Available online at www.sciencedirect.

com

Safety Science 47 (2009) 88–96

www.elsevier.com/locate/ssci

Logistic regression model for prediction of roof fall risks in bord

and pillar workings in coal mines: An approach
Sanjay Kumar Palei *, Samir Kumar Das
Department of Mining Engineering, Indian Institute of Technology, Kharagpur 721 302, India

Received 18 September 2007; received in revised form 18 December 2007; accepted 14 January 2008

Abstract

The roof fall hazards are the major problems in underground coal mines, which are generally unpredictable due to the associated uncer-
tainties arising out of the complexity of geological conditions and variability in mining parameters. During the six year period, 1996–2001,
253 Indian coal miners lost their lives and 401 are seriously injured in 490 different roof fall accidents. In India, 32.7% of the total fatal
injuries in coal sectors are due to roof fall in bord and pillar method of workings. This paper attempts to predict the severities of roof fall
accidents based on some major contributing parameters using the binary logistic regression model. In total, 128 roof fall accidents for the
last few years from five underground coal mines in India having bord and pillar method of working are analyzed for this study. The results
revealed that wider gallery width is more prone to major and serious accidents than narrower gallery width. Thin seams are more amenable
to major accidents in comparison to thick seams. Unsupported or partially supported roofs have higher risk for contributing major as well
as serious accidents. Deep workings are more prone to major accidents as compared to shallow depth workings.
Ó 2008 Elsevier Ltd. All rights reserved.

Keywords: Accident; Bord and pillar; Logistic regression; Risk; Roof fall

1. Introduction
Roof falls continue to be one of the greatest single haz-
ards faced by underground coal miners. The hazardous
nature of roof falls in underground coal mine operations
can be illustrated from the national statistics of mine acci-
dents. During the six year period, 1996–2001, in India, 253
coal miners lost their lives and 401 are seriously injured in
490 different roof fall accidents. 32.7% of the total fatal
injuries in Indian coal sectors are due to roof fall in bord
and pillar method of working; whereas only 1% of the total
fatal injuries due to roof fall are occurring in longwall
method of working (DGMS, 1996–2001). The roof fall
accident rates are less in longwall faces, since they are very
few in number, mostly mechanized covering the working
*
Corresponding author. Present address: Department of Mining Engi-
neering, Institute of Technology, Banaras Hindu University, Varanasi 221
005, Uttar Pradesh, India. Tel./fax: +91 542 2369442.
E-mail address: sanjay_palei@rediffmail.com (S.K. Palei).
faces with powered roof supports and galleries with the
combination of roof bolting, props and bars, and also areas
of exposure due to drivage of galleries are much smaller
than the bord and pillar method of working (Das, 1984).
In underground coal mining, bord and pillar is the method
of working, mostly preferable for flat tabular deposits in
thin seams, where bords or galleries are driven in the solid
coal to form pillars in the development workings. ‘Pillars’
of coal are left behind to support the roof and prevent col-
lapse. Additional support is provided through the use of
roof bolts and props. In some cases, the pillars are removed
partly or fully in a later operation known as bord and pillar
depillaring.
The complexity of geological deposit and variability of
mining parameters leads to the occurrences of unwanted
roof falls. There are various studies on finding relationship
between the roof falls and geological conditions, stress
state and mine layout (Molinda et al., 2000; Das, 2000;
Phillipson, 2003). Phillipson (2003) emphasized on geolog-
ical mapping for ultimate benefit of understanding the

0925-7535/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ssci.2008.01.002
 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96
geology and projecting the zones of adverse ground condi-
tions for the prevention of fatality. The identification of
many slickenside features as faults and shear zones, which
are true structural geologic features, increases the impor-
tance of basic structural geologic mapping in coal mines.
Interpreting slickenside formation in terms of structural
geology can allow areas of poor ground conditions to be
projected ahead of mining, so that additional support mea-
sures can be planned in advance. Das (2000) studied the
behaviour of coal measure roof rocks during mining, con-
sidering the geology and physico-mechanical rock proper-
ties. Molinda et al. (2000) performed simple regression
analysis with some significant geotechnical variables like
overburden, bolt strength, bolt capacity, grout length, den-
sity, entry width, coal mine roof rating (CMRR) and inter-
section span for predicting the roof fall rate. The results
showed that there is relationship existing between intersec-
tion span and entry width; CMRR and overburden;
CMRR and bolt length. For stronger roofs (high CMRR),
shorter bolts are required, and there is a positive correla-
tion between CMRR and depth of cover. Depth of cover
is one of the major factors for the ground falls, as it is an
indirect representation of the horizontal stress (Dolinar
et al., 2001; Molinda et al., 2000) affecting the roof geology
and the pillar strength. In most of the countries, mine
workings having depths less than, or equal to 120 m are
regarded as shallow depth category and more than
120 m, as deep mines (Dolinar et al., 2001). In their study,
for the guidelines of roof bolt design, though they found
statistically, the effect of depth of cover was relatively weak
with all other geotechnical variables remaining the same,
stated deeper mines are more likely to have high roof fall
rates. Besides depth of cover, another important factor is
the width of gallery. The relationship of gallery width with
factor of safety of pillars has been well established and the
immediate roof rocks may be classified as weak, or strong
(Das, 1994). The management of roof fall risks in under-
ground coal mines is explained by Duzgun and Einstein
(2004) with the probability of roof fall, and cost of conse-
quences by relative cost criterion. They provided two alter-
native methods of ‘‘do nothing” and go for ‘‘support
improvement” by the cost-benefit analysis in their decision
framework. A comparative study is conducted for a group
of five underground coal mines in India for the reliability of
mine for roof fall accidents and estimation of roof fall risks
by Monte Carlo method (Palei and Das, 2005). Their
results revealed that the probability of roof fall of a mine
did not identify the overall risk, because the other compo-
nent of risk, i.e. the costs of consequences of such roof fall
accidents also play an important role.
Accident ratio (i.e. reportable injury:serious
injury:fatality) studies indicated that there are many more
minor injuries experienced than serious injuries and that
for the same serious injury there have been numerous
property-damage accidents and there are a large number
of near-miss cases too. Bhattacherjee (2000) identified that
for a group of six Indian coal mines, the four year ratio
89
of reportable injury to serious injury to fatality was
50:5:1. A special attention is paid in this paper on the
assessment of the risk of roof fall accidents of different
severity in the coal mines having bord and pillar method
of workings. The major contributing parameters may be
categorized into different groups, which are responsible
for unwanted outcomes. The contributing parameters
has been grouped into two classes namely (i) geo-mining
variables, such as gallery width, mining height, depth of
cover, seam thickness, roof support status, and immediate
roof and (ii) location variables like roof fall location and
the individual mine. A binary logistic regression model is
used for this study.
2. Case study
2.1. Mine description
Five underground coal mines are considered for the pur-
pose of study and data are collected from those mines. All
the mines have bord and pillar method of workings. Out of
the five mines, three are located in Eastern India (relevant
data collected from these mines from 1991 to 2004) and
two are in Southern India (relevant data collected from
these mines from 1996 to 2004). Each mine operates six
days a week and three shifts a day. The roof supporting
is done by systematic support rules (SSR) as specified by
the Directorate General of Mines Safety. The roof supports
used are mainly props, wooden cogs, and steel cogs rein-
forced with roof bolting and stitching. The average overall
output per man-shift (OMS) for the case study mines dur-
ing those period are 0.92, 0.51, 0.57, 0.93 and 0.77 tonnes
for Mines 1–5, respectively.
2.2. Data and analysis
Since very few fatal incidents (11 in total) occurred dur-
ing the aforementioned study period in the specified mines,
the fatal and serious accidents are merged into a single cat-
egory namely major accident. The roof fall accidents
involving one or more reportable injuries are encoded here
as minor accidents. The dependent variable for this study is
the degrees of roof fall accident, and is categorized as
major, serious, and minor. The variable degree of accident
is important as it measures the injury severity, which is also
an indirect measure of the vulnerability, briefly the cost of
an accident. A more severe injury involves higher cost of
operation compared to a minor injury. This is due to pay-
ment of workers’ compensation, medical treatment, down-
times, machinery breakdown, reduced production rates,
and social costs. The terms like major accident, serious
accident and minor accident used in the logistic regression
model may be defined as mentioned below:
Major accident: Major accident may be defined as the
accident involving at least either a fatal injury, or at least
one serious bodily injury, which increases the cost of
consequence.
90 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96

Serious accident: The accident involving one, or more Table 2

serious bodily injuries, and may contain reportable injuries; Description and categorization of independent variables used in logistic
regression analysis
and not involving fatal injuries.
Minor accident: Accidents involving one or more report- Variables and category description Indicator co-variates used in logistic
regression
able injuries only, and not involving fatal or serious inju-
ries; and the expected cost of consequence is very less in Depth of cover (m) X1
DEPTH1 6 120a 0
compared to major or serious accidents. DEPTH2 > 120 1
Based on the categories of the severity of injury, two
Gallery width (m) X1
cases A and B are considered. Cases A and B dealt with
WIDTH1 < 4.2a 0
major versus minor accidents, and serious versus minor WIDTH2 P 4.2 1
accidents, respectively. The occurrences of the roof fall
Seam thickness (m) X1
accidents for both the cases (cases A and B) in different cat- SEAMTH1 Less thick 6 4.8 1
egories are presented in Table 1 for the case study mines. SEAMTH2 Thick seams > 4.8a 0
Immediate roof X1
2.3. The logistic regression model IMMRF1 Weaka 0
IMMRF2 Strong 1
The logistic regression model assumes that a roof fall
Mining height (m) X1
will have a degree of accident during the working with a HEIGHT1 6 2.5a 0
probability P (0 < P < 1). In order to understand the mod- HEIGHT2 > 2.5 1
eling process, a binary variable is used. A binary variable Z Roof support status X1
is defined as follows: Z = 1, provided there is a roof fall SUPPST1 Supporteda 0
that resulted at least a severe accident; Z = 0, provided SUPPST2 Unsupported 1
there is a roof fall causing a minor accident. Location X1
Two binary variables for the degree of accident are gen- LOCAT1 Facea 0
erated in this model. First, the degree of accident is coded LOCAT2 Outbye 1
as ‘1’ (Z = 1) for all major roof fall accidents, and ‘0’ Mine variables X1 X2 X3 X4
(Z = 0) for minor roof fall accidents (case A). Similarly, MINE1 Mine 1 1 0 0 0
the variable is coded as ‘1’ for all serious roof fall accidents, MINE2 Mine 2 0 1 0 0
and ‘0’ for minor roof fall accidents (case B). MINE3 Mine 3 0 0 1 0
MINE4 Mine 4 0 0 0 1
The independent variables are width of gallery, mining MINE5 Mine 5a 0 0 0 0
height, depth of cover, seam thickness, roof support status, a
Reference category.
immediate roof, face and specific mine, which are coded as

dummy variables. A description of the categories of these

Table 1 variables with indicator coding scheme is presented in
Description of roof fall accidents for the two cases of the case study mines
Table 2.
Mine Id Case A Case B The dichotomous nature of the dependent variable
Major Minor Total Serious Minor Total directs the logistic regression model to be used for this
Mine 1 2 18 20 2 18 20 analysis. The logistic regression analysis presents a unique
Mine 2 8 15 23 4 15 19 compliment to multivariate regression in its ability to uti-
Mine 3 11 12 23 9 12 21 lize binary dependent variable. Following the established
Mine 4 5 34 39 3 34 37
Mine 5 9 14 23 6 14 20
coding scheme (Table 2) of the mine variables, the logistic
model for all the four dummy variables is specified as
Total 35 93 128 24 93 117
follows:

Table 3
Spearman’s correlations for major versus minor accidents (case A)
Depth Gallery width Seam thickness Immediate roof Height Face Support status Mine
Depth 1.000
Gallery width 0.253* 1.000
Seam thickness 0.597* 0.388* 1.000
Immediate roof 0.011 0.126 0.063 1.000
Height 0.138 0.100 0.152 0.114 1.000
Face 0.022 0.071 0.035 0.084 0.016 1.000
Support status 0.073 .060 0.014 0.009 0.107 0.219 1.000
Mine 0.623* 0.301* 0.607* 0.601 0.017 0.037 0.027 1.000
*
Correlation is significant at the 0.01 level (2-tailed).
 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96 91

Table 4
Spearman’s correlations for serious versus minor accidents (case B)
Depth Gallery width Seam thickness Immediate roof Height Face Support status Mine
Depth 1.000
Gallery width 0.276* 1.000
Seam thickness 0.603* 0.420* 1.000
Immediate roof 0.005 0.134 0.042 1.000
Height 0.109 0.123 0.164 0.131 1.000
Face 0.033 0.065 0.011 0.104 0.000 1.000
Support status 0.122 0.092 0.060 0.075 0.107 0.227 1.000
Mine 0.662* 0.312* 0.642* 0.086 0.020 0.065 0.025 1.000
*
Correlation is significant at the 0.01 level (2-tailed).

P ðX 1 ; X 2 ; X 3 ; X 4 Þ ¼ 1=ð1 þ exp ðb0 þ b1 X 1 þ b2 X 2
þ b3 X 3 þ b4 X 4 ÞÞ
where P(Xl, X2, X3, X4), probability of roof fall accidents
of the mine categories; Xl, X2, X3, and X4 dummy variates
for the mine categories; and b1, b2, b3, and b4 for corre-
sponding parameters of Xl, X2, X3, and X4.
The b parameters of the logistic regression model are
estimated by maximum likelihood method suggested by
Cox (Cox, 1970). The test that the b coefficients are zero
or not is governed by Wald Statistics, which follows a
chi-square distribution (SPSS 10). Similarly, for the other
variables, the probability of roof fall accidents of the
respective categories of parameters can be computed using
the binary logistic regression model.

Table 5
Logistic regression model coefficients, significance level, and adjusted odds ratio for the specific cases
Variables and category description Case specifications for roof fall accidents
(Major versus minor accidents) (Serious versus minor accidents)
b Significance level exp(b) b Significance level exp(b)
Depth of cover (m)
DEPTH1 6 120a (0) (1) (0) (1)
DEPTH2 > 120 1.721 0.047 5.6* 1.959 0.061 7.1
Gallery width (m)
WIDTH1 < 4.2a (0) (1) (0) (1)
WIDTH2 P 4.2 1.504 0.017 4.5* 1.908 0.013 6.7*
Seam thickness (m)
SEAMTH1 6 4.8 2.062 0.029 7.9* 1.409 0.152 4.1
SEAMTH2 > 4.8a (0) (1) (0) (1)
Immediate roof
IMMRF1 Weaka (0) (1) (0) (1)
IMMRF2 Strong 0.231 0.644 0.8 0.501 0.403 0.6
Mining height (m)
HEIGHT1 6 2.5a (0) (1) (0) (1)
HEIGHT2 > 2.5 0.032 0.957 1.0 0.166 0.812 1.2
Roof support status
SUPPST1 Supporteda (0) (1) (0) (1)
SUPPST2 Unsupported 1.181 0.026 3.3* 1.317 0.048 3.7*
Roof fall location
LOCAT1 Facea (0) (1) (0) (1)
LOCAT2 Outbye 0.479 0.368 0.6 0.202 0.743 0.8
Mine variables
MINE1 Mine 1 2.378 0.050 0.09 1.152 0.420 0.32
MINE2 Mine 2 2.429 0.026 0.09* 2.362 0.075 0.09
MINE3 Mine 3 0.856 0.378 0.43 0.032 0.978 0.97
MINE4 Mine 4 1.789 0.030 0.17* 1.976 0.049 0.14*
MINE5 Mine 5a (0) 0.051 (1) (0) 0.075 (1)
Constant 2.763 0.008 0.06* 3.664 0.003 0.03*
*
Results are significant at 0.05 probability level.
a
Reference category.
92 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96

3. Results accidents (case B) with 117 observations are tabulated

In this section, the correlation between the indepen-
dent variables are tested and the results obtained from
the logistic regression model are explained. The correla-
tion coefficients between the independent variables for
case A and case B are presented in Tables 3 and 4,
respectively. The logistic model is utilised in establishing
the significant differences between categories of the vari-
ous variables.
3.1. Correlation between independent variables
Though the logistic regression model assumes the inde-
pendent variables are not dependent on each other, every
time, it is difficult to find such a model. The Spearman’s
correlations for major versus minor accidents (case A)
with 128 observations are tabulated in Table 3; whereas
Spearman’s correlations for only serious versus minor
in Table 4. From Tables 3 and 4, the Spearman’s correla-
tion coefficients are found significant in some of the cases,
at 0.01 significance level and are marked with an asterisk.
The discussions below for the correlation coefficients are
equally valid for both the cases (cases A and B), as the
direction of the coefficients are the same for both and
the correlations are significant for the same variables.
The gallery width and depth of cover are found to have
a negative correlation; this may be due to the fact that
at greater depths, generally lower gallery widths are pre-
ferred because of the higher concentration of in-situ stres-
ses around the openings. The seam thickness is positively
correlated with depth of cover, i.e. thickness of the seam
of the deposits are more at greater depths. Again, seam
thickness is found to be negatively correlated with gallery
width. The mine variable is found to be positively corre-
lated with depth of cover and seam thickness; and nega-
tively correlated with the width of gallery.

Table 6
Comparison of original model and nested model after dropping the variable ‘Depth of cover’ for the major versus minor accidents (case A)
Variables and category description Major versus minor roof fall accidents
Original model Nested model
Significance level exp(b) [95% CI] Significance level exp(b) [95% CI]
Depth of cover (m)
DEPTH1 6 120a (1) EXM
DEPTH2 > 120 0.047 5.6* [1.0–30.5] EXM
Gallery width (m)
WIDTH1 < 4.2a (1) (1)
WIDTH2 P 4.2 0.017 4.5* [1.3–5.5] 0.011 4.8* [1.4 -16.5]
Seam thickness (m)
SEAMTH1 6 4.8 0.029 7.9* [1.2–50.2] 0.144 3.1 [0.7–3.7]
SEAMTH2 > 4.8a (1) (1)
Immediate roof
IMMRF1 Weaka (1) (1)
IMMRF2 Strong 0.644 0.8 [0.3–2.1] 0.671 0.8 [0.3–2.1]
Mining height (m)
HEIGHT1 6 2.5a (1) (1)
HEIGHT2 > 2.5 0.957 1.0 [0.3–3.3] 0.617 1.3 [0.4–4.1]
Roof support status
SUPPST1 Supporteda (1) (1)
SUPPST2 Unsupported 0.026 3.3* [1.1–9.2] 0.018 3.4* [1.2–9.5]
Roof fall location
LOCAT1 Facea (1) (1)
LOCAT2 Outbye 0.368 0.6 [0.2–1.8] 0.378 0.6 [0.2–1.7]
Mine variables
MINE1 Mine 1 0.050 0.09 [0.01–1.00] 0.003 0.04 [0.01–0.31]
MINE2 Mine 2 0.026 0.09* [0.01–0.75] 0.005 0.05* [0.01–0.42]
MINE3 Mine 3 0.378 0.43 [0.06–2.85] 0.099 0.22 [0.04–1.33]
MINE4 Mine 4 0.030 0.17* [0.03–0.84] 0.010 0.12* [0.03–0.61]
MINE5 Mine 5a 0.051 (1) 0.009 (1)
Constant 0.008 0.06* 0.996 1.00
EXM: Excluded from the original model.
*
Results are significant at 0.05 probability level.
a
Reference category.
 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96 93

3.2. Logistic model results
The SPSS package is used for the logistic model runs.
The relationships of degree of roof fall accidents with the
contributing parameters are tabulated in Table 5 with the
logistic model coefficients (b) for two case specifications
(cases A and B). Cases A and B represent the ‘major versus
minor roof fall accident’ and ‘serious versus minor roof fall
accident’, respectively, taking the major contributory fac-
tors like geotechnical and roof fall location variables of
the case study mines. An asterisk in Table 5 indicates the
significant parameters at 0.05 probability level of
significance.
The major geotechnical and location variables are
included in the model. Though other parameters like rock
mass rating (RMR), bolt length, method of working may
have some effects, these parameters are not included in
the model. It is very difficult to get the RMR value at
each accident place, since its true value is rarely measured
at accident places. The occupational demographic charac-
teristics like age and experience, which may have some
effect for these roof fall accidents, are not considered
due to the limitations of the model. However, the mean
age of the workers affected by the injury are 40.4, 39.0,
36.6, 42.3, and 42.4; and the mean experience for the
injured workers are 14.4, 14.5, 11.8, 17.8 and 17.3 from
Mine 1 to Mine 5, respectively. Regular safety trainings
are provided to control the accidents and safety weeks
are organised for the awareness of workers by the mine
management.
Amongst the mine categories, there are significant differ-
ences in the occurrence of severity of roof fall accidents for
the cases A and B. The risk factors for the various catego-
ries of geo-mining parameters and location variables are
shown in Table 5.
From Tables 3 and 4, it is clear that there may be some
intermediary effects of the variables, as the correlation is
existing between the independent variables, which may

Table 7
Comparison of original model and nested model after dropping the variable ‘Depth of cover’ for the serious versus minor accidents (case B)
Variables and category description Serious versus minor roof fall accidents
Original model Nested model
Significance level exp(b) Significance level exp(b) [95% CI]
Depth of cover (m)
DEPTH1 6 120a (1) EXM
DEPTH2 > 120 0.061 7.1 [0.9–55.1] EXM
Gallery width (m)
WIDTH1 < 4.2a (1) (1)
WIDTH2 P 4.2 0.013 6.7* [1.5–30.1] 0.007 7.7* [1.7–34.5]
Seam thickness (m)
SEAMTH1 6 4.8 0.152 4.1 [0.6–28.1] 0.508 1.7 [0.3–8.0]
SEAMTH2 > 4.8a (1) (1)
Immediate roof
IMMRF1 Weaka (1) (1)
IMMRF2 Strong 0.403 0.6 [0.2–1.9] 0.480 0.7 [0.2–2.1]
Mining height (m)
HEIGHT1 6 2.5a (1) (1)
HEIGHT2 > 2.5 0.812 1.2 [0.3–4.6] 0.533 1.5 [0.4–5.4]
Roof support status
SUPPST1 Supporteda (1) (1)
SUPPST2 Unsupported 0.048 3.7* [1.0–13.7] 0.032 4.0* [1.1–14.2]
Roof fall location
LOCAT1 Facea (1) (1)
LOCAT2 Outbye 0.743 0.8 [0.2–2.7] 0.781 0.8 [0.27–2.7]
Mine variables
MINE1 Mine 1 0.420 0.32 [0.02–5.2] 0.023 0.07* [0.01 -0.70]
MINE2 Mine 2 0.075 0.09 [0.01–1.3] 0.009 0.04* [0.01 -0.46]
MINE3 Mine 3 0.978 0.97 [0.1–9.7] 0.297 0.36 [0.05–2.48]
MINE4 Mine 4 0.049 0.14* [0.02–0.99] 0.012 0.09 [0.02–0.59]
MINE5 Mine 5a 0.075 (1) 0.021 (1)
Constant 0.003 0.03* 0.996 1.00
EXM: Excluded from the original model.
*
Results are significant at 0.05 probability level.
a
Reference category.
94 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96

cause some bias in estimating the parameters by the logistic
regression model, or simply the endogeneity in the logistic
regression, and are discussed below.
Hence, the original model (with all the hypothesized
variables) and the nested model (dropping the variables
depth of cover and mine) are tested. The variable depth
of cover which is correlated with three variables like gallery
width, seam thickness and the mine variable, when
dropped from the original model; in the nested model,
the widened galleries are found to be 4.8 [95% confidence
interval (CI) is 1.4–16.5] times more prone as compared
to the reference category for major roof fall accidents (case
A). But in the original model for the same case A, this value
is 4.5 [95% CI is 1.3–5.5]. Similarly, in the original model
for the unsupported roof the adjusted odds ratio is 3.3
[95% CI is 1.1–9.2], and for the nested model the value is
3.4 [95% CI is 1.2–9.5]. Similar are the cases for the mine
variable and shows significance for the same variable in
both the original and nested model (Table 6). Similarly,
for case B (Table 7), when the variable depth of cover is
dropped from the original model, in the nested model wid-
ened galleries are found to be 7.7 [95% CI is 1.7–34.5] times
more prone to serious roof fall accidents compared to the
reference category; whereas in the original model this value
is 6.7 [95% CI is 1.5–30.1]. Similarly, in the original model
for case B for the unsupported roof, the adjusted odds ratio
is 3.7 [95% CI is 1.0–13.7], and for the nested model the
adjusted odds ratio is 4.0 [95% CI is 1.1–14.2]. Similar
are the cases for the mine variable and shows significance
for the same variable in both the original and nested model
for case B. The detailed results are presented in Tables 6
and 7 for the cases A and B, respectively, when the variable
depth of cover is excluded from the original model.
The variable ‘mine’ also shows significant correlations
with the variables like depth of cover, width of gallery
and thickness of seam. Hence, this ‘mine’ variable is tested
after dropping it from the original model. In the nested
model, for case A, widened galleries are found to be 3.4
[95% CI is 1.2–9.5] times more prone compared to the ref-
erence category; whereas in the original model this is

Table 8
Comparison of original model and nested model after dropping the variable ‘Mine’ for major versus minor accidents (case A)
Variables and category description Major versus minor roof fall accidents
Original model Nested model
Significance level exp(b) [95% CI] Significance level exp(b) [95% CI]
Depth of cover (m)
DEPTH1 6 120a (1) (1)
DEPTH2 > 120 0.047 5.6* [1.0–30.5] 0.004 8.8* [2.0–38.4]
Gallery width (m)
WIDTH1 < 4.2a (1) (1)
WIDTH2 P 4.2 0.017 4.5* [1.3–5.5] 0.019 3.4* [1.2–9.5]
Seam thickness (m)
SEAMTH1 6 4.8 0.029 7.9* [1.2–50.2] 0.017 6.0* [1.4–26.2]
SEAMTH2 > 4.8a (1) (1)
Immediate roof
IMMRF1 Weaka (1) (1)
IMMRF2 Strong 0.644 0.8 [0.3–2.1] 0.449 0.7 [0.3–1.7]
Mining height (m)
HEIGHT1 6 2.5a (1) (1)
HEIGHT2 > 2.5 0.957 1.0 [0.3–3.3] 0.116 0.5 [0.2–1.2]
Roof support status
SUPPST1 Supporteda (1) (1)
SUPPST2 Unsupported 0.026 3.3* [1.1–9.2] 0.070 2.4 [0.9–6.1]
Roof fall location
LOCAT1 Facea (1) (1)
LOCAT2 Outbye 0.368 0.6 [0.2–1.8] 0.401 0.7 [0.2–1.8]
Mine variables
MINE1 Mine 1 0.050 0.09 [0.01–1.00] EXM
MINE2 Mine 2 0.026 0.09* [0.01–0.75] EXM
MINE3 Mine 3 0.378 0.43 [0.06–2.85] EXM
MINE4 Mine 4 0.030 0.17* [0.03–0.84] EXM
MINE5 Mine 5a 0.051 (1) EXM
Constant 0.008 0.06* 0.005 0.08*
EXM: Excluded from the original model.
*
Results are significant at 0.05 probability level.
a
Reference category.
 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96 95

found to be 4.5 [95% CI is 1.3–5.5]. Similarly, in the ori-
ginal model for the unsupported roof adjusted odds ratio
is 3.3 [95% CI is 1.1–9.2]; and for the nested model this
variable do not have statistical significance (Table 8).
For the same case A, depth of cover in the original model
is 5.6 [95% CI is 1.0–30.5] times more amenable to the ref-
erence category, but in the nested model this value is
found to be 8.8 [95% CI is 2.0–38.4]. Similar is the case
for the seam thickness. Thin seams are 7.9 [95% CI is
1.2–50.2] times more prone to major roof fall accidents,
and in the nested model this value is found to be 6.0
[95% CI is 1.4–26.2]. The detailed comparison of results
is presented in Table 8.
Similarly, while dropping the mine variable from the
original model for testing the case B, in the nested model,
widened galleries are found to be 5.3 [95% CI is 1.5–19.1]
times more prone to serious roof fall accidents in com-
pared to the reference category; whereas, in the original
model this is found to be 6.7 [95% CI is 1.5–30.1]. Simi-
larly, in the original model for the unsupported roof
adjusted odds ratio is 3.7 [95% CI is 1.0–13.7]; and in
the nested model this variable does not have statistical
significance. The detailed results and comparisons for case
B are presented in Table 9.
Chi-square is the difference in the likelihood ratios
(2LL) for the two models (original and nested), and the
degrees of freedom is the difference in degrees of freedom
for the two models. The difference of this change in chi-
square for the mine variable is 11.296 at 4 degrees of free-
dom and shows the p-value of 0.023. For the variable depth
of cover, similar results are obtained, i.e. the change in chi-
square is 4.815 at 1 degrees of freedom having p-value of
0.028. This indicates that both the variables matter signifi-
cantly in predicting the response variable, degree of roof
fall accident. Hence, though there is some correlation exist-
ing between the independent variables, the overall model
has good fit for the dataset as the model chi-square, Hos-
mer–Lemeshow test, and Wald statistics for individual
parameters performs well. The results are presented in
Table 5 and briefly summarized below.

Table 9
Comparison of original model and nested model after dropping the variable ‘Mine’ for serious versus minor accidents (case B)
Variables and category description Serious versus minor roof fall accidents
Original model Nested model
Significance level exp(b) [95% CI] Significance level exp(b) [95% CI]
Depth of cover (m)
DEPTH1 6 120a (1) (1)
DEPTH2 > 120 0.061 7.1 [0.9–55.1] 0.856 0.9 [0.3–2.7]
Gallery width (m)
WIDTH1 < 4.2a (1) (1)
WIDTH2 P 4.2 0.013 6.7* [1.5–30.1] 0.010 5.3* [1.5 -19.1]
Seam thickness (m)
SEAMTH1 6 4.8 0.152 4.1 [0.6–28.1] 0.105 3.6 [0.8–17.3]
SEAMTH2 > 4.8a (1) (1)
Immediate roof
IMMRF1 Weaka (1) (1)
IMMRF2 Strong 0.403 0.60 [0.2–1.9] 0.668 0.4 [0.2–1.9]
Mining height (m)
HEIGHT1 6 2.5a (1) (1)
HEIGHT2 > 2.5 0.812 1.2 [0.3–4.6] 0.170 0.5 [0.2–1.4]
Roof support status
SUPPST1 Supporteda (1) (1)
SUPPST2 Unsupported 0.048 3.7* [1.0–13.7] 0.154 2.2 [0.7–6.8]
Roof fall location
LOCAT1 Facea (1) (1)
LOCAT2 Outbye 0.743 0.8 [0.2–2.7] 0.856 0.9 [0.3–2.7]
Mine variables
MINE1 Mine 1 0.420 0.32 [0.02–5.2] EXM
MINE2 Mine 2 0.075 0.09 [0.01–1.3] EXM
MINE3 Mine 3 0.978 0.97 [0.1–9.7] EXM
MINE4 Mine 4 0.049 0.14* [0.02 -0.99] EXM
MINE5 Mine 5a 0.075 (1) EXM
Constant 0.003 0.03* 0.003 0.05*
EXM: Excluded from the original model.
*
Results are significant at 0.05 probability level.
a
Reference category.
96 S.K. Palei, S.K. Das / Safety Science 47 (2009) 88–96
The results show that wider gallery width (P4.2 m) is
4.5 [95% CI is 1.3–5.5] times more prone to major accidents
than narrower gallery width and 6.7 [95% CI is 1.5–30.1]
times more prone to serious accidents. This may be due
to the fact that rock load is dependent on width of galleries.
Widened galleries experience less pillar safety factor. Less
thick seams (64.8 m) are more amenable to major roof fall
accidents and 7.9 [95% CI is 1.2–50.2] times more amenable
to major accidents in comparison to thick seams (>4.8 m).
Unsupported, or partially supported roofs are 3.3 [95% CI
is1.1–9.2] times higher prone to major accident than the
supported roofs and 3.7 [95% CI is 1.0–13.7] times higher
risky in relation to serious accident. Deep workings
(>120 m) are 5.6 [95% CI is 1.0–30.5] times more prone
to major accident as compared to shallow depth category
(6120 m). The reason may be due to the concentrations
of more in-situ stresses at deep workings. Variables like
mining height, workplace locations and roof conditions
have not had statistical significance at the mentioned prob-
ability level in the logistic model. MINE5 is 1/0.09 times
more risky for major accidents than that of MINE2 and
1/0.17 times that of MINE4. Again, MINE5 is 1/0.14 times
more prone to serious accidents in comparison to MINE4.
A solemn attention should be paid on the support design
based on the depth of super incumbent strata restricting
the gallery width to reduce the risks of roof fall.
4. Conclusion
This paper focused on evaluation of severities of roof fall
accidents in bord and pillar method of working through the
application of logistic regression model. Again, logistic
regression model deals with probability function and treats
each individual as a separate observation to estimate the
probabilities. On the basis of the model results, the unsup-
ported, or partially supported roofs are more prone to major
accidents than the supported roofs, but the earlier study by
Biswas and Zipf (2003) only noted more ground fall inci-
dents in supported areas (67.6%) than the unsupported
areas (32.4%) in their taxonomic analysis. The variable like
workplace locations have not had statistical significance in
this study, whereas Maiti (2003) in his study of occupational
risk of underground coal miners found outbye locations are
more accident-prone workplaces than the faces. The specific
conclusions for the case study mines based on the available
roof fall dataset are summarized below:
(i) Unsupported or partially supported roofs are more
prone to major accident than the supported roofs.
(ii) Deep workings (>120 m) are more prone to major
accidents in comparison to shallow depth workings
(6120 m).
(iii) Widened galleries (P4.2 m) are more prone to major
accidents than the narrower width galleries (<4.2 m).
(iv) Less thick seams (64.8 m) are more amenable to
major accident in comparison to thick seams
(>4.8 m).
(v) A solemn attention should be paid on the support
design based on the depth of super incumbent strata,
restricting the gallery width to reduce the risks of roof
fall.
References
Bhattacherjee, A., 2000. Coal mine safety – where do we go from here.
Journal of Mines, Metals and Fuels XLVIII, 215–221.
Biswas, K., Zipf, R.K., 2003. Root causes of ground fall related incidents
in US Mining industry. In: 22nd International Conference on Ground
Control in Mining, pp. 335–343.
Cox, D.K., 1970. The Analysis of Binary Data. Chapman and Hall,
London.
Das, S.K., 1984. Accidents at Moonidih – a comparative analysis of
conventional and powered support faces. The Indian Mining and
Engineering Journal XXIII (1&2), 6–9.
Das, S.K., 1994. Modern Coal Mining Technology. Lovely Prakashan,
Dhanbad, India.
Das, S.K., 2000. Observations and classification of roof strata behaviour
over longwall coal mining panels in India. International Journal of
Rock Mechanics and Mining Sciences 37 (4), 585–597.
DGMS, Directorate General of Mines Safety (Data source), 1996–2001.
Dhanbad, India.
Dolinar, D.R., Mark, C., Molinda, G.M., 2001. Design of Primary Roof
Support Systems in US Coal Mines Based on the Analysis of Roof Fall
Rates. NIOSH Pitsbury Laboratory, USA.
Duzgun, H.S.B., Einstein, H.H., 2004. Assessment and management of
roof fall risks in underground coal mines. Safety Science 42, 23–41.
Maiti, J., 2003. Development of risk indices for underground coal mine
workers in India. Mining Technology, Transactions of the Institution
of Mining and Metallurgy 112 (2), A119–A124.
Molinda, G.M., Mark, C., Dolinar, D., 2000. Assessing coal mine roof
stability through roof fall analysis. Information Circular, IC9453.
Department of Health and Human Services, Center for Disease
Control and Prevention, NIOSH.
Palei, S.K., Das, S.K., 2005. Study of reliability and estimation of risk due
to roof fall in underground coal mines. Proceedings of the Interna-
tional Symposium on Advances in Mining Technology and Manage-
ment. Indian Institute of Technology, Kharagpur, India, pp. 575–578.
Phillipson, S.E., 2003. Control of coal bed decollement-related slickensides
on roof falls in North American Late Paleozoic coal basins. Interna-
tional Journal of Coal Geology 53, 181–195.
SPSS, Statistical Package for Social Sciences Inc., 1993. 444 N, Michigan
Avenue, Chicago, Illinois, 60611, USA.