INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)

ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

Effects of Missing Data on the Parameters of Multiple Regression

Model (MRM)
Etaga Harrison. O*, Ngonadi Lilian, Aforka Kenechukwu F and Etaga Njideka C
Department of Statistics, Nnamdi Azikiwe University, Awka, Anambra State, Nigeria
*Corresponding author
Received: 27 January 2023; Revised: 15 February 2023; Accepted: 21 February 2023; Published: 21 March 2023

Abstract: Multiple Regression Models are used in prediction the nature of relationship between one dependent variable and more
than more independent variables. There are so many assumptions the guide the estimation of the parameters of the model. The
interpretations of parameters are always subjected to the nature of data involved. Missing values tends to limit the fullness of
information in analysis. It is therefore necessary to check for the effect of missing data on the parameters of the Multiple Regression
Model. Data were simulated using Binomial, Geometric, Normal and Exponential Distribution. The simulation was done at different
sample sizes of 15, 25, 50 and 100. The level of missingness was moderated at 5%, 10%, 25% and 35%. Two methods of handling
missing data were employed, listwise deletion and Mean imputation. Data were analysis using multiple regression and Analysis of
Variance. The results shows that the least Mean Square Error (MSE) were obtained at different level of missingness depending of
the distribution. There was a significant effect on the parameters of the multiple Regression base on sample sizes.
Keywords: Multiple, Regression, Model, Missing data, MSE
I. Introduction
Researchers have faced the problem of missing quantitative data at some point in the work. Missing data can occur if research
informants refuse or forgot to answer a survey question or there might be lost of files as well as data might not be recorded properly.
Given the high cost of collecting data, there cannot be wastage of effort of starting all over or to wait until soundproof methods of
collecting information are developed. In statistics, Missing data/value is an occurrence when there is no data value stored for the
variable in an observation.
Regression analysis is the study of the nature of relationship between dependent variable(s) and independent variable(s). The Simple
Linear regression involves just one dependent and one independent variable. The situation where there exist one dependent variable
and more than one independent variable is referred to as Multiple Regression (MR). When estimating the parameters of the Multiple
Regression, Least Squares Method (LSM) is used most often. There are various factors that can affect the signs or magnitudes of
the parameter(s), one of such is that of missing data. There is need to adequately address the problem of missing data before analysis
the data to avoid reaching wrong conclusions.
When treating missing data, the most common method and the easiest to apply is the use of only those cases with complete
information. An alternative to complete case analysis, there is the use of the mean as a replacement of the missing value. More
recently, there are methods that are based on distributional models for the data (such as maximum likelihood and multiple
imputation).
Methods of analyzing missing data require assumptions about the nature of the data and about the reasons for the missing
observations that are not often acknowledged. There is need to carefully considered the required assumptions before treating missing
data. Missing data can lead to problems that affect the interpretation and inference of research results, the understanding and
explanation of conclusions made, the strength of the study design, the validity of conclusions about the relationship between
variables and may limit the representativeness of the sample.
Avoiding missing data is the optimal means of handling incomplete observations. During data collection phase, the researcher has
the opportunity to make decisions about what data to collect, and how to monitor data collection. The scale and distribution of the
variables in the data and the reasons for missing data are two critical issues for applying the appropriate missing data techniques.
This paper there seek to evaluate the effects of missing data on the parameter estimates of the Multiple Regression.
II. Literature
There are various forms of missingness. Rubin (1976) introduced the term “Missing Completely At Random” (MCAR). MCAR is
a situation where the events that led to the particular data item being missing are independent both of observable variables and of
unobservable parameters of interest and occur entirely at random. There is another missingness referred to as Missing AT Random
(MAR), which occurs when the missingness is not random, but where missingness can be fully accounted for by variables where

www.rsisinternational.org Page 63
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

there exists complete information. This is when the probability that the responses that are missing depends on the set of observed
responses, but not related to the specific missing values.
Sunbul (2018) opined in his study aimed at investigating the impact of different missing data handling methods on model parameter
estimation and classification accuracy. Simulated data were used and the data were generated by manipulating the number of items
and sample size. In the generated data, two different missing data mechanisms (Missing Completely At Random and Missing At
Random) were created according to three different amounts of missing data. The generated missing data was completed by using
methods of treating missing data as incorrect, person mean imputation, two-way imputation, and expectation-maximization
algorithm imputation.
Graham (2009) noted that most missing data are due to survey non-response, which can vary from an intentional decision
(discarding a survey or skipping sensitive items) to a rather more complex reason(s). The problem of missing data is relatively
common in almost all researches and can have significant effects on the conclusions that can be drawn from the data.
Pigott (2001) reviewed methods of handling missing data in a research study. He observed that many researchers use ad hoc methods
such as complete case analysis, available case analysis (pairwise deletion), or single-value imputation. Though these methods are
easily implemented, they require assumptions about the data that rarely hold in practice. Model-based methods such as maximum
likelihood using the EM algorithm and multiple imputation hold more promise for dealing with difficulties caused by missing data.
Rubin (1976) introduced the term “Missing Completely At Random “(MCAR) to describe data where the complete cases are a
random sample of the originally identified set of cases.
Little and Rubin, Schafer (1997) discuss methods that can be used for non-ignorable missing data. They observed that ruling out a
non-ignorable response mechanism can simplify survey items.
Schafer (1997) reports on simulation studies that provide evidence of the robustness with missing data and using missing variable
code as a predictor in a regression model. Penny et al (2012 stated in their book that the simplest approach to missing data, and the
one that is the default in virtually all statistical packages is the method known to statistician as complete case analysis but more
commonly known among social scientist as listwise deletion, in this method, cases are deleted from the sample if they have missing
data on any of the variables in the analysis to be conducted.
Peter et al (2015) observed that Missing data are part of almost all researches and introduce an element of ambiguity into data
analysis. It follows that there is need to consider them appropriately in order to provide an efficient and valid analysis. they compared
6 different imputation methods: Mean, K-nearest neighbors (KNN), fuzzy K-means (FKM), singular value decomposition (SVD),
Bayesian principal component analysis (bPCA) and multiple imputations by chained equations (MICE). Comparison was performed
on four real datasets of various sizes (from 4 to 65 variables), under a missing completely at random (MCAR) assumption, and
based on four evaluation criteria: Root mean squared error (RMSE), unsupervised classification error (UCE), supervised
classification error (SCE) and execution time. Our results suggest that bPCA and FKM are two imputation methods of interest
which deserve further consideration in practice.
Nakai and Weiming (2011), observed that in well-controlled situations, missing data always occur in longitudinal data analysis.
Missing data may degrade the performance of confidence intervals, reduce statistical power and bias parameter estimate. They
review and discuss general approaches for handling miss data in longitudinal studies. They started by first illustrating the mechanism
of missing data. Then focused on the methods for handling missing values in longitudinal data analysis. In the end, they summarized
and discussed the characteristics of each method
Howell (2000) observed that the treatment of missing data has been an issue in statistics for some time, but it has come to the fore
in recent years. He noted that the current interest in missing data stems mostly from the problems caused in surveys and census
data. He suggested several methods of treating missing values and elaborately gave some examples of how to deal with missingness.
Allison (2001) uses the example of ‘missingness’ for data on income being dependent on marital status. Perhaps unmarried couples
are less likely to report their income than married ones. Unmarried couples probably have lower incomes than married ones, and it
would at first appear that missingness on income is related to the value of income itself. But the data would still be MAR if the
conditional probability of missingness were unrelated to the value of income within each marital category. Here the real question is
whether the value of the dependent variable determines the probability that it will be reported, or whether there is some other
variable (X) where the probability of missingness on Y is conditional on the levels of X. To put it more formally, data are MAR if
p(Y missing |Y,X) = p(Y missing | X).
Dempster, Laird and Rubin (1977). Although, this method was considered to be more superior to the previous adhoc methods in
that they produced better estimates with smaller and acceptable standard errors. Finally, in the late 80’s, more superior methods

www.rsisinternational.org Page 64
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

such as Multiple Imputation, were developed. These methods were proved to be flexible and produced smaller standard errors as
compared to earlier methods.
Cohen and Cohen (1983), have suggested that when the missing data is on the dependent variable, the subject may be dropped from
the analysis. However, if the missing data is among the independent variables, it might be intrusive to determine what proportion
of the data is missing.
Truxillo (2005) has suggested that the EM covariance matrix and vector of means can also be used as input for procedures that
entail inference, but that one must then use a nominal sample size that properly accounts for the fact that some data are missing.
In Little’s test of MCAR (Little 1988), the data yi, (i = 1, 2,...,n) are modeled as p-dimensional multivariate normal with mean vector
μ and covariance matrix Σ, with part of the components in yi’s missing. When the normality is not satisfied, Little’s test still works
in the asymptotic sense for quantitative random vectors yi’s but is not suitable for categorical variables
De Silva et al (2007) observed that the presence of missing values in rainfall data is a common problem in the process of data
analysis. They suggested ways of treating missing data in temperature and rainfall data so as to obtain the best method.
III. Method
Data for this study were simulated from four different distributions. Two discrete and two continuous distributions were considered.
The distributions considered are Binomial, Geometric, Normal and Exponential distributions. The following specifications were
used
i. Binomial (p= 0.25, n=15, 25, 50, 100)
ii. Geometric (p = 0.5, n=15, 25, 50, 100)
iii. Normal (µ = 9.7, 𝜎 = 1.3, n=15, 25, 50, 100)
iv. Exponential (µ = 6.9, n=15, 25, 50, 100)
Level of missingness considered
The computations and analysis were done for distributions at various level of missingness. The level of missingness considered
varied from 5% to 35%. To be specific, the different level of missingness are; 5%, 10%, 25% and 35%.
Missing Data Method considered.
Two methods of handling missing data were considered in other to examine their effect on the parameters of a Multiple Regression
Model (MRM). The two methods considered are Mean Imputation and listwise deletion methods
Multiple Regression Model.
Regression Models are statistical procedures that deals with the nature of relationship between dependent variable(s) and
Independent Variable(s). Multiple regression model that deals with one dependent and more than one independent variables.
Model specification
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 . . . 𝛽𝑛 𝑥𝑛 + 𝑒𝑖 . . . . . (1)
Where
Y1 is the dependent variable
Β0 is the intercept
βi are the regression coefficients that represent the changes in y relative to a ne – unit change in xi’s
ei is the residual term.
The parameters of the model βi ae estimated using
𝛽 = (𝑋 ′ 𝑋)−1 (𝑋′𝑌) . . . . . . . . (2)
Where

www.rsisinternational.org Page 65
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023
𝑛 𝑛 𝑛

𝑛 ∑ 𝑋1 ∑ 𝑋2 . . . ∑ 𝑋𝑛
𝑖=1 𝑖=1 𝑖=1
𝑛 𝑛 𝑛 𝑛

∑ 𝑋1 ∑ 𝑋 21 ∑ 𝑋𝑖 𝑋2 . . . ∑ 𝑋1 𝑋𝑛
(𝑋 ′ 𝑋)−1
= 𝑖=1
. 𝑖=1
. 𝑖=1
. 𝑖=1
.
. . . ... .
. . . ... .
𝑛 𝑛 𝑛 𝑛
.. .. ..
∑ 𝑋𝑛 ∑ 𝑋𝑛 𝑋1 ∑ 𝑋𝑛 𝑋2 ∑ 𝑋2𝑛
[ 𝑖=1 𝑖=1 𝑖=1 𝑖=1 ]

∑𝑌
𝑖=1
𝑛

∑ 𝑋1 𝑌
(𝑋′𝑌) = 𝑖=1
..
.
𝑛

∑ 𝑋𝑛 𝑌
[ 𝑖=1 ]
Model Performance
The performances of the MR model were determined using Regression Analysis of Variance ANOVA given in Table 1.
Table 1: Regression ANOVA for model adequacy
Source of Variation Dere of Freedom Sum of Squares Mean Squares (MS) F-Ratio
Regression k-1 SSR 𝑆𝑆𝑅 𝑀𝑆𝑅
𝑀𝑆𝑅 = 𝐹 =
𝐾−1 𝑀𝑆𝐸
Error N-K SSE 𝑆𝑆𝐸
𝑀𝑆𝐸 =
𝑁−𝐾
Total N-1 SST

𝑆𝑢𝑚 𝑜𝑓 𝑆𝑞𝑢𝑎𝑟𝑒𝑠 𝑜𝑓 𝑅𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛 (𝑆𝑆𝑅) = 𝛽′𝑋′𝑌 − 𝑛𝑌̅ 2 . . . 3

𝑆𝑢𝑚 𝑜𝑓 𝑆𝑞𝑢𝑎𝑟𝑒𝑠 𝑜𝑓 𝐸𝑟𝑟𝑜𝑟 (𝑆𝑆𝐸) = 𝑌′𝑌 − 𝛽′𝑋′𝑌 . . . . 4
𝑆𝑢𝑚 𝑜𝑓 𝑆𝑞𝑢𝑎𝑟𝑒𝑠 𝑜𝑓 𝑇𝑜𝑡𝑎𝑙 (𝑆𝑆𝑇) = 𝑌′𝑌 − 𝑛𝑌̅ 2 . . . . . 5
Mean Sum of Squares Regression (MSR)

𝑆𝑆𝑅
𝑀𝑆𝑅 = . . . . . . .6
𝐾−1
Mean Sum of Squares Error (MSE)
𝑆𝑆𝐸
𝑀𝑆𝐸 = . . . . . . 7
𝑁−𝐾
IV. Results
Multiple regression models were fitted for the data simulated at different sample sizes. The F-Values, MSR, MSE and P-Values
where obtained. The model with the lowest MSE is considered as the best.

www.rsisinternational.org Page 66
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

Table 2: Estimation Results for Complete Data Sets by Distribution

Sample Distribution B0 B1 B2 F-Values MSR MSE P-
size Value
15 Binomial 7.69 -0.435 -0.073 0.62 2.9360 4.7107 0.553
25 4.71 -0.056 0.131 0.18 0.7702 4.2863 0.837
50 5.168 -0.067 -0.029 1.14 0.3655 2.5908 0.869
100 5.224 0.0038 -0.017 0.02 0.0643 3.6630 0.983

15 Geometric 1.59 0.741 -0.362 3.28 15.010 4.576 0.075

25 2.57 -0.133 0.092 0.37 1.7995 4.8437 0.694
50 2.887 0.016 -0.194 1.21 4.11227 3.40629 0.308
100 2.618 -0.014 -0.070 0.35 0.91717 2.60336 0.704

15 Normal 4.96 0.358 0.182 1.03 1.7354 1.6896 0.387

25 9.08 0.121 -0.042 0.17 0.21985 1.26686 0.842
50 9.12 0.101 -0.023 0.28 0.44113 1.56528 0.756
100 8.58 0.0238 0.083 0.38 0.56528 1.49900 0.687

15 Exponential 6.29 0.034 -0.105 0.10 59.08 58.573 0.905

25 8.05 -0.281 0.106 0.97 46.38 47.83 0.395
50 7.74 -0.184 -0.034 1.23 62.820 51.223 0.303
100 6.22 -0.128 0.0108 0.83 31.937 38.5476 0.440

For the Complete data set, the Parameters of the Multiple Regression were estimated for all the distributions and their MSE obtained.
The results showed that the Normal distribution at the sample size of 25 had the least MSE of 1.26686. It was also noticed that at
all sample sizes, the Normal distribution had the least MSE. This buttressed the fact that some assumptions guiding Multiple
regression has some link to Normal distribution.
Table 3: Estimation Results for Binomial Data Sets by degree of missingness and method of handling
Level Method Sample β0 β1 β2 F-Values MSR MSE P-Value
Size
5% Listwise 12 9.89 -0.775 -0.246 1.94 7.832 4.0373 0.199
22 3.99 -0.063 0.243 0.5 2.2406 4.4914 0.615
41 4.25 0.104 -0.029 0.23 0.58254 2.4892 0.792
85 4.995 0.0187 0.035 0.06 0.2167 3.8277 0.945
Mean 15 8.91 -0.617 -0.19 2.3 6.8138 3.3644 0.175
Imputation
25 4.47 -0.064 0.178 0.32 1.322 4.1916 0.733
50 4.518 0.068 -0.021 0.12 0.27552 2.26351 0.886
100 5.199 0.0052 -0.016 0.02 0.06112 3.60634 0.983
10% Listwise 10 7.01 -0.387 -0.0044 0.32 1.93746 6.01787 0.735
18 4.83 -0.095 0.167 0.25 0.913 3.6449 0.782
37 5.11 -0.079 -0.025 0.11 0.35105 3.0652 0.892
71 5.023 -0.018 0.056 0.14 0.58401 4.05987 0.866

www.rsisinternational.org Page 67
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

Mean 15 7.12 -0.378 0.064 0.53 2.1816 4.1364 0.603

Imputation
25 5.67 -0.154 0.054 0.29 1.003 3.4481 0.75
50 5.466 -0.082 -0.07 0.31 0.7397 2.378 0.734
100 5.152 -0.022 0.026 0.08 0.2577 3.2576 0.924
25% Listwise 3 7.8 0 -0.4 0 1.333 0
9 2.55 -0.273 1.182 1.23 7.2727 5.9091 0.357
17 7.65 -0.149 -0.389 0.8 2.7014 3.3956 0.417
34 4.75 0.128 0.034 0.38 1.4899 3.9162 0.687
Mean 15 7.59 -0.169 -0.241 0.62 1.7215 0.27737 0.554
Imputation
25 5.58 -0.26 0.168 0.81 2.865 3.519 0.456
50 4.94 0.006 -0.004 0 0.00205 1.74229 0.999
100 5.332 0.0039 -0.0011 0 0.00295 2.52558 0.999
35% Listwise 3 7.7 0.189 0.62 0.07 1.451 19.841 0.934
4 4.1 -0.15 0.4 1
13 7.61 -0.3 -0.143 0.92 1.5774 1.7153 0.43
22 5.4 -0.18 0.109 0.52 2.7248 5.2371 0.603
Mean 15 7.66 -0.432 0.101 0.41 1.0207 2.4632 0.67
Imputation
25 3.94 0.178 -0.026 0.21 0.51317 2.48744 0.815
50 5.12 -0.018 -0.037 0.05 0.07132 1.53353 0.955
100 5.233 -0.0583 0.059 0.44 1.173 2.6775 0.647

For the Binomial data set, the Parameters of the Multiple Regression were also estimated for all the distributions and their MSE
also obtained. It was observed that when listwise deletion was used, the least MSE of 1.7153 was obtained at 35% level of
Missingness. When mean imputation was used, the least MSE of 1.53353 was obtained at 35% level of Missingness. Overall, The
Mean Imputation method gave the least MSE of 1.53353 for the Binomial Distribution. This shows that missingness of data can
affect the significant of the model.
Table 4: Estimation Results for Geometric Data Sets by degree of missingness and method of handling
Level Method Sample β0 β1 β2 F-Values MSR MSE P-Value
Size
5% Listwise 12 1.25 0.845 -0.452 0.2 1.0964 5.4277 0.819
22 2.31 -0.071 0.1 0.2 1.0964 5.4277 0.819
41 2.269 -0.03 -0.089 0.76 0.9951 1.3135 0.476
86 2.669 -0.075 -0.053 0.35 0.7827 2.2473 0.707
Mean 15 1.75 0.77 -0.447 4.24 16.638 3.923 0.04
Imputation
25 2.31 -0.77 0.116 0.31 1.5317 4.8541 0.734
50 2.759 -0.084 -0.104 0.7 2.015 2.889 0.503
100 2.574 -0.078 -0.0687 0.41 0.86718 2.09346 0.662
10% Listwise 9 0.8 -0.0048 -1.082 6.35 24.971 3.935 0.033
16 3.6 1.488 0.142 1.33 7.289 5.494 0.299
36 2.871 0.025 -0.177 0.62 2.5702 4.117 0.542
73 2.603 -0.024 -0.062 0.2 0.58274 2.91877 0.819

www.rsisinternational.org Page 68
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

Mean 15 1.67 0.895 -0.0837 4.82 18.473 3.831 0.029

Imputation
25 2.8 -0.273 0.101 0.7 3.215 4.575 0.506
50 2.783 0.024 -0.145 0.65 2.1005 3.2137 0.525
100 2.446 0.036 -0.0562 0.31 0.7529 2.4133 0.733
25% Listwise 6 -3.18 0.529 4.31 24.66 29.615 1.2 0.014
8 6.43 -0.669 -0.038 0.78 9.2614 11.8704 0.507
20 2.904 -0.229 -0.167 0.84 1.74 2.034 0.448
39 2.205 0.06 -0.018 0.1 0.27252 2.66007 0.903
Mean 15 0.88 0.783 0.061 5.09 16.5673 3.2539 0.025
Imputation
25 2.44 -0.174 0.163 0.89 0.3927 4.399 0.424
50 3.126 -0.107 -0.152 0.85 2.599 3.065 0.435
100 2.177 0.105 -0.0081 0.51 1.1201 2.18309 0.6
35% Listwise 3 0 2 1 2.33
4 2.32 0.113 -0.03318 0.18 0.8784 4.9932 0.86
10 1.61 0.325 0.174 1.23 3.8943 3.1931 0.349
29 2.504 -0.127 -0.005 0.39 0.67428 1.72267 0.68
Mean 15 3.51 -0.153 0.484 0.24 1.183 4.8778 0.788
Imputation
25 1.79 -0.02 0.165 0.21 0.79883 3.85636 0.814
50 2.938 0.073 -0.214 1.67 4.365 2.6201 0.2
100 2.496 0.002 -0.06 0.18 0.32027 1.77833 0.835

For the Geometric data set, the Parameters of the Multiple Regression were also estimated for all the distributions and their MSE
also obtained. It was observed that when listwise deletion was used, the least MSE of 1.200 was obtained at 25% level of
Missingness. When mean imputation was used, the least MSE of 1.77833 was obtained at 35% level of Missingness. Overall, the
listwise gave the lowest MSE of 1.200 at the 25% level of missingness. This shows that missingness of data can affect the significant
of the model.
Table 5: Estimation Results for Normal Data Sets by degree of missingness and method of handling
Level Method Sample β0 β1 β2 F-Values MSR MSE P-Value
Size
5% Listwise 12 4.44 0.285 0.308 0.61 1.19 1.16 0.563
21 9.16 0.231 -0.154 0.42 0.5923 1.4129 0.663
41 9.85 0.051 -0.049 0.12 0.1758 1.4358 0.885
87 8.75 -0.0034 0.095 0.41 0.61767 1.51295 0.666
Mean 15 7.22 0.225 0.089 0.4 0.706 1.764 0.679
Imputation
25 9.34 0.175 -0.122 0.31 0.3816 1.2509 0.74
50 9.73 0.05 -0.037 0.11 0.1675 1.5335 0.897
100 8.59 0.0137 0.09 0.41 0.60858 1.47203 0.663
10% Listwise 9 4.69 0.54 0.052 0.75 1.55829 2.08069 0.512
16 10.28 0.177 -0.223 0.27 0.4011 1.5009 0.77
35 9.86 0.132 -0.134 0.82 1.3109 1.6003 0.45

www.rsisinternational.org Page 69
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

71 9.15 -0.013 0.061 0.14 0.21896 1.54941 0.868

Mean 15 5.79 0.475 -0.02 1.73 210286 1.21738 0.0219
Imputation
25 9.61 0.167 -0.154 0.38 0.4456 1.1688 0.687
50 9.4 0.128 -0.086 0.65 0.9669 1.4905 0.527
100 8.98 0.0079 0.0571 0.17 0.22786 1.30281 0.84
25% Listwise 3 -13.3 1.949 0.1591 2.82954
10 12.2 -0.09 -0.142 0.26 0.18536 0.71218 0.778
17 12.5 -0.236 -0.009 0.64 0.95935 1.50272 0.543
45 8 -0.186 0.356 2.67 4.082 1529 0.081
Mean 15 0.31 0.36 0.642 1.71 2.448 1.429 0.221
Imputation
25 9.1 0.16 -0.089 0.43 0.4143 0.9717 0.658
50 11.04 -0.158 0.043 0.66 0.7706 1.1663 0.521
100 9.23 -0.063 0.091 0.48 0.551 1.1649 0.622
35% Listwise 3 0.7156 0.9261 0.1097 1.20688
4 2.7 0.2066 0.603 14.37 1.9664 0.1369 0.183
11 13.79 -0.097 -0.383 2.53 1.0266 0.4375 0.158
22 9.87 0.077 -0.098 0.23 0.2625 1.1521 0.798
Mean 15 9.64 0.088 -0.05 0.11 0.17326 1.63866 0.9
Imputation
25 8.81 -0.052 0.181 0.94 0.47171 0.5041 0.407
50 7.59 0.228 -0.019 1.33 1.25975 0.94411 0.273
100 9.81 0.0227 -0.037 0.1 0.09002 0.90819 0.906

For the Normal data set, the Parameters of the Multiple Regression were also estimated for all the distributions and their MSE also
obtained. It was observed that when listwise deletion was used, the least MSE of 0.1365 was obtained at 35% level of Missingness.
When mean imputation was used, the least MSE of 0.5041 was obtained at 35% level of Missingness. Overall, the listwise gave the
lowest MSE of 0.1365 at the 35% level of missingness. This shows that missingness of data can affect the significant of the model.
Table 6: Estimation Results for Exponential Data Sets by degree of missingness and method of handling.
Level Method Sample β0 β1 β2 F-Values MSR MSE P-Value
Size
5% Listwise 12 6.38 0.021 -0.079 0.04 3.791 76.8096 0.961
22 8.7 -0.298 0.128 1.04 51.14 50.09 0.372
41 8.32 -0.201 -0.107 1.17 67.21 57.56 0.322
85 6.22 -0.124 0.0131 0.63 26.1984 41.9085 0.538
Mean 15 6.26 0.019 -0.091 0.07 4.375 58.6456 0.929
Imputation
25 8.24 -0.298 0.14 1.25 56.3 45.2 0.307
50 7.69 -0.184 -0.08 1.14 56.39 49.55 0.327
100 6.47 -0.142 0.0022 0.95 36.0933 37.9739 0.39
10% Listwise 9 8.41 -0.039 -0.336 0.17 17.183 100.461 0.847
17 6.78 -0.446 0.456 3.7 161.7 43.7 0.051
35 7.15 -0.153 0.001 0.78 32.3982 41.3921 0.466
73 6.51 -0.037 -0.104 0.52 19.1319 36.9509 0.598

www.rsisinternational.org Page 70
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

Mean 15 6.94 -0.085 -0.058 0.12 6.692 55.317 0.887

Imputation
25 7.67 -0.336 0.183 2.02 86.09 42.62 0.157
50 7.74 -0.175 -0.007 1.08 54.549 50.489 0.349
100 5.72 -0.013 0.0065 0.01 0.4009 37.4961 0.989
25% Listwise 5 63.1 -3.939 -12.27 15.75 252.56 16.03 0.06
11 8.37 -0.566 0.501 2.58 153.22 59.4 0.137
20 8.49 -0.229 0.041 1.4 69.655 49.843 0.274
10 11.15 -0.43 -0.207 8.34 33.725 4.043 0.014
Mean 15 6.04 0.204 -0.173 0.37 19.67 52.89 0.697
Imputation
25 6.87 -0.441 0.434 4.31 144.74 33.61 0.026
50 8.96 -0.195 -0.011 1.35 60.276 44.658 0.269
100 5.594 -0.1257 0.0686 1.81 28.649 15.794 0.168
35% Listwise 5 1.52 -0.1363 0.904 9.3 28.838 3.102 0.097
6 16.68 -1.49 0.211 2.65 65.45 24.66 0.217
8 19.4 -1.316 -0.22 0.152 160.67 105.68 0.305
27 9.24 -0.078 -0.292 0.043 36.145 84.795 0.658
Mean 15 7.13 -0.163 0.233 0.29 14.51 49.27 0.75
Imputation
25 9.38 -0.367 -0.002 0.41 12.9954 31.3236 0.665
50 8.86 -0.211 -0.065 2.03 69.05 .33.938 0.142
100 7.38 -0.1544 -0.1379 2.27 64.521 24.442 0.109

For the Exponential data set, the Parameters of the Multiple Regression were also estimated for all the distributions and their MSE
also obtained. It was observed that when listwise deletion was used, the least MSE of 3.102 was obtained at 35% level of
Missingness. When mean imputation was used, the least MSE of 15.794 was obtained at 25% level of Missingness. Overall, the
listwise gave the lowest MSE of 3.102 at the 35% level of missingness. This shows that missingness of data can affect the significant
of the model.
The summary of the results can be seen in Table 7.
Table 7: Summary of Results
Least-Final
SN Distribution Level Method Least MSE Least-best MSE
MSE
1 Complete - - 1.26686 1.26686 1.26686
2 Binomial Listwise 1.7153
3 Binomial Mean Imputation 1.53353 1.53353
4 Geometric Listwise 1.2000 1.2000
5 Geometric Mean Imputation 1.77833
6 Normal Listwise 0.1365
7 Normal Mean Imputation 0.5041 0.5041 0.5041
8 Exponential Listwise 3.102 3.102
9 Exponential Mean Imputation 15.794

www.rsisinternational.org Page 71
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

The results in Table 7 shows that when Binomial and Geometric distributions were compared, Binomial had a Least MSE of 1.53353
under mean imputation as against Geometric distribution with MSE of 1.77833. But when Listwise deletion method was used,
Geometric distribution had the least MSE of 1.2000 as against Binomial MSE of 1.7153.
Comparing Binomial and Normal distributions, Normal had a Least MSE of 0.5041 under mean imputation as against Binomial
distribution with MSE of 1.53353. Also, when Listwise deletion method was used, Normal distribution had the least MSE of 0.1365
as against Binomial MSE of 1.7153.
When Binomial and Exponential distributions were compared, Binomial had a Least MSE of 1.53353 under mean imputation as
against Exponential distribution with MSE of 15.794. While under Listwise deletion method, Binomial distribution had the least
MSE of 1.7153 as against Exponential distribution with MSE of 3.102.
Comparing Geometric and Normal distributions, Normal had a Least MSE of 0.5041 under mean imputation as against Geometric
distribution with MSE of 1.77833. Also when Listwise deletion method was used, Normal distribution had the least MSE of 0.1365
as against Geometric with MSE of 1.2000.
When Geometric and Exponential distributions were compared, Geometric distribution had a Least MSE of 1.77833 under mean
imputation as against Exponential distribution with MSE of 15.794. While under Listwise deletion method, Geometric distribution
had the least MSE of 1.2000 as against Exponential distribution with MSE of 3.102.
Comparing Exponential and Normal distributions, Normal had a Least MSE of 0.5041 under mean imputation as against
Exponential distribution with MSE of 15.794. Also when Listwise deletion method was used, Normal distribution had the least
MSE of 0.1365 as against Exponential distribution with MSE of 3.102.
V. Conclusion
The results of the analysis shows that missing data significantly affect the parameters as well as the model significance for various
distributions. More specifically, for the complete data set, the results showed that the Normal distribution at the sample size of 25
had the least MSE of 1.26686. It was also noticed that at all sample sizes, the Normal distribution had the least MSE among the
four distributions considered. Exploring the results of the Binomial Distribution, it was observed that when listwise deletion was
used, the least MSE of 1.7153 was obtained at 35% level of Missingness. When mean imputation was used, the least MSE of
1.53353 was obtained at 35% level of Missingness. Overall, The Mean Imputation method gave the least MSE of 1.53353 for the
Binomial Distribution.
For the Normal Distribution, it was observed that when listwise deletion was used, the least MSE of 0.1365 was obtained at 35%
level of Missingness. When mean imputation was used, the least MSE of 0.5041 was obtained at 35% level of Missingness. Overall,
the listwise gave the lowest MSE of 0.1365 at the 35% level of missingness. For the Exponential Distribution, it was observed that
when listwise deletion was used, the least MSE of 3.102 was obtained at 35% level of Missingness. When mean imputation was
used, the least MSE of 15.794 was obtained at 25% level of Missingness. Overall, the listwise gave the lowest MSE of 3.102 at the
35% level of missingness.
From the varying values of least MSE’s and the sample size effects, it is very important to handle missing values with care and to
adhere to the assumptions (if any) on the use of any method of estimations of model parameters so as to obtain consistent and
unbiased results.
Reference
1. Allison, P (2001) Missing Data: Quantitative Application in the Social Science. Thousand Oaks, CA: Sage. Vol. 136.
2. Brigga A., Clark T, Wolstenholme, J, Clarke P (2003) Missing Data. Health Economics 12, 377-392
3. Catherine Truxillo (2005) A Comparison of Missing Data Handling Methods. SAS® Institute Inc, Cary, NC.
4. Cohen J and Cohen P (1975): Applied Multiple Regression and Correlation Analysis for the Behavioural Sciences. New
York : John Wiley
5. Dempster A. P, Laird, N. M and Rubin D. B. (1977): Maximum Likelihood Estimation from Incomplete data via the EM
algorithm. Journal of the Royal Statistical Association B39, 1-38
6. Graham, J. W (2009) Missing data analysis: making it work in the real world. Annu Rev Psychol 60, 549 – 576.
7. Jones, M. P (1996) Indicator and Stratification methods for missing explanatory variables in Multiple Linear Regression.
Journal of the American Statistical Association, 91, 222-230
8. Kajornrit, Jesada; Wong, Kok Wai; Fung, Chun Che (2012) A comparative Analysis of soft computing techniques used to
estimate missing precipitation records. 9th Biennial Conference of the International Telecommunications Society (ITS):
Bangkok, Thailand.

www.rsisinternational.org Page 72
 INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 231-2705 | DOI: 10.51244/IJRSI | Volume X Issue II February 2023

9. Little, R. J. A (1988). A test of missing completely at Random or multivariate data with missing values. Journal of the
American Statistical Association 83, 1198 – 1202.
10. Little, R. J. A (1992) Regression with Missing X’s: A review. Journal of the American Statistical Association 87, 1227 –
1237.
11. Nakai M and Weiming Ke (2011) Review of methods for handling Missing Data in Longitudinal Data analysis. Int journal
of Math. Analysis. Vol 5, no 1. 1-13.
12. De Silva R. P., Dayawansa and Ratnasiri (2007). A comparison of method used n Estimating missing Rainfall data. The
Journal of Agricultural Sciences, 2007, Vol 3. No 2.
13. Rubin D. B, (1976) Inference and missing data. Biometrika, 63, 581 592
14. Rubin D B (1987). Multiple Imputation for Nonresponse in Surveys. New York. John Wiley and Sons.
15. SAS Institute, 2005. Multiple Imputation for missing data. Concepts and New Approaches. A Useful Overview of the
different methods to deal with Missing Data Using SAS
16. Schmitt p, Mandel J, Guedj M (2015) A Comparison of Six Methods of missng data Imputation. J. Biomet Biostat 6: 24
doi:10.4172/2155-6180.1000224
17. Schafer, J. L (197). Analaysis of incomplete Multivariate data. New York: Chapman & Hall
18. Sunbul Secil Omur(2018) The Impact of Diffrnt Missing Data handling Methods on DINA Model. Internatiopnal Journal
of Evaluation and Research in education (IJER) Vol 7. No.1. pp 77 – 86 ISSN: 2252 – 8822
19. Therese D. pigott (2001) A Review of methods of Missing Data. Educational research and evaluation, Vol 7. No 4 pp. 353
– 388.

www.rsisinternational.org Page 73