Sequential Analysis: Past, Present and Future

(An Invited Talk in the FDP on Recent Developments in Statistics

and its Applications in Data Science at VIT Chennai)

Neeraj Joshi

Assistant Professor, Department of Mathematics, IIT Delhi

Email: [email protected]

March 28, 2024

Contents
• Background
• Strategy
• Example: Testing of Hypothesis
• Example: Fixed-Width Confidence Interval Estimation
• Example: Minimum Risk Point Estimation
• Example: Bounded Risk Point Estimation
• Solutions: Sequential and Multistage Sampling
• Sequential and Multistage Estimation of an Inverse Gaussian Mean
• Applications of Sequential Analysis with Future Directions
• References
Background
• Mahalanobis (1940) showed the importance of sampling in steps and developed
sampling designs for estimating the acreage of jute crops in the whole state of
Bengal. This seminal development in large-scale survey sampling of national
importance was regarded by many, including Abraham Wald, as the forerunner of
sequential analysis.
• Wald and his collaborators systematically developed theory and methodology of
sequential tests in the the early 1940s to reduce the number of sampling
inspections without compromising the reliability of the terminal decisions. This
culminated in a true classic of our time (Wald, 1947).
Background (Contd.)
• Sequential analysis began its march with deep-rooted applied motivations in
response to demands for efficient testing of anti-aircraft gunnery and other
weapons during World War II. Those methodologies were especially crucial then.
Fewer sampling inspections with accurate outcomes were essential to gain an
advantage in the front line. These developments were classified in the early to
mid-1940s.
• Methodological researchers caught on and began applying sequential analysis to
solve a wide range of practical problems from inventory, queuing, reliability, life
tests, quality control, designs of experiments, and multiple comparisons, to name a
few.
Background (Contd.)
• In the 1960s through 1970s, researchers in clinical trials realized the relevance of
emerging adaptive designs and optimal stopping rules. The area of clinical trials
continues to be an important beneficiary of some of the basic research in
sequential methodologies. The basic research in clinical trials has also enriched the
area of sequential sampling designs.
• Some important books and monographs on sequential analysis are Wald (1947),
Ghosh et al. (1997), Mukhopadhyay et al. (2004), Govindarajulu (2004), and
Mukhopadhyay and de Silva (2009).
Strategy
• Sequential Analysis is a method of statistical inference, in which we collect
information about an unknown parameter θ by taking random samples in a
sequence.
• We take one or a few observations at a time, and a common characteristic among
such sampling designs is that the total number of observations collected at
termination is a positive integer-valued random variable usually denoted by N.
• At each stage, the decision to stop or continue sampling depends on the results of
observations previously made.
Strategy (Contd.)
• In fixed sample size methodology, we estimate θ by some statistic, say Tn , obtained
from a random sample X1 , X2 , ..., Xn of fixed size n which is known to us.
• However, in sequential experiments, we make use of a stopping variable N and
random samples X1 , X2 , ..., XN . Then we estimate θ by the corresponding randomly
stopped estimator TN .
• This technique significantly reduces the number of observations required for the
inference part, which is advantageous as it reduces both time and cost respectively.
• Besides this, there exist several challenging problems of statistical inference which
can not be solved with the fixed-sample size methodologies and one is bound to
adopt sequential sampling designs to tackle them.
Example: Testing of Hypothesis
Let X1 , X2 , ..., Xn be a random sample of size n from N(µ, σ 2 ). Let us assume that, µ is
unknown but σ 2 is known. The problem is to test a simple null hypothesis
H0 : µ = µ0 against a simple alternative H1 : µ = µ1 (> µ0 ). We know from the
classical theory of Neyman and Pearson (1933) that, based on a random sample
X1 , X2 , ..., Xn of size n, the most powerful test for testing H0 against H1 at a
significance level α is given by:
√
n( X n− µ0 )
Reject H0 , if, > zα , (1)
σ
where α (0 < α < 1) is the type I error probability, Xn is the sample mean and zα is
the upper 100α% point of the standard normal distribution.
Example: Testing of Hypothesis (Contd.)
The type II error probability β is given by:
√ 
− µ0 )
n ( Xn
β=P ≤ zα | µ = µ1 ,
σ
√ √ 
n ( Xn − µ 1 ) n(µ1 − µ0 )
=P ≤ zα − | µ = µ1 ,
σ σ
 √ 
n(µ1 − µ0 )
= P Z ≤ zα − | µ = µ1 ,
σ
 √ 
n(µ1 − µ0 )
= Φ zα − . (2)
σ
Example: Testing of Hypothesis (Contd.)
Here Z is the standard normal variable. Now, if we fix n in advance, the type II error
probability β of the best α-level test is also fixed, and it may not be very small.
However, an experimenter may decide that, type II error probability should be smaller
than β , where β (0 < β < 1) is preassigned. Now, arbitrarily fixing the sample size n
in advance is not necessarily going to help in achieving the goal. Given (2), one
requires  √ 
n(µ1 − µ0 )
Φ zα − ≤ β = Φ(z1−β ) = Φ(−zβ ). (3)
σ
Example: Testing of Hypothesis (Contd.)
So, we must determine the appropriate sample size n satisfying the equation
√
n(µ1 − µ0 )
zα − ≤ −zβ (4)
σ
It leads us to  2
zα + zβ
n ≥ σ 2 = n∗ . (5)
µ1 − µ0
However, in case of unknown σ 2 , no fixed-sample procedure can provide the
solution. Hence, one needs to adopt a sequential strategy to solve this problem.
Example: Fixed-Width Confidence Interval
Estimation
• For specified d (> 0) and α (0 < α < 1), suppose one wishes to construct a
confidence interval of width 2d for the mean µ of a N(µ, σ 2 ) population having the
coverage probability at least (1 − α).
√
• If we consider the fixed sample size confidence interval [Xn ± Sn tα/2 / n] of the
required coverage probability (1 − α), tα/2 being the upper 100α/2% point of the
student‘s t-distribution with (n − 1) degrees of freedom, then the width of the
interval is proportional to Sn and is beyond our control. Here Sn stands for the
sample standard deviation.
Example: Fixed-Width Confidence Interval
Estimation (Contd.)
• On the other hand, if we fix the width to be 2d and consider the interval [Xn ± d],
√
then the coverage probability is 2Φ(d n/σ) − 1. Thus, for known σ , the sample
size needed is the smallest positive integer n ≥ n∗∗ , where n∗∗ = (a/d)2 σ 2 and ‘a’
is the upper 100α/2% point of a standard normal population.
• But since σ 2 is often unknown, hence one can not have control over the coverage
probability.
• These two cases lead us to the conclusion that the simultaneous control of the
width and coverage probability is not feasible as long as the sample size is fixed.
We need a sequential technique to achieve the goal.
Example: Minimum Risk Point Estimation
Again we consider a N(µ, σ 2 ) population. The problem is to estimate the unknown
mean µ. Based on the sample size n, the customary estimator of µ is the sample
mean Xn . Having recorded X1 , X2 , ..., Xn , suppose that the loss in estimating µ by Xn is
given by
L(Xn , µ) = A(Xn − µ)2 + cn, (6)
where A and c are known positive constants. Here (Xn − µ)2 represents the loss due
to estimation of µ and c stands for the cost per unit observations. The risk Rn (c)
corresponding to the loss function (6) is given by

Rn (c) = E[L(Xn , µ)]

Aσ 2
= + cn. (7)
n
Example: Minimum Risk Point Estimation (Contd.)
We aim to minimize this risk. The sample size n0 which minimizes the risk function (7)
is given by
 1/2
A
n0 = σ. (8)
c

Clearly, the magnitude of the optimal fixed sample size n0 remains unknown since σ 2
is often unknown. Hence no fixed sample size procedure will minimize the risk (7)
under the loss function (6). Again, we resort to a sequential sampling technique.
Example: Bounded Risk Point Estimation
Now we consider the problem of bounded risk point estimation of the mean µ of a
N(µ, σ 2 ) population. Let the loss in estimating µ by Xn be

L(Xn , µ) = A(Xn − µ)2 , (9)

where A is a known positive constant. The risk corresponding to the loss function (9) is
given by
Aσ 2
Rn (A) = . (10)
n
For specified w(> 0), suppose one wishes that the risk Rn (A) should not exceed w.
Example: Bounded Risk Point Estimation (Contd.)
The sample size required to achieve this goal is the smallest positive integer n ≥ n00 ,
where,  
A
n00 = σ2 . (11)
w

Again, since the optimal sample size n00 depends on σ 2 , no fixed sample size
procedure can achieve the goal in case of unknown σ 2 , and we intend to develop a
sequential sampling strategy.
Solutions: Sequential and Multistage Sampling
• For the hypothesis testing problem, Dantzig (1940) proposed a test procedure,
whose power function is independent of σ . He proved the non-existence of fixed
sample size procedures for this problem.
• Stein (1945) proposed a two-sample test having the property that the size of the
second-stage sample depends on the result of the first sample, thus making the
sample size a random variable.
• Wald (1947) developed the sequential probability ratio test (SPRT) to deal with this
problem. The sample size in Wald’s SPRT turns out to be smaller on average than
the one seen in (5).
Solutions: Sequential and Multistage Sampling
(Contd.)
• Wald and Wolfowitz (1948) proved that among all the tests (fixed sample or
sequential) for testing a simple null hypothesis against a simple alternative, SPRT
minimizes the average sample number under null and alternative hypotheses.
• The development of purely sequential estimation methodologies started with the
path-breaking papers of Anscombe (1950, 1952, 1953), Ray (1957), and Chow and
Robbins (1965).
• Anscombe (1950) provided the large sample theory of sequential estimation.
• Chow and Robbins (1965) developed a sequential procedure for constructing a
fixed-width confidence interval for the mean of an arbitrary population.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Starr (1966a) also proposed a sequential procedure for fixed-width confidence
interval estimation of normal mean and derived a formula for obtaining the exact
distribution of stopping time.
• The problem of minimum risk point estimation of the mean µ of a normal
population having an unknown variance σ 2 under absolute error loss function and
linear cost of sampling was originally considered by Robbins (1959).
• Starr (1966b) considered the ’asymptotic risk efficiency’ as an optimality criterion
for the sequential procedure associated with the minimum risk point estimation
problem.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Starr and Woodroofe (1969) introduced the ’regret’ (difference b/w the risk of
sequential procedure and the risk of corresponding fixed sample size procedure)
as an optimality criterion. It means that a sequential procedure is ’optimal’ if its
’regret’ is asymptotically bounded.
• Woodroofe (1977) introduced the concept of ‘second-order approximations’ in the
area of sequential estimation. It was a breakthrough in the true sense.
• In the theory of second-order approximations, one may be able to study the
behavior of remainder terms after the optimum position achieved by a fixed
sample size procedure.
Solutions: Sequential and Multistage Sampling
(Contd.)
• The literature surrounding purely sequential methodologies is very vast and still
being explored. For an overview, one may refer to Mukhopadhyay and de Silva
(2009) and the references therein.
• Some recent advancements include, Zacks (2009), Zacks and Khan (2011), De and
Zacks (2016), Mukhopadhyay and Hu (2017), Mukhopadhyay and Bapat (2017),
Bapat (2018a,b), Lalehzari et al. (2018), Mukhopadhyay and Bhattacharjee (2018),
Hu and Mukhopadhyay (2019), Mukhopadhyay and Banerjee (2019),
Mukhopadhyay and Wang (2019), Mahmoudi et al. (2019), Chaturvedi et al. (2019a)
and others.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Dantzig (1940) established the non-existence of any fixed sample size procedure to
deal with the problem of confidence interval estimation of normal mean.
• To handle this problem, Stein (1945, 1949) developed a two-stage procedure.
• The properties of Stein’s two-stage procedure were further studied by Ruben
(1961).
• Seelbinder (1953) and Moshman (1958) discussed the problem of choosing a
first-stage sample size in Stein’s procedure.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Stein’s two-stage procedure is simple to apply, as it requires only two stages and
achieves the target value of the coverage probability exactly but it has some
drawbacks.
• First of all, it is not ’asymptotically efficient’ in the sense of Chow and Robbins
(1965).
• Secondly, this procedure does not utilize a second-stage sample size for estimation
of nuisance parameter (variance in normal case), and thus, there is a loss of
information obtainable from the sample.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Moreover, the ’cost of ignorance’ of the nuisance parameter does not remain
asymptotically bounded and sometimes it leads us to considerable oversampling
(Hall, 1981 & 1983).
• To make Stein’s two-stage procedure ’asymptotically efficient’, Mukhopadhyay
(1980), proposed a ’modified’ two-stage procedure.
• On comparing two-stage procedures with purely sequential, we find that both have
their advantages and disadvantages.
Solutions: Sequential and Multistage Sampling
(Contd.)
• On one side, the two-stage procedure is simple to apply and achieves the exact
coverage probability, it is not ’asymptotically efficient’ and its ’cost of ignorance’ for
the nuisance parameter does not remain bounded.
• Purely sequential procedure is the most economical in terms of average sample
size. It is ‘asymptotically efficient’ but it is complicated to apply due to the changing
nature of sample space at each stage of sampling. It does not achieve the exact
coverage probability.
Solutions: Sequential and Multistage Sampling
(Contd.)
• To combine the advantages of two-stage and purely sequential procedures, Hall
(1981) developed a three-stage procedure to construct a fixed-width confidence
interval for a normal mean with unknown variance.
• Hall (1981), through his theoretical and numerical findings established that if a
single stage is appended to Stein’s two-stage procedure, the resulting three-stage
procedure becomes strongly competitive to both the two-stage and purely
sequential procedures.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Hall’s three-stage procedure is simple to apply, it is ‘asymptotically efficient’, the
‘cost of ignorance’ of the variance remains bounded and one can achieve the target
value of the coverage probability as closely as one pleases at the cost of only a
finite number of observations.
• Later on, for the same reasons behind the use of a three-stage procedure, Hall
(1983) developed an ‘accelerated’ sequential procedure, in which the number of
stages can be reduced by a predetermined factor at the cost of only a finite
number of observations.
Solutions: Sequential and Multistage Sampling
(Contd.)
• To deal with the problem of fixed-width confidence interval estimation of the
normal mean, Liu (1997) developed a k-stage (k ≥ 3) procedure which includes the
three-stage procedure of Hall (1981) as a special case.
• Liu (1997) established that, with a suitable value of k, the k-stage procedure can not
only be as efficient as the fully sequential procedure of Anscombe (1953) and Chow
and Robbins (1965) in terms of sample size, but also requires at most k sampling
operations. He obtained the associated second-order asymptotics as well.
Solutions: Sequential and Multistage Sampling
(Contd.)
• Some of the latest notable works in this direction comprise Zacks and
Mukhopadhyay (2006, 2009), Aoshima et al. (2011), Zhang (2013), Mahmoudi and
Shahraki (2015), Roughani and Mahmoudi (2015), Mukhopadhyay and Bapat
(2016), De and Zacks (2016), Mukhopadhyay et al. (2018), Mukhopadhyay and Hu
(2018), Chaturvedi et al. (2019b, 2020), Khalifeh et al. (2020) and others. One may
also refer to Mukhopadhyay and de Silva (2009).
Sequential and Multistage Estimation of an Inverse
Gaussian Mean
Let X1 , X2 , ..., Xn be an independent and identically distributed (i.i.d.) sequence of
inverse Gaussian (IG) random variables with a common density function
 1/2 ( )
2
λ −λ (x − µ)
f(x; µ, λ) = exp , x > 0, (12)
2π x3 2xµ2

where, µ > 0 and λ > 0 are the mean and scale parameters respectively. The
density function given in (12) is known as the probability density function (pdf) of an
IG distribution. The variance of the distribution is µ3 /λ. As λ → ∞, the IG
distribution behaves like a normal distribution.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Based on the random sample X1 , X2 , ..., Xn , the maximum likelihood estimators
(MLEs) of µ and 1/λ are given by

1X
n

bn = Xn =
µ Xi , (13)
n
i=1

 
1X
n
1 1 1
= − , (14)
b
λn n Xi Xn
i=1
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
• This distribution was first developed by Schrodinger (1915) to model the first
passage time of a Brownian motion with positive drift.
• Over many years, this distribution is well considered as an alternative to a Gaussian
model with many common inferential properties.
• It is also used for modeling positively skewed data, apart from the well-known
models such as Pareto, gamma, lognormal, folded normal etc.
• An IG distribution finds its use in many real-life applications such as in actuarial
science for claim cost analysis, in hydrology for analysis of floods, in meteorology
to model the distribution of wind speed or in physiology to study the blood
circulation and body fluids.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
• One may refer to Folks and Chhikara (1978), Chhikara and Folks (1989), Seshadri
(1993, 1999), Johnson et al. (1994) for relevant details and applications of an IG
model.
• An R package is also developed by Leiva et al. (2008) which covers various aspects
of a general class of inverse Gaussian models.
• SPRT for the mean of the distribution was developed by Edgeman and Salzburg
(1991).
• Joshi and Shah (1990) constructed a SPRT for the mean assuming a known
coefficient of variation.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
• Using the concept of prescribed ‘proportional closeness’ and zero-one loss
function, Chaturvedi (1985) developed a sequential procedure for estimating an
inverse Gaussian mean and established that his procedure is ‘asymptotically
efficient’ and ‘asymptotically consistent’ in the sense of Chow and Robbins (1965).
• An asymptotically efficient sequential estimation of an inverse Gaussian mean has
been considered by Chaturvedi et al. (1991).
• Bapat (2018a) developed a sequential procedure for the bounded risk estimation
of an inverse Gaussian mean under a weighted squared-error loss function and
obtained first and second-order asymptotics for the expected sample size and the
associated risk.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Suppose the loss incurred in estimating µ by Xn be

2
A ( Xn− µ)
L(µ, Xn ) = + cnλk , (15)
µ3
where A, c and k are fixed known positive constants.
Let Wt be a standard Brownian motion with drift ν . Then the first passage time for a
fixed level α has an IG distribution. Further, if the drift ν = 0, the mean µ tends to
infinity and also reduces the penalty. On the other hand, if the mean µ is close to 0,
the penalty is severe. This is a reasonable loss function under different real-life
applications, where the mean µ can fluctuate a lot.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
• Additionally, the benefit of taking such a loss function is that the risk function
becomes free from µ and depends only on the unknown scale parameter λ.
• The cost function Cn (> 0) is Cn = cnλk which has to depend on the problem in
hand and to be decided by the experimenter. A reasonable belief is that the cost of
each observation should go up (or down) with decreasing (increasing) variance or
standard deviation.
• We obtain the corresponding risk function by taking expectation across (15) as
follows:

Rn (c) = E[L(µ, Xn )],

A
= + cnλk . (16)
nλ
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
The value n = n0 minimizing the risk given in (16) is given by
 1/2   k+1
A 1 2
n0 = . (17)
c λ
Now clearly, the magnitude of n0 is unknown even though the expression is known.
We hence resort to developing purely sequential and multistage (two-stage,
three-stage, and accelerated sequential) sampling strategies to estimate the unknown
mean µ.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy
We first fix our pilot sample size m(> 1) and obtain a sequence of samples
X1 , X2 , ..., Xm from (12). We then proceed to gather one additional observation as
needed until the sampling is terminated according to the rule:
(  1/2   k+1 )
A 1 2
N = inf n ≥ m : n ≥ . (18)
c bn
λ

N estimates n0 from (17). We thus have our final dataset (N, Xi ), i = 1, 2, ..., N and
Pi=N
one can estimate the unknown mean µ by the sample mean XN = N−1 i=1 i
X.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy (First-Order Asymptotics)
For the estimation rule (N, XN ) with N defined in (18), for each fixed value of µ, λ, c, k,
we have as c → 0,
1. N
→ 1 w.p. 1.
n0
h s i
2. E N
n0
→ 1 for all s > 0 (asymptotic first-order efficiency).
h
i
s
3. E nN0 → 1 for all s > 0, if m > 1 + k+1
2 s.
RN (c)
4. Rn0 (c)
→ 1, if m > 5k + 6 (asymptotic first-order risk efficiency).
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy (Second-Order Asymptotics)
For the estimation rule (N, XN ) with N defined in (18), for each fixed value of µ, λ, c, k
we have as c → 0,

2
k+1 (k + 3)
E [ N] = n0 + ν+1− + o(c−1/2 ), when m > k + 2,
2 4

2
RN (c) (k + 1)
=1+ + o(c), when m > 5k + 6.
Rn0 (c) 4n0
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Simulation Analysis
• We first generated a set of pseudo random observations at-a-time from the IG
distribution with choices of µ = 5 and λ = 2.
• We obtained the corresponding results from 10,000 replications, where results are
tabulated for different combinations of k, m and n0 .
• We fix the value of A = 100 in each case. We also must ensure m > 5k + 6 to
satisfy the conditions of second-order asymptotics.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Simulation Analysis
Simulation results from 10,000 replications for the purely sequential methodology (18) with µ = 5, λ = 2, A = 100

(k, m) n0 c x, sx n, sn n/ n 0 z, sz ξ

(1,12) 100 0.0025 4.9928,0.0081 97.54,0.1503 0.9754 1.0157,0.0008 0.01

150 0.0011 5.0072,0.0064 147.74,0.1796 0.9849 1.0085,0.0004 0.0066

250 0.0004 4.9958,0.0050 247.89,0.2260 0.9915 1.0043,7 × 10−5 0.004

400 0.0002 5.0010,0.0039 397.47,0.2828 0.9936 1.0026,4 × 10−5 0.002

Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Simulation Analysis
Simulation results from 10,000 replications for the purely sequential methodology (18) with µ = 5, λ = 2, A = 100

(k, m) n0 c x, sx n, sn n/ n0 z, sz ξ

(2,17) 100 0.0012 5.0095,0.0084 94.05,0.2394 0.9405 1.0594,0.0019 0.0225

150 0.0005 5.0035,0.0068 145.28,0.2828 0.9685 1.0286,0.0013 0.015

250 0.0002 4.9982,0.0051 245.03,0.3459 0.9801 1.0126,0.0008 0.009

400 7.81 × 10−5 4.9980,0.0039 395.66,0.4308 0.9891 1.0063,0.0001 0.0056

Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Real Data Analysis
• We now demonstrate the applicability of our proposed purely sequential
procedures using a real data set from reliability studies.
• This data set was reported by Birnbaum and Saunders (1958, 1969) and
corresponds to the fatigue life, measured in cycles of 101 aluminum coupons of
type 6061-T6. These aluminum sheets were cut parallel to the direction of rolling
and oscillating at 18 cycles per second. These were then exposed to pressure at a
stress level of 31,000 psi. All coupons were tested until they failed. The full data
consists of the fatigue life of 101 aluminum coupons.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Real Data Analysis
• An IG model fits very well to the data with a p-value of 0.4603. One can also refer to
Leiva et al. (2008) for an affirmation of this claim. Treating these data as the
universe, the MLEs were µ b = 133.73 and λ b = 4587.16.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Purely Sequential Strategy: Real Data Analysis
Analysis of fatigue life data using purely sequential procedure (18) with m = 12, A = 200, µ b = 4587.15
b = 133.73, λ
k n0 c x n n/ n0 z

1 50 3.8 × 109 137.69 45 0.90 1.0045

60 2.64 × 10−9 134.38 57 0.95 1.0013
70 1.93 × 10−9 134.27 66 0.94 1.0017
80 1.48 × 10−9 134.98 75 0.93 1.0132
2 50 8.28 × 10−13 135.95 42 0.84 1.0152
60 5.75 × 10−13 134.88 56 0.93 1.0012
70 4.22 × 10−13 131.85 57 0.84 1.0211
80 3.23 × 10−13 132.49 73 0.91 1.0041
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy
We start with a sample of size m ≥ 2, where, as in Hall (1981), m is chosen in such a
manner that m = o(c−1/2 ) as c → 0 and lim supc→∞ ( nm ) < 1. Based on these m
bm . Recall that, 1/λ
bm
0
Pm  −1 −1

observations, we compute 1/λ = m−1 i=1
Xi − Xm is the
MLE of 1/λ. We define the stopping rule as:
( $  1   k+1 % )
A2
1 2
N1 = max m, +1 , (19)
c b
λm
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy
We estimate our fixed sample size n0 given in (17) by this N1 . Now, if N1 = m, then we
will not require any more observations in the second stage. However, if N1 > m, we
sample the difference N1 − m at the second stage by obtaining an additional set of
observations Xm+1 , ..., XN1 .
From the full dataset obtained by combining both stages, we propose to estimate µ
by XN1 . Now the expression for the associated risk with the terminal estimator XN1 is
given by:     
n0 N1
RN1 (c) = cλk n0 E +E .
N1 n0
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy
For the two-stage procedure defined in (19), as c → 0,
N1
= 1 a.s.,
n0

and for s > 0,  s   s

1 1
E = + o(c1/2 ).
b
λm λ
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy: Second-Order Asymptotics
For the two-stage procedure defined in (19), as c → 0, we have,
1
E ( N1 ) = n0 + + o(1),
2
1
E(N21 ) = n20 + n0 + + o(c−1/2 ),
3
2c
RN1 (c) = Rn0 (c) + + o(c1/2 ).
n0
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy: Simulation Study
Simulation results from 10000 replications of the two-stage methodology (19) with m = 15, k = 3

n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 5, λ = 0.8, A = 10
50 0.0097 5.03, 0.022 50.78, 0.38 1.0 0.50 0.50
100 0.0024 5.01, 0.017 99.29, 0.75 0.9 0.25 0.25
350 0.0002 5.00, 0.008 351.95, 2.80 1.0 0.07 0.07
500 0.0001 5.00, 0.007 495.14, 3.80 0.9 0.05 0.05
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy: Simulation Study
Simulation results from 10000 replications of the two-stage methodology (19) with m = 15, k = 3

n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 3, λ = 0.2, A = 0.1
50 0.025 2.99, 0.020 50.48, 0.37 1.0 0.02 0.02
100 0.0062 2.99, 0.015 100.29, 0.77 1.0 0.01 0.01
350 0.0005 3.00, 0.008 353.42, 2.77 1.0 0.01 0.01
500 0.0002 2.99, 0.006 503.12, 3.95 1.0 0.01 0.01
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Two-Stage Sampling Strategy: Real Data Analysis
Analysis of fatigue life data using two-stage procedure (19) with m = 6, A = 500, µ b = 4587.15
b = 133.73, λ

k n0 c x n n / n0 z ξ
1 50 9.5 × 10−9 129.61 54 1.08 0.0043 0.0043
70 4.84 × 10−9 130.96 58 0.82 0.0031 0.0031
80 3.71 × 10−9 133.99 86 1.07 0.0027 0.0027
2 50 2.07 × 10−12 131.81 40 0.80 0.0044 0.0043
70 1.05 × 10−12 133.60 59 0.84 0.0031 0.0031
80 8.09 × 10−13 131.96 89 1.11 0.0027 0.0027
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy
Let η ∈ (0, 1) be specified. We start with a pilot sample X1 , X2 , ..., Xm of size m(≥ 2),
where m is chosen in such a manner that m = o(c−1/2 ) as c → 0 and
lim supc→∞ ( nm ) < 1. Then we define:
0

(
$   1   k+1 % )
A 2 1 2
L = max m, η +1 . (20)
c bm
λ

If L = m, we do not take any more observations at the second stage. But we sample
the difference L − m at the second stage if L > m.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy
We denote the combined data from both stages as, X1 , X2 , ..., XL . We define our final
sample size as:
( $   1   k+1 % )
A 2 1 2
N2 = max L, η +1 . (21)
c bL
λ
If N2 = L, we do not need any more observations at the third stage. But, if N2 > L, we
sample the difference N2 − L at the third stage, i.e., we take additional observations
XL+1 , ..., XN2 at the third stage.
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy
Combining observations from all three stages, our final data set would be X1 , ..., XN2 .
After stopping, we estimate µ by XN2 , incurring the risk,
    
n0 N2
RN3 (c) = cλ k n0 E +E .
N2 n0
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy: Second-Order Asymptotics
For the three-stage procedure defined in (20)-(21), as c → 0,
k+2 1
E ( N3 ) = n0 − + + o(1),
2η 2
 
2(η − k) − 3 (k + 1)2 − η(k + 2) 1
E(N23 ) = n20 + n0 + + + o(c1/2 ),
2η 2η 2 3
k
 
cλ 3k
RN3 (c) = Rn0 (c) − 1− + o(c1/2 ).
η 4
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy: Simulation Study
Simulation results from 10000 replications of the three-stage methodology (20)-(21) with m = 15, k = 3

n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 5, λ = 0.8, A = 10
50 0.0097 4.99, 0.02 54.32, 0.34 1.08 0.50 0.48
100 0.0024 5.00, 0.01 85.67, 0.51 0.85 0.25 0.21
350 0.0001 4.99, 0.01 313.93, 1.21 0.89 0.07 0.07
500 0.0001 4.99, 0.01 467.11, 1.46 0.93 0.05 0.04
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy: Simulation Study
Simulation results from 10000 replications of the three-stage methodology (20)-(21) with m = 15, k = 3

n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 3, λ = 0.2, A = 0.1
50 0.025 2.99, 0.01 53.22, 0.34 1.06 0.02 0.01
100 0.0062 2.98, 0.01 84.57, 0.46 0.84 0.01 0.01
350 0.0005 3.00, 0.01 327.39, 1.15 0.93 0.02 0.01
500 0.0002 2.99, 0.01 460.18, 1.55 0.92 0.02 0.01
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Three-Stage Sampling Strategy: Real Data Analysis
Analysis of fatigue life data using three-stage procedure (20)-(21) with
b = 4587.15
b = 133.73, λ
m = 6, A = 500, µ
k n0 c x n n / n0 z ξ
1 50 9.5 × 10−9 135.31 60 1.20 0.0044 0.0042
70 4.84 × 10−9 131.16 66 0.94 0.0031 0.0030
80 3.71 × 10−9 130.73 69 0.86 0.0027 0.0026
2 50 2.07 × 10−12 137.04 41 0.82 0.0044 0.0042
70 1.05 × 10−12 130.11 70 1 0.0031 0.0030
80 8.09 × 10−13 133.67 71 0.88 0.0027 0.0025
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy
Let ρ ∈ (0, 1) be specified. We start with a sample of size m(≥ 2) from the given
population where as in Hall (1983) m is chosen in such a manner that m = o(c1/2 ) as
c → 0 and lim supc→∞ ( nm ) < 1. We start sampling sequentially with the stopping
0
time Q defined by,
(  1/2   k+1 )
A 1 2
Q = inf n ≥ m; n ≥ ρ . (22)
c b
λn

Based on these Q observations, we compute b1 .

λQ
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy
Then, we jump ahead and take N3 − Q observations, where,
 $  ! k+1 % 
 A
1/2
1
2 
N3 = max Q, +1 . (23)
 c b
λQ 

After stopping, we estimate µ by XN5 , incurring the risk

    
n0 N5
RN3 (c) = cλ k n0 E +E . (24)
N5 n0
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy: Second-Order Asymptotics
For the accelerated sequential procedure defined in (22)-(23) and m > k + 3, as
c → 0,

(k + 1)(k + 3) 1
E ( N3 ) = n0 − + + o(1), (25)
4ρ 2
2
(k + 1)
Var(N3 ) = n0 + o(c1/2 ), (26)
2ρ

cλ k
RN3 (c) = Rn0 (c) − k2 + 3 + o(c1/2 ). (27)
4ρ
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy: Simulation Study
Simulation results from 10000 replications of the ‘accelerated’ sequential
methodology (22)-(23) with m = 15 and k = 3
n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 5, λ = 0.8, A = 10
50 0.0097 5.01, 0.06 52.26, 1.1 1.0 0.50 0.53
100 0.0024 4.99, 0.04 109.29, 1.0 1.0 0.25 0.25
350 0.0001 4.99, 0.02 364.82, 1.0 1.0 0.07 0.07
500 0.0001 5.01, 0.01 499.60, 3.1 0.9 0.05 0.05
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy: Simulation Study
Simulation results from 10000 replications of the ‘accelerated’ sequential
methodology (22)-(23) with m = 15 and k = 3
n0 c x, sx̄ n̄, sn̄ n̄/n0 z̄ ξ

µ = 3, λ = 0.2, A = 0.1
50 0.025 2.98, 0.05 51.09, 0.8 1.0 0.02 0.02
100 0.0062 2.97, 0.04 101.36, 1.5 1.0 0.01 0.01
350 0.0005 3.03, 0.02 346.51, 4.9 0.9 0.02 0.02
500 0.0002 2.97, 0.01 499.36, 2.7 0.9 0.02 0.02
Sequential and Multistage Estimation of an Inverse
Gaussian Mean (Contd.)
Accelerated Sequential Strategy: Real Data Analysis
Analysis of fatigue life data using ‘accelerated’ sequential procedure (22)-(23) with
m = 6, A = 500, µ b = 133.73 and λ b = 4587.15
k n0 c x n n / n0 z ξ
1 50 9.5 × 10−9 134.311 47 0.94 0.0043 0.0045
70 4.84 × 10−9 132.46 75 1.07 0.0031 0.0031
80 3.71 × 10−9 134.78 82 1.02 0.0027 0.0027
2 50 2.07 × 10−12 130.85 54 1.08 0.0043 0.0045
70 1.05 × 10−12 135.64 75 1.07 0.0031 0.0031
80 8.09 × 10−13 131.74 87 1.08 0.0027 0.0029
Applications of Sequential Analysis with Future
Directions
• Cyber-Security (Polunchenko et al., 2012).
• Analysis of financial time series (Alp and Demetrescu, 2010.
• Marketing (Anderton et al., 1980).
• Regime-switching models (Carvalhoa and Lopes, 2007).
• Study of optimal clinical trials and novel therapies (Bartroff et al., 2008).
• Psychology and education (Hoffman, 1992).
• Change point detection (Gossman et al., 2022).
• Reliability and Life Testing (Joshi et al. 2022).
Applications of Sequential Analysis with Future
Directions (Contd.)
• In studies involving data analysis, it is often necessary to at least roughly determine
the sample size to estimate how long data collection will take and how much
money or resources will be needed.
• Therefore, a natural question arises: how many observations are needed to obtain
accurate or reliable estimators of the model parameters?
• This question is becoming more common as the concept of “big data” gains traction
among researchers in all fields.
• While textual, audio, and visual data collection through web scrapping is entering
the toolbox of social science scholars, data retrieval can be an essential part of the
research. To answer this question, sequential estimation procedures emerge.
References
• Aljarrah, M.A., Famoye, F., and Lee, C. (2015). A new Weibull-Pareto distribution. Communications in Statistics - Theory and Methods, 44: 4077-4095.

• Anis, M.Z. (2008). Estimating the mean of normal distribution with known coefficient of variation. American Journal of Mathematical and Management Sciences, 28:
469-487.

• Anscombe, F.J. (1950). Sampling Theory of the Negative Binomial and Logarithmic Series Distributions. Biometrika, 37: 358-382.

• Anscombe, F.J. (1952). Large-sample theory of sequential estimation. Proceedings of the Cambridge Philosophical Society, 48: 600-607.

• Anscombe, F.J. (1953). Sequential Estimation. Journal of the Royal Statistical Society: Series B, 15: 1-29.

• Aoshima, M., Mukhopadhyay, N., and Kobayashi, Y. (2011). Two-Stage Procedures for Estimating the Difference of Means when the Sampling Cost is Different.
Sequential Analysis, 30: 160-171.

• Arnholt, A.T. and Hebert, J.L. (1995). Estimating the Mean with Known Coefficient of Variation. The American Statistician, 49: 367-369.

• Azen, S.P. and Reed, A.H. (1973). Maximum likelihood estimation of correlation between variates having equal coefficients of variation. Technometrics, 15: 457-462.

• Banerjee, S., and Mukhopadhyay, N. (2016). A general sequential fixed-accuracy confidence interval estimation methodology for a positive parameter: illustrations
using health and safety data. Annals of the Institute of Statistical Mathematics, 68: 541–570.

• Bapat, S.R. (2018a). On Purely Sequential Estimation of an Inverse Gaussian Mean. Metrika, 81: 1005-1024.

• Bapat, S.R. (2018b). Purely Sequential Fixed Accuracy Confidence Intervals for P(X < Y) under Bivariate Exponential Models. American Journal of Mathematical and
Management Sciences, 37: 386-400.

• Bhargava, R.P. and Srivastava, M.S. (1973). On Tukey’s Confidence Intervals for the Contrasts in the Means of the Intraclass Correlation Model. Journal of the Royal
Statistical Society: Series B, 35: 147-152.
• Bhat, B.R. (1981). Modern Probability Theory, Wiley Eastern Limited.

• Bhattacharjee, D. and Mukhopadhyay, N. (2012). On SPRT and RSPRT for the Unknown Mean in a Normal Distribution with Equal Mean and Variance. Sequential
Analysis, 31: 108-134.

• Bhattacharya, P.K. and Mallik, A. (1973). Asymptotic Normality of the Stopping Times of Some Sequential Procedures. Annals of Statistics, 1: 1203-1211.

• Birnbaum, Z.W. and Saunders, S.C. (1958). A Statistical Model for Life Length of Material. Journal of the American Statistical Association, 53: 151-160.

• Birnbaum, Z.W. and Saunders, S.C. (1969). Estimation for a Family of Life Distributions. Journal of Applied Probability, 6: 319-327.

• Brazauskas, V. and Ghorai, J. (2007). Estimating the Common Parameter of Normal Models with Known Coefficient of Variation: A Sensitivity Study of Asymptotically
Efficient Estimators. Journal of Statistical Computation and Simulation, 77: 663-681.

• Buonaguidi, B. and Muliere, P. (2016). Optimal sequential testing for an inverse Gaussian process. Sequential Analysis, 35: 69-83.

• Chattopadhyay, B. and Mukhopadhyay, N. (2013). Two-Stage Fixed-Width Confidence Intervals for a Normal Mean in the Presence of Suspect Outliers. Sequential
Analysis, 32: 134-157.

• Chaturvedi, A. (1985). Sequential estimation of an inverse Gaussian parameter with prescribed proportional closeness. Calcutta Statistical Association Bulletin, 34:
215-219.

• Chaturvedi, A. (1986). Addendum to Sequential Estimation of an Inverse Gaussian Parameter with Prescribed Proportional Closeness. Calcutta Statistical Association
Bulletin, 35: 211-212.

• Chaturvedi, A. (1987). Sequential point estimation of regression parameters in a linear model. Annals of the Institute of Statistical Mathematics, 39: 55-67.

• Chaturvedi, A. (1988). On Sequential procedures for the point estimation of the mean of a normal population. Annals of the Institute of Statistical Mathematics, 40:
769-783.

• Chaturvedi, A., Alam, W., and Chauhan, K. (2013). Robustness of the sequential testing procedures for the parameters of zero-truncated negative binomial, binomial
and Poisson distributions. Journal of the Indian Statistical Association, 51: 313-328.

• Chaturvedi, A. and Gupta, R. (1994). “Linking the Estimation and Ranking and Selection Problems through Sequential Procedures: The Normal Case.” Statistics and
Probability Letters, 20:273-285.

• Chaturvedi, A. and Gupta, R. (1996). On multi-stage procedures for estimating the largest mean of k normal populations having unequal and unknown variances.
Scandinavian Actuarial Journal, 1: 64-78.

• Chaturvedi, A. and Tomer, S. (2003). Three-stage and ‘accelerated’ sequential procedures for the mean of a normal population with known coefficient of variation.
Statistics, 37: 51-64.
• Chaturvedi, A., Bapat S.R., and Joshi, N. (2019a). Multi-stage procedures for the minimum risk and bounded risk point estimation of the location of negative
exponential distribution under the modified LINEX loss function. Sequential Analysis, 38: 135-162.

• Chaturvedi, A., Bapat S.R., and Joshi, N. (2019a). Multi-stage procedures for the minimum risk and bounded risk point estimation of the location of negative
exponential distribution under the modified LINEX loss function. Sequential Analysis, 38: 135-162.

• Chaturvedi, A., Bapat S.R., and Joshi, N. (2019b). A k-Stage Procedure for Estimating the Mean Vector of a Multivariate Normal Population. Sequential Analysis, 38:
369–384.

• Chaturvedi, A., Bapat S.R., and Joshi, N. (2020). Second-order approximations for a multivariate analog of Behrens-Fisher problem through three-stage procedure.
Communications in Statistics - Theory and Methods, 49: 3466-3480.

• Chaturvedi, A., Bapat, S.R., and Joshi, N. (2021). Two-stage and sequential procedures for estimation of powers of parameter of a family of distributions. Sequential
Analysis, 40: 170-197. doi: 10.1080/07474946.2021.1912510.

• Chaturvedi, A., Chattopadhyay, S., Bapat, S.R., and Joshi, N. (2019c). Sequential point estimation procedures for the parameter of a family of distributions.
Communications in Statistics - Simulation and Computation. doi: 10.1080/03610918.2019.1612432.

• Chaturvedi, A., Pandey, S.K., and Gupta, M. (1991). On a class of asymptotically risk-efficient sequential procedures. Scandinavian Actuarial Journal, 1: 87-96.

• Chhikara, R.S. and Folks, J.L. (1989). The Inverse Gaussian Distribution: Theory, Methodology and Applications, New York: Marcel Dekker Inc.

• Chow, Y.S. and Robbins, H. (1965). On the Asymptotic Theory of Fixed Width Sequential Confidence Intervals for the Mean. Annals of Mathematical Statistics, 36: 457-462.

• Chow Y. S., Robbins, H., and Teicher, H. (1965). Moments of randomly stopped sums. Annals of Mathematical Statistics, 36: 789-799.

• Dantzig, G.B. (1940). On the non-existence of tests of Student’s hypothesis having power functions independent of σ . Annals of Mathematical Statistics, 11: 186-192.

• Davies, O.L. and Goldsmith, P.L. (1976). Statistical Methods in research and Production, Longman Group Ltd. London.

• De, S.K. (2014). Modified Three-Stage Sampling for Fixed-Width Interval Estimation of the Common Variance of Equicorrelated Normal Distributions. Sequential
Analysis, 33: 87-111.

• De, S.K. and Mukhopadhyay, N. (2019). Two-stage fixed-width and bounded-width confidence interval estimation methodologies for the common correlation in an
equi-correlated multivariate normal distribution. Sequential Analysis, 38: 214-258.

• De, S.K., and Zacks, S. (2016). Two-stage and sequential estimation of parameter N of binomial distribution when p is known. Sequential Analysis, 35: 440-452.

• Dutang, C. and Charpentier, A. (2016). CASdatasets: Insurance datasets (Official website), http://cas.uqam.ca/, version 1.0-6 (2016-05-28).
• Edgeman, R.L. and Salzburg, P.M. (1991). A Sequential Sampling Plan for the Inverse Gaussian Mean. Statistical Papers, 32: 45-53.

• Ehrenfeld, S. and Littauer, S.B. (1964). Introduction to Statistical Methods, McGraw-Hill, New York, USA.

• Fatima A and Roohi A (2015). Transmuted exponentiated Pareto-I distribution. Pakistan Journal of Statistics, 32: 63-80.

• Folks, J.L. and Chhikara, R.S. (1978). The Inverse Gaussian Distribution and its Statistical Application - A Review. Journal of the Royal Statistical Society, Series B, 40:
263-289.

• Gerig, T.M. and Sen, A.R. (1980). MLE in Two Normal Samples with Equal but Unknown Population Coefficients of Variation. Journal of the American Statistical
Association, 75: 704-708.

• Ghosh, B.K. (1970). Sequential Tests of Statistical Hypotheses, Addison-Welsley: Reading.

• Ghosh, M. and Mukhopadhyay, N. (1975). Asymptotic Normality of Stopping Times in Sequential Analysis. Unpublished Report.

• Ghosh, M. and Mukhopadhyay, N. (1979). Sequential Point Estimation of the Mean when the Distribution is Unspecified. Communications in Statistics - Theory and
Methods, 8: 637-652.

• Ghosh, M. and Mukhopadhyay, N. (1981). Consistency and Asymptotic Efficiency of Two-Stage and Sequential Procedures. Sankhya, Series A, 43: 220-227.

• Gleser, L.J. and Healy, J.D. (1976). Estimating the Mean of a Normal Distribution with Known Coefficient of Variation. Journal of the American Statistical Association, 71:
977-981.

• Gonzalez-Estrada, E. and Villasenor, J.A. (2018). An R package for testing goodness of fit: goft. Journal of Statistical Computation and Simulation, 88: 726-751.

• Govindarazulu, Z. (1981). The Sequential Statistical Analysis, American Science Press: Columbus.

• Guo, H. and Pal, N. (2003). On a normal mean with known coefficient of variation. Calcutta Statistical Association Bulletin, 54: 17-29.

• Hald, A. (1952). Statistical Theory with Engineering Applications, John Wiley and Sons, New York.

• Hall, P. (1981). Asymptotic Theory of Triple Sampling for Estimation of a Mean. Annals of Statistics, 9: 1229-1238.

• Hall, P. (1983). Sequential Estimation Saving Sampling Operations. Journal of the Royal Statistical Society, Series B, 45: 219-223.

• Haner, D.M. and Zacks, S. (2013). On Two-Stage Sampling for Fixed-Width Interval Estimation of the Common Variance of Equi-Correlated Normal Distributions.
Sequential Analysis, 32: 368-380.
• Mukhopadhyay, N. (1974). Sequential estimation of regression parameters in Gauss-Markoff setup. Journal of the Indian Statistical Association, 12: 39-43.

• Mukhopadhyay, N. (1980). A consistent and asymptotically efficient two-stage procedure to construct fixed-width confidence intervals for the mean. Metrika, 27:
281-284.

• Mukhopadhyay, N. (1982). Stein’s two-stage procedure and exact consistency. Scandinavian Actuarial Journal, 2: 110-122.

• Mukhopadhyay, N. (1988). Sequential estimation problems for negative exponential populations. Communications in Statistics - Theory and Methods, 17: 2471-2506.

• Mukhopadhyay, N. and Banerjee, S. (2019). Sequential minimum risk point estimation (MRPE) methodology for a normal mean under Linex loss plus sampling cost:
First-order and second-order asymptotics. Sequential Analysis, 38: 461-479.

• Mukhopadhyay, N. and Bapat, S.R. (2016). Multistage point estimation methodologies for a negative exponential location under a modified linex loss function:
Illustrations with infant mortality and bone marrow data. Sequential Analysis, 35: 175-206.

• Mukhopadhyay, N. and Bapat, S.R. (2017). Purely sequential bounded-risk point estimation of the negative binomial means under various loss functions: Multi-sample
problems. Sequential Analysis, 36: 490-512.

• Mukhopadhyay, N. and Bhattacharjee, D. (2018). Sequentially estimating the required optimal observed number of tagged items with bounded risk in the recapture
phase under inverse binomial sampling. Sequential Analysis, 37: 412-429.

• Mukhopadhyay, N. and de Silva, B.M. (2009). Sequential Methods and Their Applications, Boca Raton: CRC.

• Mukhopadhyay N and Ekwo M.E. (1987). Sequential estimation problems for the scale parameter of a Pareto distribution. Scandinavian Actuarial Journal, 1: 83-103.

• Mukhopadhyay, N. and Hu, J. (2017). Confidence intervals and point estimators for a normal mean under purely sequential strategies involving Gini’s mean difference
and mean absolute deviation. Sequential Analysis, 36: 210-239.

• Mukhopadhyay, N. and Hu, J. (2018). Two-stage estimation for a normal mean having a known lower bound of variance with final sample size defined via Gini’s mean
difference and mean absolute deviation. Sequential Analysis, 37: 204-221.

• Mukhopadhyay, N. and Solanky, T.K.S. (1992). Accelerated sequential procedure for selecting the best exponential population. Journal of Statistical Planning and
Inference, 32: 347-361.

• Mukhopadhyay, N. and Solanky, T.K.S. (1994). Multistage Selection and Ranking Procedures, New York: Dekker.
• Mukhopadhyay, N. and Wang, Z. (2019). A general theory of purely sequential minimum risk point estimation (MRPE) of a function of the mean in a normal
distribution. Sequential Analysis, 38: 480-502.

• Mukhopadhyay, N., Chaturvedi, A., and Malhotra, A. (2018). Two-stage procedures for the bounded risk point estimation of the parameter and hazard rate in two
families of distributions. Sequential Analysis, 37: 69-89.

• Mukhopadhyay, N., de Silva, B.M., and Waikar, V.B. (2010). On Two-Stage Confidence Interval Procedures and Their Comparisons for Estimating the Difference of
Normal Means. Sequential Analysis, 29: 44-68.

• Wald, A. (1947). Sequential Analysis, John Wiley and Sons, New York.

• Wald, A. and Wolfowitz, J. (1948). Optimum character of the sequential probability ratio test. Annals of Mathematical Statistics, 19: 326-339.

• Woodroofe, M. (1977). Second-Order Approximation for Sequential Point and Interval Estimation. Annals of Statistics, 5: 984-995.

• Zacks, S. (1966). Sequential estimation of the mean of a lognormal distribution having a prescribed proportional closeness. Annals of Mathematical Statistics, 31:
1688-1696.

• Zacks, S. (1971). Theory of Statistical Inference, John Willey and Sons, New York, USA.

• Zacks, S. (1974). Closeness and the expected sample size of sequential procedures for estimating tail probabilities in exponential distributions. Communications in
Statistics - Theory and Methods, 3: 105-120.

• Zacks, S. (2009). The Exact Distributions of the Stopping Times and Their Functionals in Two-Stage and Sequential Fixed-Width Confidence Intervals of the Exponential
Parameter. Sequential Analysis, 28: 69–81.

• Zacks, S. and Khan, R.A. (2011). Two-Stage and Sequential Estimation of the Scale Parameter of a Gamma Distribution with Fixed-Width Intervals. Sequential Analysis,
30: 297-307.

• Zacks, S. and Mukhopadhyay, N. (2006). Bounded risk estimation of the exponential parameter in a two-stage sampling. Sequential Analysis, 25: 437-452.

• Zacks, S. and Mukhopadhyay, N. (2009). On Exact and Asymptotic Properties of Two-Stage and Sequential Estimation of the Normal Mean Under Linex Loss.
Communications in Statistics - Theory and Methods, 38: 2992-3014.

• Zhang, C. (2013). Exact Evaluation of Two-Stage Procedures for Estimating the Difference Between Two Normal Means. Sequential Analysis, 32: 111-133.
THANK YOU