals, often especially noticeably with small samples.Revise before your examAll 1 exam tip from this module in one list
- For variance, the larger chi-square critical value is the denominator of the lower endpoint; the smaller critical value is the denominator of the upper endpoint. Taking reciprocals reverses the inequality.
Test yourself on this module
Questions from these notes, one at a time, with feedback. Your weak topics are saved.