STATISTICAL ESTIMATION

Structure
18.0 Objectives
18.1 Introduction
18.2 Statistical Background
18.3 Concept of Statistical Inference
18.4 Point Estimation
18.5 Confidence Interval for Known Variance
18.6 Confidence Interval for Unknown Variance
18.7 Let Us Sum Up
1 8.8 Key Words
18.9 Some Useful Books
18'.10 Answers/Hints to Check Your Progress Exercises

After going through this Unit you will be in a position to:

explain the concept of estimation;
distinguish between point estimate and interval estimate;
estimate confidence interval for a parameter; and
explain the concept of confidence level.

18.1 INTRODUCTION
Many times due to certain constraints such as inadequate funds or manpower or time
we-are not in a position to survey all the units in a population. In such situations we
take resort to sampling, that is, we survey only a part ofthe population. On the basis of
the'information contained in the sample we try to draw conclusions about the population.
This process is called statistical inference. We must emphasise that statistical inference
is widely applied in economics as well as in many other fields such as sociology,
psychology, political science, medicine, etc. For example, before election process starts
or just before declaration of election results many newspapers and television channels
conduct exit polls. The purpose is to predict election results before the actual results
are declared. At that point of time, it is n ~ t ' ~ o s s i bfor
l e the surveyors to ask all the
voters about their preferences for electoral candidates -the time is too short, resources
are scarce, manpower is not available, and a complete census before election defeats
the very purpose of election!
In the above example the surveyor actually does not know the result, which is the
outcome of votes cast by all the voters. Here all the voters taken together comprise the
population. The surveyor has collected data from a representative sample of the
population, not all the voters. On the basis of the information contained in the sample,
(s)he is making forecast about the entire population.
In this Unit we deal with the concept of statistical inference and methods of statistical
estimation. Parameter, as you know, is a function of population units while statistic is
a hnction of sampling units. There could be a number ofparameters and corresponding
Statistical Inference statistics. However, in order to keep our presentation simple, we will confine ourselves
mostly to arithmetic mean.

18.2 STATISTICAL BACKGROUND.

In the previous two blocks we have discussed two important aspects: theoretical
probability distributionsand sampling techniques. These two aspects form the basis of
statistical inference.
In Unit 14, Block 5 we explained the concept of a random variable. We learnt that Xis
arandom variable if it assumes values x, ,x, ,.........,x, with correspondingprobabilities
p, , p 2,......,p, attached to it. Here the probability of occurrence of x, is p, , the
probability of occurrence of x, is p, , and so on. If the values x, ,x, ,.........,X,, are
discrete we call X a discrete random variable and find out the probability for isolated
values of X. On the other hand, ifX is a continuous random variable we can find out
the probability ofxwithin certain range such that P(a I X I b) = p, .
In Units 14 and 15 of Block 5 we discussed theoretical discrete probability distributions
(such as binomial and Poisson) and continuous probability distributions (such as normal
and t). We learnt that if the range of X increases infinitely then these probability
distributions approach normal distribution.Thus normal distribution is a limiting case
of these probability distributions and is considered as a sort of ideal among probability
distributions.
The normal distribution is defined by two parameters: mean ( p )and standard deviation
( a ). If the probabilities associated with a random variable are distributed according
to normal distribution (that means, ifxfollows normal distribution), we can find out
the probability of P(a I X _< b) = p, by using the equation for its probability
distribution function.
A problem encountered here is that p and a can take any values and finding out
corresponding probability is time consuming. This problem is tackled by subtracting
p from the normal variable and dividing it by a.This way we obtain the 'standard

normal variate', z = -, which has mean= 0 and standard deviation = 1. By plotting

a
the probabilities for different values of z on a graph paper we obtain 'standard normal
curve' which is symmetrical and area under the curve is = 1. Remember that in the

-'
case of standard normal curve we measure z = -on the x-axis and probability of
a
occurrence of z, that is p(z) ,on the y-axis. Thus if we,consider a particular segment
of the normal curve (bounded by two values of z, say, z, and z, ) the area under the
curve gives its probability. Remember that normal curve is different from the frequency
curve considered in Block 1 of this course. Area under the normal curve qoes not give
frequencies; it gives probabilities.
In Unit 16 of Block 6 we leant that v e j often it is not possible to study the entire
population and we undertake a sample survey. If the samplejs drawn in a random
manner through appropriateprobability attached to each population unit and the sample
size is not very small, the sample can be a representative one ofthe popu1ation:Recall
that we can draw a number of samples from a given population and each sample
provides us with a sample mean. Thus the sample means can be arranged in the form
of a frequency distribution, called the 'sampling distribution'.
We know from Unit 16, Block 6 that sample.mean (x) assumes different values and Statistical Estimation

for each value we can attach a probability. Thus sample mean can be considered as a
random variable. In real life situations we have a finite population and the number of
samples (and therefore the number of sample means) is finite. In this case z is a
discrete random variable but when there are infinite number of samples, x could be a
continuous random variable.
Now let us consider another important concept discussed in Unit 16: the central limit
theorem. It says that sampling distribution of is normal if the population from
which the sample is drawn is normal. However, sampling distribution of is
approximately normal if sample size (n) is large, even if the parent population (that is,
population from which it is drawn) is not normal.,If the parent population is
approximately normal then sampling distribution of sample means is approximately
normal even when sample size is small.
We know that dispersion of sample means is smaller in value than dispersion of the
parent population from which the sample is drawn. Recall that the standard deviation
of the sampling distribution is called standard error. Thus if the population has a
0.
standard deviation o then the standard error of sample mean is -
&
From the above we learn that sample mean can be considered as a,random variable and
it approximates normal distribution when sample size is large. Usually we consider a
sample to be large in size if n > 30 .For small samples ( n 2 30 ), sampling distribution
of sample means is similar to student's t distribution. Recall that in the case of t
distribution the shape of the probability curve changes according to its 'degrees of
freedom'.

18.3 CONCEPT OF STATISTICAL INFERENCE

As mentioned earlier, statistical inference deals with the methods of drawing conclusions
about the population characteristics on the basis of information contained in a sample
drawn from the population. Remember that population mean is not known to us, but
we know the sample mean. In statistical inference we would be interested in answering ,
two types of questions. First, what would be the value of the population mean? The
answer lies in making an informed guess about the population mean. This aspect of
statistical inference is called 'estimation'. The second question pertains to certain
assertion made about the population mean. Suppose a manufacturer of electric bulbs
claims that the mean life of electric bulbs is equal to 2000 hours. On the basis of the
sample information, can we say that the assertion is not correct? This aspect of statistical
inference is called hypothesis testing.
Thus statistical inference has two aspects: estimation and hypothesis testing. We will
discuss about statistical estimation in the present Unit while testing of hypothesis will
be taken up for discussion in the next Unit.
Fig. 18.1 below summarises different aspects of statistical inference. A crucial factor
before us is whether we know the population variance or npt. Of course when we do
not know the population mean, how do we know the population variance? We begin
with the case where population variance is known, because it will help us in explaining
the concepts. Later on we will take up the more realistic case of unknown population
variance.
Estimation could be of two types: point estimation and interval estimation. In point
estimation we estimate the value of the population parameter as a single point. On the
other hand, in the case of interval estimation we estimate lower and upper bounds
around sample mean within which population mean is likely to remain.
Statistical Inference The assertion or claim made about the population mean would be in the form of a null
hypothesis and its counterpart, alternative hypothesis. We will explain these concepts
and the methods oftesting of hypothesis in the next Unit.

Hypo! hesii
Estimation

when variance when variance

is not known

Fig. 18.1: Statistical Inference

Check Your Progress 1

1) Explain the following concepts.
a) standard normal variate
b) random variable
c) sampling distribution
d) central limit theorem

.......................................................................................................................
2) State whether the following statements are true or false.
a) Normal distribution is a limiting case of binomial distribution.
b) Standard deviation of sampling distribution of a statistic is termed as
standard error.
c) Poisson distribution is an example of continuous distribution.
d) Statistical estimation is a p& of statistical inference.
 Statistical Estimatio~l
18.4 POINT ESTIMATION
As mentioned earlier we do not know the parameter value and wankto guess it by using
sample information. Obviously the best guess will be the value of the sample statistic.
For example, if we do not know the population mean the best guess would be the
sample mean. Here in this case we use a single value or point as 'estimate' of the
parameter.
In Unit 16 we have explained the concepts of estimate and estimator. Also we have
pointed out the distinction between the two. Recall that estimator is the formula and
estimate is the particular value obtained byusing the formula. For example, if we use
1
sample mean for estimation of population mean, then - C xi is the estimator. Suppose
?il
I collect data on a sample, and put the sampling units to this formula and obtain a
particular value for sample mean, say 120. Then 120 is an estimate of population
mean. It is possible that you draw another sample from the same population, use the
1
formula for sample mean, that is -C xi ,and obtain a different value, say 123. Here
n i
both 120 and 123 are estimates ofpopulation mean. But in both the cases the estimator
1
is the same, which is - C x, . Remember that the term
statistic, which is used to mean
n
a function of sample values, is a synonym for estiinator.
There may be situations when you would find moie than one potential estimator
(alternative formulae) for a parameter. In order to choose the best among these
estimators, we need to follow certain criteria. Based on these criteria an estimator
should fulfill certain desirable properties. There are quite a few desirable properties
for an estimator, but the most important is its unbiasedness.
Unbiasedness means that an estimate may be higher or lower than the unknown value
of the parameter. But the expected value of the estimate should be equal to the parameter.
For example, sample mean ( 3)may fluctuate ftom sample to sample but on an average
it would be equal to population mean. In other words, E(2) = p .

"
However, 1
1

i=l
bi - P' is not an unbiased estimator of the population variance

unbiased estimator of .2 . Usually a sample is less dispersed t h q the population from

which it is drawn. Therefore, there is a tendency for the sample standard deviations to
be little less than population standard deviation o .In order to rectifj this condition we
artificially inflate s by dividing by a smaller number (n-I), instead of n.
The point estimate is quite important for testing of hypothesis, as we will see in
Unit 19.

18.5 CONFIDENCE INTERVAL FOR

KNOWN VARIANCE
We have seen above that in point estimation, we estimate the parameter by a single
value, usually the corresponding sample statistic. The point estimate may not be realistic
in the sense that the parameter value may not exactly be equal to it. An alternative
procedure is to give an interval, which would hold the parameter with certain probability.
Statistical Inference Here we specify a lower limit and an upper limit within which the parameter value is
likely to remain. Also we specify the probability of the parameter remaining in the
interval. We call the intervalas 'confidence interval' and theprobability of the parameter
remaining within this interval as 'confidence level' or 'confidence coefficient'.
Let us take an example. Suppose you are asked to estimate the average income of
people in Raigarh district of Chhattisgarh state. You collected data from a sample of
500 households and found the average income (say, 2 ) to be Rs. 18250 per annum.
This sample average may not be equal to the actual average income of Raigarh district
of Chhattisgarh ( p ) because of sampling error. Thus we are not sure whether average
income of the above district is Rs. 18250 or not. On the other hand, it will be more
sensible if we say that average income of Raigarh district of Chhattisgarh is between
Rs,17900 and Rs. 18600 per annum. Also we may specify that'the probability that
average income will remain within these limits is 95 per cent. Thus our confidence
intkrval in this case is Rs.17900-18600 and the confidence level or confidence coefficient
is 95 per cent. \

Here a question may be shaping up in your mind, 'How do we find out the confidence
ingerval and confidencecoeEcient?' Let us begin with confidence coefficient. We know
that the sampling distribution of ,- for large samples is normally distributed withmean
0-
, and
U standard error , where n is the size of the sample. By transforming the
&
2-p
sample mean ( = -) we obtain standard normal variate, which has zero mean
o/&
and unit variance. The standard normal curve is symmetrical and therefore, the area
under the curve for 0 5 z 5 oo is 0.5 which is presented in the form of atable (See Table
15.1 in Unit 15 of Block 5). Let us assume that we want our confidence coefficient to
be 95 per cent (that is, 0.95). Thus we should find out a range for z which will cover
0.95 area of the standard normal curve. Since distribution of z is symmetrical, 0.475
area should remain to the right and 0.475 area should remain to the left of z = 0 . If
look into normal area table (Table 15.1) we find that 0.475 area is covered when
z = 1.96. Thus the probability that z ranges between -1.96 to 1.96 is 0.95. From this
information let us work out backward and find the range within which p will remain.
We find that

Let us interpret the above. Recall that each sample would provide us with a different
value of 2 . Accordingly, the confidence interval would be different. In each case
the confidence interval would contain the unknown parameter or it would not.
Equation (18.2) means that if a large number of random samples, each of size n,
 are drawn from the given population and if for each such sample the interval Statirtiarl %timation
!

(3

+1 is determined, then in about 95% of the cases, the interval

1 will include the population mean P .

The confidence coefficient is denoted by (1 -a ) where a is the level of significance
(we will discuss the concept of 'level of significance' in Unit 19).Confidence coefficient
could take any value. We can ask for a confidence level of say 81 per cent or 97 per
cent depending upon how precise our conclusions should be. However, conventionally
two confidence levels are frequently used, namely, 95 per cent and 99 per cent. Of
course at times we take 90 per cent confidence level also, though not frequently.
Let us find out the confidence interval when confidence coefficient( 1-a ) = 0.99. In
'
this case 0.495 area should remain on either side of the standard normgl curve. If we
look into the normal area table (Table 15.1) we find that 0.495 area is cove~edwhen z
= 2.58.

! Thus

L By rearranging the terms in the above we find that

Equation (18.4) implies that 99 per cent confidence interval for p is given by

i By looking into the normal area table you can work out the confidence interval for
confidence coefficient of 0.90 and find that

I We observe from (18.2), (18.4) and (1 8.5) that as the interval widens, the chance for

I
the interval holding a population parameter (in this case p ) increases.
The two limits of the confidence interval are called conJdence limits. For example, for

I
Q
95 per cent confidence level we have the lower confidence limit as 2 - 1.96- and the
J;;
(3
upper confidence limit as 2 + 1 - 9 6. The confidence coefficient can be interpreted
J;;
as the confidence or trust that we place in these limits for actually holding p .

Example 18.1
A paper company wants to estimate the average time required for a-new machine to
produce a ream of paper. A random sample of 36 reams shows an average production
time of 1.5 minutes per ream of paper. The population standard deviation is known to
be 0.30 minute. Construct an interval estimate with 95% confidence limits.
The information given is
x = 1.5, a = 0.30 and n = 36
 C

Statistical Inference Sihce n = 36 ( > 30), we can take the sample as a large sample and accordingly ,- is
a
normally distributed with mean p and standard error -.Now, the standard error is
&
a 0.30
- 0.05
&-,I%
The 95% confidence interval is given by

Thius with 95% confidence, we can state that the average production time for the new I
machine will be between 1.402 minutes and 1.598 minutes. Here, 1.402 is the lower 1
confidence limit and 1.598 is the upper confidence lim!t. I
18.6 CONFIDENCE INTERVAL FOR
UNKNOWN VARIANCE
In l$e previous Section we estimated confidence interval for population mean on the
asSumption that population variance is known. It is a bit unrealistic that we do not
know population mean (we want to estimate it) but know population variance. A realistic
case would be the assumption that both population mean and variance are unknown.
On the basis of sample mean and variance we want to find out confidence interval for
population mean.
Sincethe population standard deviation ( o ) is not known we use the sample standard
deviation ( s ) in its place. However, in such a case the sampling distribution of 5 is
not normal, rather it follows student's t distribution. The standard error of the sample
S
maans would be -
J;;
Like the standard normal variate z, the t-distribution has a mean of zero, is symmetrical
about mean and ranges between - a, to a, . But its variance is greater than 1. Actually
its variance changes according to degrees of freedom. However, when n > 30 the t-
distribution has a variance very close to 1 and thus resembles z-distribution.
The t-statistic, like the z-statistic, is calculated as

By looking into the area table for t-dis&ution (see Table 15.3 in Unit 15) we find the
probability values for the confidence level that we require. The degrees of freedom is
( n - 1). Thus the confidence interval would be

Example 18.2
The mean weight (in kilogram) of 20 children are found to be 15 with a standard
deviation of 4. On the basis of the above information estimate 95 per cent confidence
interval for mean weight of the population from which the sample is drawn. Assume Statistical Estimation
7

that population is normally distributed:

Since population is normal and sample size is small we apply t-distribution for estimation
of confidence interval. Since n = 20 we have degrees of freedom (d.f.) = 19. We move
down the first column of Table 15.3 till we reach the row corresponding to 19. Since
we need 95 per cent confidence interval we should leave 0.025 area on each side o f t =
0 as we did in the previous Section. Thus for 19 degrees of ffeedom and a = 0.025 we
find that t-value is 2.093. I

Hence the confidence interval is

Similarlyyou can find out confidence intervals for different sample sizes and confidence
coefficients.
Let us summarise.the rules for application of z or t statistic for estimation of confidence
interval.
1) If sample size is large (n>30) apply z-statistic - it does not matter whether
i) parent population is normal or not, and ii) variance is known or not.
2) If sample size is small (n 2 30) check whether i) parent population is normal, and
ii) variance is known.
a) If parent population is not normal apply nonpararnetric tests.
b) If parent population is normal and'variance is known apply z.
c) If parent population is normal and variance is not known apply t.
In Fig. 18.2 we present the above in the form of a chart.

Fig. 18.2: Selection of Proper Test Statistic

Statistical Inference Check Your Progress 2

1) A sample of 50 employees were asked to provide the distance commuted by them

to reach office. If sample mean was found to be 4.5 km. find 95 percent confi-
dence interval for the population. Assume that population is normally distributed
with a variance of 0.36.

.2) For a sample of 25 students in school the mean height was found to be 95 cm.
with a standard deviation of 4 cm. Find the 99 percent confidence interval.

3) State whether the following statements are true or false.

a) When parent population is not normal and sample size is small we use
t-distribution to estimate confidence interval.
b) The range of t-distribution is 0 to infinity.
c) When confidence level is 90 per cent, level of significance is 10 per cent.

18.7 LET US SUM UP

Drawing conclusions about a population on the basis of sample informati0n.i~called
statistical inference. Here we have basically two things to do: estimation andlypothesis
testing. Inthis unit we took up the first issue while the second one will be discussed in
the remaining units of the block.
An estimate of an unknown parameter could be either a point or an interval. Sample
mean is usually taken as a point estimate of population mean. On the other hand, in
intervalestimationwe construct two limits (upper and lower) around the sample mean.
We can say with stipulated level of confidence that the population mean, which we do
not know, is likely to remain within the confidence interval. In order to construct
confidence interval we need to know the population variance or itsestimate. When we

T
ow population variance, we apply normal distribution to construct the confidence
i terval. In cases where population variance is not known, we use student's t for the
bove purpose. Remember that when sample size is large (n>30) t-distribution
approximatesnormal distribution. Thus for large samples, even if population variance
is not known, we can use normal distribution for constructionof confidence interval on
the basis of sample mean and sample variance.

18.8 KEY WORDS

Confidence Level : It gives the percentage (probability) of samples where the
population mean would remain within the confidence
interval around the sample mean. If a is the significance
level the confidence level is ( 1 - a ).
Estimation : It is the method of prediction about parametet value on the
. basis of statistic.
 Estimator : It is another name given to statistic in the theory of Statistical Estimation
estimation.
Parameter : It is a measure of some ,characteristicof the population.
Population It is the entire collection of units of a specified type in a
given place and at a particular point of time.
Random Sampling : It is a procedure where every member of the population
has a definite chance or probability of being selected in the
sample. It is also called probability sampling.
Sample : It is a sub-set of the population. It can be drawn from the
I
i
population in a scientikc manner by applying the rules of
probability so that personal bias is eliminated. Many
samples can be drawn from a population and there are
many methods of d r a ~ n agsample.
Sampling Error : In the samplingmethod, we try to approximate some feature
of a given population from a sample drawn from it. Now,
since in the sample all the members of the population are
not included, howsoever close the approximation is, it is
not identical to the required population feature and some
error is committed. This error is called the sampling error.
Significance Level : There may be certain samples where population mean
would not remain within the confidence interval around
sample mean. The percentage (probability) of such cases
is called significance level. It is usually denoted by a .
When a = 0.05 (that is, 5 percent) we can say that in 5 per
cent cases we are likely to reach an incorrect decision or
commit Type I error. Level of significance could be at any
level but it is usually taken at 5 percent or 1 percent level.
Statistic : It is a function of the values of the units that are included
in the sample. The basic purpose of a statistic is to estimate
some population parameter.
Sampling Distribution : It is the relative frequency or probability distribution of
the values of a statistic when the number of samples tends
to infinity.
Standard Error : It is the standard deviation of the sampling distribution of
a statistic.
Statistical Inference : It is the process of concluding about an unknown population
from a known sample drawn from it.
Problem of Estimation : We may be interested in some feature of the population
that is completely unknown to us and we want to make
some intelligent guess about it on the basis of a random
sample drawn from the population. This problem of
statistical inference is known as the problem of estimation.
-
18.9 SOME USEFUL BOOKS
Nagar, A. L. and Das, R. K., 1989, Basic Statistics: Oxford University Press, Delhi,
Chapter 9.
Newbold, P., 1991, Statisticsfor Businsss and Economics (Third Edition): Prentice
Hall, New Jersey, Chapters 6, 7, 8 and 9.
Keller, G, and B. Warrack, 1991, Essentials of Business Statistics, Wordsworth
Publishing Co., California, Chapters 7 and 8.
Statlstlcal Inference
18.10 ANSWERSIHINTS TO CHECK YOUR
PROGRESS EXERCISES
Check Your Progress 1

1) Go through Section 18.2 and answer.

2) a) true b) true c) true d) true

Chcick Your Progress 2

1) Since it is large sample we applyz-statistic. The confidence interval is

4.40 1p 14.60
2) Since it is small sample and population variance is not given we apply t-statistic
with degrees of freedom 24. The tabulated value o f t at 99 per cent confidence
level is 2.49. The confidence interval is 93.01 5 p I 96.99.
3) a) false b) false c) true