The Flight of the Space Shuttle Challenger

On January 28, 1986, the space shuttle Challenger took o on the 25th ight in NASA's space shuttle
program. Less than 2 minutes into the ight, the spacecraft exploded, killing all on board. A Presidential
Commission was formed to explore the reasons for this disaster.
First, a little background information: the space shuttle uses two booster rockets to help lift it into
orbit. Each booster rocket consists of several pieces whose joints are sealed with rubber O{rings, which are
designed to prevent the release of hot gases produced during combustion. Each booster contains 3 primary
O{rings (for a total of 6 for the orbiter). In the 23 previous ights for which there were data (the hardware
for one ight was lost at sea), the O{rings were examined for damage.
One interesting question is the relationship of O{ring damage to temperature (particularly since it was
(forecasted to be) cold | 31o F | on the morning of January 28, 1986). There was a good deal of discussion
among the Morton Thiokol engineers the previous day as to whether the ight should go on as planned or
not (an important point is that no statisticians were involved in the discussions). A simplied version of one
of the arguments made is as follows. There were 7 previous ights where there was damage to at least one
O{ring. Consider the following table. The entry p^ is the frequency estimate of the probability of an O{ring
failing for that ight.

Ambient temperature p^

53o .333
57o .167
58o .167
63o .167
70o .167
70o .167
75o .333

If you look at the table above, there's no apparent relationship between temperature and the probability of
damage; higher damage occurred at both lower and higher temperatures. Thus, the fact that it was going to
be cold on the day of the ight doesn't imply that the ight should be scrubbed. (In fact, this table was not
actually constructed the night of January 27th, but was rather given later by two Thiokol sta members as
an example of the reasoning in the pre{launch debate. The actual charts faxed from the Thiokol engineers
to NASA that night were considerably less informative than even this seriously awed table.)
Unfortunately, this analysis is completely inappropriate. The problem is that it is ignoring the 16 ights
where there was no O{ring damage, acting as if there is no information in those ights. This is clearly
absurd! If ights with high temperatures never had O{ring damage, for example, that would certainly tell
us a lot about the relationship between temperature and O{ring damage! In fact, here is a scatter plot of
the frequency estimates of the probability of O{ring damage versus temperature for all of the ights:

c 1999, Jerey S. Simono 1

 0.3
Proportion failing

0.2

0.1

0.0

50 60 70 80
Temperature

The picture is very dierent now. With the exception of the one observation in the upper right of
the plot, there is a clear inverse relationship between the probability of O{ring damage and the ambient
temperature | lower temperature is associated with higher probability of failure (the unusual observation
is the ight of the Challenger from October 30 through November 6, 1985; one way that it was dierent was
that the two O{rings damaged in that ight suered only \blow{by" [where hot gases rush past the O{ring],
while in all of the other ights damaged O{rings suered \erosion" [where the O{rings burn up], as well as
(possibly) blow{by). A plot of this kind would certainly have raised some alarms as to the advisability of
launching the shuttle. Unfortunately, such a plot was never constructed.
Here is the full set of data:
Row Temp Damaged O-rings

1 53 2 6
2 57 1 6
3 58 1 6
4 63 1 6
5 66 0 6
6 67 0 6
7 67 0 6
8 67 0 6
9 68 0 6
10 69 0 6
11 70 0 6
12 70 0 6
13 70 1 6
14 70 1 6
15 72 0 6
16 73 0 6

c 1999, Jerey S. Simono 2

 17 75 0 6
18 75 2 6
19 76 0 6
20 76 0 6
21 78 0 6
22 79 0 6
23 81 0 6

Logistic regression can be used to analyze the relationship between temperature and the probability of
O{ring failure more precisely. In this case, the number of failures is the target variable (which MINITAB calls
Success, remember), and the program is told that the number of trials is given in a variable O-rings (which
is 6 for each ight here). Here is the output of the logistic analysis:
Binary Logistic Regression

Link Function: Logit

Response Information

Variable Value Count

Damaged Success 9
Failure 129
O-rings Total 138

Logistic Regression Table

Odds 95% CI
Predictor Coef StDev Z P Ratio Lower Upper
Constant 5.085 3.052 1.67 0.096
Temperat -0.11560 0.04702 -2.46 0.014 0.89 0.81 0.98

Log-Likelihood = -30.198
Test that all slopes are zero: G = 6.144, DF = 1, P-Value = 0.013

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 13.572 14 0.482
Deviance 11.956 14 0.610
Hosmer-Lemeshow 5.677 4 0.225

Table of Observed and Expected Frequencies:

(See Hosmer-Lemeshow Test for the Pearson Chi-Square Statistic)

Group
Value 1 2 3 4 5 6 Total
Success
Obs 0 2 2 0 2 3 9
Exp 0.3 0.6 1.6 1.8 2.0 2.7
Failure
Obs 18 22 34 30 16 9 129
Exp 17.7 23.4 34.4 28.2 16.0 9.3

Total 18 24 36 30 18 12 138

c 1999, Jerey S. Simono 3

 Measures of Association:
(Between the Response Variable and Predicted Probabilities)

Pairs Number Percent Summary Measures

Concordant 759 65.4% Somers' D 0.38
Discordant 315 27.1% Goodman-Kruskal Gamma 0.41
Ties 87 7.5% Kendall's Tau-a 0.05
Total 1161 100.0%

The slope coecient has the following natural interpretation: each increase in temperature by one degree
Fahrenheit is associated with an estimated multiplication of the relative odds of an O-ring failure
P (O ; ring fails)
P (O ; ring does not fail)
by exp(;:1156) = 0:891, or roughly an 11% decrease. This value is given in the output under Odds Ratio,
along with a 95% condence interval. If this interval does not contain 1, there is signicant predictive power
of the predictor on the probability of success (at a .05 level).
There are two other tests given related to the strength of the predictive power of temperature for
probability of an O{ring failure. The z {statistic of ;2:46 for Temperature corresponds to a t{statistic
in linear regression, and is called a Wald statistic (it is equivalent to the odds ratio condence interval
comparison mentioned above). The G{statistic, given as testing that all slopes are zero, corresponds to the
F {statistic for overall signicance in linear regression. Note that for least squares linear regression these
two tests are equivalent when there is one predictor, but here the tail probabilities are slightly dierent,
demonstrating that the two tests are not exactly equivalent.
The three goodness{of{t tests are designed to test whether the logistic model ts the data adequately.
All three are based on a 2 {test construction. For each value of temperature given in the data (there are
J = 16 distinct values in these data), let p^j be the tted probability of O{ring failure, let fj be the observed
number of O{rings that failed, and let nj be the number of O{rings at risk for that temperature (6 for each
ight at that temperature). Note that looking at the data this way means that all ights at a given ambient
temperature are pooled together and treated as indistinguishable. These values can be obtained as Storage
from a logistic t, and are as follows:
Row Temperature NOCC1 NTRI1 EPRO1

1 53 2 6 0.260787
2 57 1 6 0.181787
3 58 1 6 0.165220
4 63 1 6 0.099940
5 66 0 6 0.072783
6 67 0 18 0.065357
7 67 * * *
8 67 * * *
9 68 0 6 0.058640
10 69 0 6 0.052575
11 70 2 24 0.047106
12 70 * * *
13 70 * * *
14 70 * * *
15 72 0 6 0.037749
16 73 0 6 0.033767
17 75 2 12 0.026985
18 75 * * *

c 1999, Jerey S. Simono 4

 19 76 0 12 0.024110
20 76 * * *
21 78 0 6 0.019229
22 79 0 6 0.017166
23 81 0 6 0.013671

Note that the nj values range from 6 to 24. The sum of the nj values is the total sample size, or here 132.
The Pearson goodness{of{t statistic equals
X (fj ; nj p^j )2
X2 = ;
j
nj p^j

while the deviance statistic equals  

X fj
G2 = 2 fj ln :
j
nj p^j

When the nj values are reasonably large, each of these statistics follows a 2 distribution on J ; p ; 1 degrees
of freedom, where p is the number of predictors in the model, under the null hypothesis that the logistic
regression model ts the data. Thus, a small tail probability suggests that the linear logistic regression model
is not appropriate for the data. Here both tests have high tail probabilities, indicating no problem with the
linear logistic model.
Unfortunately, these tests are not trustworthy when the nj values are small (the nj = 6 values here are
marginal). This is the justication for the third goodness{of{t test, the Hosmer{Lemeshow test. In this
test, all of the 138 observations are ordered by estimated O{ring failure probability (of course for these data
all of the O{rings for a given ight have the same value of Temp, and therefore the same estimated probability
of O{ring failure). The observations are then divided into g roughly equisized groups; g is usually taken to
be 10, except when that would lead to too few observations in each group (as is the case here, where g = 6).
Based on this new categorization of the data there are values of fj , nj and nj p^j , all of which are given in
the Hosmer{Lemeshow table in the output. Then, the Hosmer{Lemeshow goodness{of{t test is the usual
Pearson goodness{of{t test based on the new categorization, which is compared to a 2 distribution on
g ; 2 degrees of freedom. It can be seen that the Hosmer{Lemeshow test also does not indicate a lack of t
here. Even the Hosmer{Lemeshow test is suspect, however, when its expected counts for either group are
too small (less than two or three, say), which is the case here.
The statistical signicance and goodness{of{t of this model are comforting, of course, but does tem-
perature provide predictive power of any practical importance? Some guidance to answer this question is
given in the output under Measures of Association. Consider the tted logistic regression model, with
resultant tted probabilities of O{ring failure p^ for each of the n = 138 observations. There are n1 = 9 ob-
served O{ring failures, and n0 = 129 observed non{failures. Consider each of the pairs (i; j ) of observations
where one observation is a failure (i) and the other is a non{failure (j ). There are 9  129 = 1161 such
pairs, each of which has a corresponding pair (^pi ; p^j ). We would like the estimated probability of failure to
be higher for the observed failure observation than for the observed non{failure observation; that is, p^i > p^j .
Such pairs are called concordant. If for a given pair p^i < p^j , the pair is called discordant. We would like to
have a high percentage of concordant pairs, and a low percentage of discordant pairs. Here there are 65.4%
concordant pairs and 27.1% discordant ones, a reasonably good performance. There are no formal cutos
for what constitutes a \good enough" performance here, but observed values can be compared for dierent
possible models to assess relative practical performance. The statistics Somers' D, Goodman{Kruskal and
Kendall's a are dierent ways of summarizing these concordancies and discordancies, with higher values
indicating more concordancy (e.g., D is the dierence between concordant and discordant pairs).
Just as is true for other regression models, unusual observations can have a strong eect on a tted
logistic regression model. Among the diagnostics that are available for logistic regression are three that
roughly correspond to the standardized residuals (here the standardized Pearson residuals), Cook's distance
(here the standardized Delta{beta [
 ]) and leverage values. Here are the values here:

c 1999, Jerey S. Simono 5

 Row Temperature SPRE1 DSBE1 HI1

1 53 0.56258 0.29503 0.482442

2 57 -0.10892 0.00340 0.222683
3 58 0.01054 0.00002 0.180967
4 63 0.56529 0.02428 0.070609
5 66 -0.70527 0.02790 0.053119
6 67 -1.21983 0.27107 0.154100
7 67 * * *
8 67 * * *
9 68 -0.62752 0.02110 0.050853
10 69 -0.59239 0.01894 0.051209
11 70 0.94160 0.23361 0.208538
12 70 * * *
13 70 * * *
14 70 * * *
15 72 -0.49902 0.01443 0.054772
16 73 -0.47134 0.01322 0.056146
17 75 3.17765 1.33690 0.116919
18 75 * * *
19 76 -0.57995 0.04523 0.118540
20 76 * * *
21 78 -0.35377 0.00799 0.060045
22 79 -0.33389 0.00712 0.059995
23 81 -0.29729 0.00555 0.059061

There is an apparent outlier at row 17, corresponding to an ambient temperature of 75o. Unfortunately,
since there are two ights at that temperature, we can't tell for sure which is actually the outlier (of course,
in this case we know what it is from the earlier graph, but in general the collapsing approach of Minitab
makes it dicult to tell which observation is actually an outlier if there are replications in the data.
For this reason, it's a good idea to try to remove the collapsing eect by forcing each observation to
have a unique set of predictor variable values, at least when looking at diagnostics. The way this is done
is by \jittering" at least one predicting variable by adding a small amount of random noise to the variable.
In fact, in this context we know that the temperature values are only given to the nearest integer value, so
this is not at all unreasonable, but even if the values were exact, we need to do this if we want to examine
diagnostics. Here are the results of a logistic t using jittered temperature:
Binary Logistic Regression

Link Function: Logit

Response Information

Variable Value Count

Damaged Success 9
Failure 129
O-rings Total 138

Logistic Regression Table

Odds 95% CI
Predictor Coef StDev Z P Ratio Lower Upper
Constant 5.085 3.052 1.67 0.096
Temperat -0.11561 0.04702 -2.46 0.014 0.89 0.81 0.98

c 1999, Jerey S. Simono 6

 Log-Likelihood = -30.198
Test that all slopes are zero: G = 6.145, DF = 1, P-Value = 0.013

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 29.985 21 0.092
Deviance 18.085 21 0.644
Hosmer-Lemeshow 8.207 6 0.223

Table of Observed and Expected Frequencies:

(See Hosmer-Lemeshow Test for the Pearson Chi-Square Statistic)

Group
Value 1 2 3 4 5 6 7 8 Total
Success
Obs 0 0 2 2 0 0 2 3 9
Exp 0.3 0.5 0.6 0.8 1.0 1.2 2.0 2.7
Failure
Obs 18 18 16 16 18 18 16 9 129
Exp 17.7 17.5 17.4 17.2 17.0 16.8 16.0 9.3

Total 18 18 18 18 18 18 18 12 138

Measures of Association:
(Between the Response Variable and Predicted Probabilities)

Pairs Number Percent Summary Measures

Concordant 759 65.4% Somers' D 0.38
Discordant 315 27.1% Goodman-Kruskal Gamma 0.41
Ties 87 7.5% Kendall's Tau-a 0.05
Total 1161 100.0%

As would be expected, the output changes very little, with the exception of the Pearson and Deviance
statistics (recall that their construction depends on how covariate patterns are dened). Here are the
diagnostics:
Row Temperature SPRE2 DSBE2 HI2 EPRO2

1 53 0.56312 0.29509 0.482020 0.260696

2 57 -0.10927 0.00342 0.222782 0.181836
3 58 0.00959 0.00002 0.181253 0.165351
4 63 0.56538 0.02428 0.070600 0.099932
5 66 -0.70541 0.02792 0.053130 0.072810
6 67 -0.66515 0.02396 0.051372 0.065376
7 67 -0.66474 0.02392 0.051361 0.065302
8 67 -0.66465 0.02392 0.051359 0.065285
9 68 -0.62743 0.02109 0.050856 0.058624
10 69 -0.59263 0.01896 0.051208 0.052615
11 70 -0.55958 0.01722 0.052131 0.047136
12 70 -0.55956 0.01722 0.052132 0.047133
13 70 1.41935 0.11080 0.052134 0.047124

c 1999, Jerey S. Simono 7

 14 70 1.41904 0.11075 0.052132 0.047135
15 72 -0.49870 0.01442 0.054790 0.037702
16 73 -0.47110 0.01321 0.056161 0.033734
17 75 -0.42012 0.01096 0.058472 0.026950
18 75 4.77345 1.41487 0.058464 0.026976
19 76 -0.39706 0.00993 0.059268 0.024123
20 76 -0.39697 0.00993 0.059271 0.024112
21 78 -0.35357 0.00799 0.060046 0.019208
22 79 -0.33384 0.00711 0.059994 0.017161
23 81 -0.29737 0.00555 0.059064 0.013678

Now it's clear that the previously mentioned ight (number 18) is a very clear outlier, with 2 of 6
O{rings damaged when the estimated probability of O{ring damage was only .027. Here is output from the
data with that ight omitted (I'm sticking with the jittered data):
Binary Logistic Regression

Link Function: Logit

Response Information

Variable Value Count

Damaged Success 7
Failure 125
O-rings Total 132

Logistic Regression Table

Odds 95% CI
Predictor Coef StDev Z P Ratio Lower Upper
Constant 8.664 3.635 2.38 0.017
Temperat -0.17684 0.05870 -3.01 0.003 0.84 0.75 0.94

Log-Likelihood = -22.039
Test that all slopes are zero: G = 10.660, DF = 1, P-Value = 0.001

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 13.398 20 0.860
Deviance 9.406 20 0.978
Hosmer-Lemeshow 3.908 6 0.689

Table of Observed and Expected Frequencies:

(See Hosmer-Lemeshow Test for the Pearson Chi-Square Statistic)

Group
Value 1 2 3 4 5 6 7 8 Total
Success
Obs 0 0 1 1 0 0 3 2 7
Exp 0.1 0.2 0.3 0.4 0.6 0.8 2.7 2.0
Failure
Obs 18 18 17 17 18 18 15 4 125
Exp 17.9 17.8 17.7 17.6 17.4 17.2 15.3 4.0

c 1999, Jerey S. Simono 8

 Total 18 18 18 18 18 18 18 6 132

Measures of Association:
(Between the Response Variable and Predicted Probabilities)

Pairs Number Percent Summary Measures

Concordant 671 76.7% Somers' D 0.61
Discordant 137 15.7% Goodman-Kruskal Gamma 0.66
Ties 67 7.7% Kendall's Tau-a 0.06
Total 875 100.0%

The strength of the relationship has gone up considerably once the outlier is removed, with there now
being an estimated 16% reduction in the odds of an O{ring being damaged with each additional degree of
temperature at launch. The goodness{of{t tests suggest no lack of t (remember, the Pearson and deviance
tests are at least marginally valid here, since there are 6 replications for each ight). Here are diagnostics:

Row Temperature SPRE3 DSBE3 HI3 EPRO3

1 53 0.02865 0.001039 0.558615 0.329680

2 57 -0.20157 0.012003 0.228058 0.195330
3 58 -0.01796 0.000071 0.180636 0.169155
4 63 0.85067 0.060061 0.076638 0.077459
5 66 -0.56312 0.021926 0.064673 0.047104
6 67 -0.51494 0.017792 0.062877 0.039768
7 67 -0.51446 0.017753 0.062861 0.039698
8 67 -0.51435 0.017745 0.062857 0.039682
9 68 -0.47077 0.014490 0.061368 0.033508
10 69 -0.43096 0.011837 0.059913 0.028277
11 70 -0.39412 0.009620 0.058318 0.023798
12 70 -0.39410 0.009619 0.058318 0.023796
13 70 2.36664 0.346845 0.058314 0.023789
14 70 2.36606 0.346697 0.058318 0.023798
15 72 -0.32907 0.006236 0.054450 0.016779
16 73 -0.30092 0.004986 0.052192 0.014103
17 75 -0.25142 0.003125 0.047109 0.009939
18 76 -0.23015 0.002463 0.044429 0.008365
19 76 -0.23006 0.002460 0.044418 0.008359
20 78 -0.19204 0.001489 0.038797 0.005873
21 79 -0.17564 0.001153 0.036032 0.004932
22 81 -0.14686 0.000682 0.030671 0.003472

There are no extreme outliers, but the low temperature cases are possible leverage points (this is not
surprising, given that most launches were at temperatures over 65o). The noteworthy 70o observations
correspond to two 70o ights where there was an O{ring failure. Omitting these two ights doesn't change
things very much (strengthening the relationship further), although a plot of the change in the Pearson
statistic versus estimated probability does show the two points as unusual.
c 1999, Jerey S. Simono 9
 Delta Chi-Square versus Probability

4
Delta Chi-Square

0.0 0.1 0.2 0.3

Probability

Binary Logistic Regression

Link Function: Logit

Response Information

Variable Value Count

Damaged Success 5
Failure 115
O-rings Total 120

Logistic Regression Table

Odds 95% CI
Predictor Coef StDev Z P Ratio Lower Upper
Constant 14.417 5.710 2.52 0.012
Temperat -0.28028 0.09864 -2.84 0.004 0.76 0.62 0.92

Log-Likelihood = -13.319
Test that all slopes are zero: G = 14.931, DF = 1, P-Value = 0.000

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 3.446 18 1.000
Deviance 2.780 18 1.000
Hosmer-Lemeshow 1.564 8 0.992

Table of Observed and Expected Frequencies:

(See Hosmer-Lemeshow Test for the Pearson Chi-Square Statistic)

Group
Value 1 2 3 4 5 6 7 8 9 10 Total
Success
Obs 0 0 0 0 0 0 0 0 2 3 5
Exp 0.0 0.0 0.0 0.0 0.1 0.1 0.2 0.2 1.1 3.4
Failure

c 1999, Jerey S. Simono 10

 Obs 12 12 12 12 12 12 12 12 10 9 115
Exp 12.0 12.0 12.0 12.0 11.9 11.9 11.8 11.8 10.9 8.6

Total 12 12 12 12 12 12 12 12 12 12 120

Measures of Association:
(Between the Response Variable and Predicted Probabilities)

Pairs Number Percent Summary Measures

Concordant 525 91.3% Somers' D 0.87
Discordant 27 4.7% Goodman-Kruskal Gamma 0.90
Ties 23 4.0% Kendall's Tau-a 0.07
Total 575 100.0%

What about the morning of January 28, 1986? Here is a plot of the logistic curve for dierent values of
temperature based on all ights except the October/November 1985 ight:
Estimated probability of O-ring failure

Scatter Plot of PROBB vs TEMPER

1.0

0.8

0.6

0.4

0.2

0.0

10 30 50 70 90

Ambient temperature at launch

Substituting into the logistic function gives a probability estimate of O{ring failure for a temperature
of 31o of .96! (This is an extrapolation, but you get the idea.) Indeed, with the benet of hindsight, it can
be seen that the Challenger disaster was not at all surprising, given data that were available at the
time of the ight. As a result of its investigations, one of the recommendations of the commission was
that a statistician be part of the ground control team from that time on. A complete (and more correct)
discussion of this material can be found in the paper \Risk Analysis of Space Shuttle: Pre{Challenger
Prediction of Failure," by S.R. Dalal, E.B. Fowlkes and B.A. Hoadley, Journal of the American Statistical
Association, 84, 945{957 (1989). Chapter 2 of Edward R. Tufte's 1997 book Visual Explanations: Images
and Quantities, Evidence and Narrative discusses the background of the disaster, and the charts used by the
Thiokol engineers in their discussions with NASA.
c 1999, Jerey S. Simono 11
 By the way, an alternative way that these data might have been presented was as a set of 138 observations
(one for each O{ring, rather than one for each ight), with a 0/1 target variable re ecting failure or non{
failure of each O{ring. This is what the dataset would look like:
Row Temp Failed

1 53 1
2 53 1
3 53 0
4 53 0
5 53 0
6 53 0
7 57 1
8 57 0
9 57 0
10 57 0
11 57 0
12 57 0
13 58 1
14 58 0
15 58 0
16 58 0
17 58 0
18 58 0
19 63 1
20 63 0
21 63 0
22 63 0
23 63 0
24 63 0
25 66 0
26 66 0
27 66 0
28 66 0
29 66 0
30 66 0
31 67 0
32 67 0
33 67 0
34 67 0
35 67 0
36 67 0
37 67 0
38 67 0
39 67 0
40 67 0
41 67 0
42 67 0
43 67 0
44 67 0
45 67 0
46 67 0
47 67 0

c 1999, Jerey S. Simono 12

 48 67 0
49 68 0
50 68 0
51 68 0
52 68 0
53 68 0
54 68 0
55 69 0
56 69 0
57 69 0
58 69 0
59 69 0
60 69 0
61 70 0
62 70 0
63 70 0
64 70 0
65 70 0
66 70 0
67 70 0
68 70 0
69 70 0
70 70 0
71 70 0
72 70 0
73 70 1
74 70 0
75 70 0
76 70 0
77 70 0
78 70 0
79 70 1
80 70 0
81 70 0
82 70 0
83 70 0
84 70 0
85 72 0
86 72 0
87 72 0
88 72 0
89 72 0
90 72 0
91 73 0
92 73 0
93 73 0
94 73 0
95 73 0
96 73 0
97 75 0
98 75 0
99 75 0
100 75 0

c 1999, Jerey S. Simono 13

 101 75 0
102 75 0
103 75 1
104 75 1
105 75 0
106 75 0
107 75 0
108 75 0
109 76 0
110 76 0
111 76 0
112 76 0
113 76 0
114 76 0
115 76 0
116 76 0
117 76 0
118 76 0
119 76 0
120 76 0
121 78 0
122 78 0
123 78 0
124 78 0
125 78 0
126 78 0
127 79 0
128 79 0
129 79 0
130 79 0
131 79 0
132 79 0
133 81 0
134 81 0
135 81 0
136 81 0
137 81 0
138 81 0

Which representation is better? It turns out not to matter; if you analyze the data in this form, where
Failed is chosen as the Response variable in the Minitab dialog box, the resultant output will be identical
to that obtained using the data represented at the level of 23 dierent ights. There is one advantage to the
earlier representation, however; since the natural way to view these data is at the ight level, rather than
the O{ring level, jittering the data in the ight{level form is more natural.

c 1999, Jerey S. Simono 14

 MINITAB commands
Logistic regression modeling is obtained by clicking on Stat ! Regression ! Binary Logistic Re-
gression. There are various ways that the data might be presented, which aect the command structure
to the program. The two most common forms are as follows:
(1) The target variable is given as the number of successes out of the number of trials (or the number of
items \at risk"). Enter the variable with the number of successes in the box next to Success:, and
enter the variable with the number of trials in the box next to Trial:.
(2) The target variable is a 0/1 variable that represents success or failure for each observation. Enter the
name of this variable in the box next to Response:.
The predicting variables for the model are entered under Model:. This includes both continuous variables and
categorical ones. Categorical variables must also be entered under Factors (optional):. Interactions with
(and between) factors are entered under Model: using the \multiplication" form as in ANOVA modeling.
Diagnostics, such as standardized Pearson residuals, leverages, and delta betas are obtained by clicking
on Storage. This dialog box also allows storage of tted success probabilities under Event probability,
the number of successes for each distinct covariate pattern under Number of occurrences of the event,
and the number of trials for each covariate pattern under Number of trials. Diagnostic plots, such as one
of Delta Chi{Square versus Event Probability, are obtained by clicking on Graphs.
To jitter a predicting variable, follow the following steps:
(1) Go to Calc ! Random Data ! Uniform. In the box labeled Generate rows of data, enter the sample
size (the number of observations, not the number of covariate patterns). Put in a new variable name
under Store in columns:, such as Jitter. Under Lower endpoint: enter a negative number close to zero,
such as ;:01. Under Upper endpoint: enter the positive version of this number (.01, for example). The
numbers you choose here should be smaller than the resolution of all of your predictors; so, for example,
if one predictor is given to three decimal digits, use ;:0001 and .0001 here.
(2) Go to the calculator, and create new versions of each predictor you wish to jitter as sums of the predictor
with Jitter. So, for example, the variable Newpred1 is determined as Predict1 + Jitter.
(3) Fit the logistic regression model using the new predictors in place of the old ones.

c 1999, Jerey S. Simono