The Monday blues.

Or are you more likely to have a heart attack on some days

of the week than others?

Introduction
A study done in Augsburg, Germany (Willich et al., 1993) looked at which days of the week people had
heart attacks. They wanted to know whether or not heart attacks are distributed equally across the days
of the week in the population. One subgroup of heart attack victims that they looked at were those who
were employed. The researchers found that 884 heart attacks from this subgroup were distributed
across the seven days of the week as shown in Table 1.

Day Sun Mon Tue Wed Thur Fri Sat

Number of heart attacks 106 160 123 115 141 10 132
7

Table 1: A distribution of the frequency of heart attacks for 884 people that were
employed at the time of their heart attack.

These data represent days of the week for 884 heart attacks in some larger population. We think of
these 884 heart attacks as a sample. Although this isn’t a random sample, this group probably is
representative of some larger population. (We will discuss this in more detail later.)

As a start, let’s first compare to one that is distributed evenly. What would this type of distribution look
like? Well, because there were 884 total heart attacks we would expect about 884/7 ≈ 126.3 heart
attacks each day. In the distribution in Table 1, Tuesday’s frequency is closest to what would be
expected, and Monday’s is the farthest from what would be expected.

Another way to rephrase the question that we want to answer is, if heart attacks were distributed
evenly in the larger population, how likely would it be for us to get a distribution as “off” as this? Is that
fairly likely to happen or is that very unlikely to happen?

Typical Variation or Noise

Let’s move away from our heart attack example for a moment to get an idea of what typical variation
looks like in a simpler situation. Suppose I rolled a fair six-sided die 120 times. What do you expect to
happen? You might say since 120/6 = 20, you would expect to get 20 ones, 20 twos, 20 threes and so on.
But do you really expect to get exactly 20 of each of these? Not necessarily. You should expect to get
close to 20 of each. And if you rolled the die 120 more times you should expect to get a distribution
different than the first; but again all the frequencies will be close to 20. But now the question is, what
does it mean to get close to 20?

To help get you to see what close to 20 might look like, we rolled a fair six-sided die 120 times and the
number of times each face landed up is shown in Table 2 in the row labeled Trial #1. You can see that
close to 20 means 1, 2 or 3 away for all of the outcomes except for the frequency of 5s. This frequency
of 14 is 6 away from what would be expected. But again, these results come from 120 rolls a fair die, so
these frequencies show the typical variation you could see or might expect. We repeated the process of
rolling a die 120 times four more times, with the results shown in the last four rows of Table 2.

Number showing 1 2 3 4 5 6
Trial #1 22 23 23 21 14 17
Trial #2 13 27 19 23 21 17
Trial #3 16 21 23 22 17 21
Trial #4 25 18 25 21 18 13
Trial #5 26 22 18 17 16 21

Table 2: Five different distributions of the number of times each face occurred in 120
rolls of a fair die.

Again, these show the type of variation you might see under this fair process. Sometimes this type of
variation is called noise. What we want to know, back in our heart attack data, are we just seeing this
sort of noise or is the observed variation from what is expected more extreme than that. If the variation
is more extreme, we might say there is something more going on than typical variation or noise. We
might say that the extra variation from what is expected is not just noise but is a signal.

Hypotheses
Now let’s get back to our heart attack example. There are two possible reasons why the distribution of
heart attacks deviates so much from one that would be equally distributed. One is that heart attacks in
the population are equally distributed across the seven days and the variation we are seeing here is just
“noise” in our data. The other reason is that heart attacks are not distributed evenly in the population so
the amount of variation we see in the sample data is more than we would expect to see if heart attacks
were evenly distributed in the population.

The two reasons we stated above are what we call hypotheses. In this example, the hypothesis that
heart attacks are distributed equally in the population is called the null hypothesis and often the
notation H0 is used to represent this hypothesis. The word null means zero, hence the subscript zero is
used in the notation. You can think of this zero as no change or in this case no change or difference from
what would be expected if heart attacks were equally distributed in the population. Formally, we could
write out the null hypothesis as:

H0: Heart attacks are distributed evenly across the days of the week for employed people in the
population.

The other hypothesis is basically the opposite, or that heart attacks are not distributed equally. We call
this hypothesis the alternative hypothesis and often the notation H a is used to represent this hypothesis.
Some people will refer to this as the research hypothesis. Formally, we could write out the alternative
hypothesis as:
Ha: Heart attacks are not distributed evenly across the days of the week for employed people in the
population.

In testing these hypotheses, we assume that the null hypothesis is true, or that all the heart attacks from
the population are distributed evenly. Then we determine the likelihood (or probability) of getting a
distribution as “off” as the one we obtained. So we now need a way to calculate this probability.

Statistic
To determine whether the variation from what is expected is more than just noise, we first need a way
to measure it. In statistical language, we need a statistic. A statistic is a measure of some attribute in a
sample. Simple statistics are things like a sample mean or a sample proportion. In this case, we want to
compare several categories, so we need a statistic that is a bit more complicated. We need something
that will measure how far away the distribution of heart attacks is from what is expected if there was an
equal distribution of heart attacks across the seven days of the week. Table 3 shows our original data for
the 884 heart attacks, though we have renamed this row the observed number of heart attacks because
this is the sample we observed. We have also added a row for the expected number of heart attacks.
This means these numbers are expected if we have a perfectly equal distribution of heart attacks across
the seven days. Because 884/3 ≈ 126.3, we have put that number in for each day. Don’t worry that this
is not a whole number. Even though we can’t have a fraction of a heart attack, we can use this in
developing our statistic.

Day Sun Mon Tue Wed Thur Fri Sat

OBSERVED number of heart attacks 106 160 123 115 141 107 132
EXPECTED number of heart attacks 126.3 126. 126. 126.3 126. 126.3 126.3
3 3 3

Table 2: The observed and expected number of heart attacks for each day of the week.

Now, how will we measure how far apart these two distributions are from one another? Subtracting
each expected frequency from its corresponding observed frequency might seem like a good place to
start. Then what should we do with the seven differences we get? We need our statistic to be a single
number not seven numbers. Perhaps we should add them up or maybe average them. Let’s give this a
try by first just adding up the differences.

( 106−126.3 ) + ( 160−126.3 ) + ( 123−126.3 ) + ( 115−126.3 )+ ( 141−126.3 ) + ( 107−126.3 )+ ( 132−126.3 )

¿−20.3+33.7+ (−3.3 )+ (−11.3 ) +14.7+ (−19.3 )+5.7

¿−0.1
What happened here? Why did we get such a small sum of all these differences? What happened was
the positive and negative differences cancelled each other out. In fact, if we didn’t round the expected
frequencies to 126.3, our sum of the differences would be exactly 0. If we simply sum the differences,
they will sum to zero every time, no matter what data set we start with. What else can we do so that we
don’t get these positive and negative differences that cancel each other out?
 Hopefully you are thinking either take the absolute values of the differences or square the differences.
Doing either of these will result in all positive values. Taking absolute values might seem the most logical
thing to do because it is simpler, however squaring will work better. In a little bit we will go through a
process to determine how unlikely our statistic would be if in fact the null hypothesis was true.
(Remember that the null hypothesis was that the population distribution of heart attacks is equally
distributed.) In fact, we are going to show you two different ways to do this. For one of those ways,
using absolute values does not work well, but squaring will. We’ll point this out to you later when it
arises.

We could just square the differences before adding them up and use that sum of squared differences as
our statistic. Just doing that will result in a fine statistic, but we can make it better. We will also divide
each squared difference by the expected frequency. We do this because the sample size matters. For
example, doesn’t it seem like 106 and 126.3 are closer together than 6 and 26.3 even though their
differences (or squared differences) are the same? After all, 26.3 is more than four times as large as 6
while 126.3 is nowhere near four times 106. By dividing by the expected frequencies we are
standardizing the statistic so that different distributions with different sample sizes give statistics that
can be compared on the same scale. We make these squared differences into relative squared
differences.

Okay, let’s put this all together. The statistic we have been describing is called the chi-square statistic
which has the symbol χ2. (Chi is a Greek letter and is pronounced ki like hi!) The formula for the χ2
statistic is the following.

( observed frequency −expected frequency )2

χ 2=∑
expected frequency

Note that the Σ in the formula is commonly called a summation symbol. It just means to add up all the
terms. (This symbol is also another Greek letter, in this case it is an upper-case sigma.) Also note in the
formula that the χ2-statistic can never be negative since the numerator (because it is squared) and the
denominator can never be negative. If all the observed frequencies were exactly the same as the
expected (like our observed distribution was equally distributed) we would end up with a χ2-statistic of
0. Can you see why? The smallest a χ2-statistic can be is 0 and the further you move the observed
frequencies away from the expected frequencies, the larger the χ2-statistic becomes.

Let’s calculate the χ2-statistic for our data.

2 ( 106−126.3 )2 ( 160−126.3 )2 ( 123−126.3 )2 ( 115−126.3 )2 ( 141−126.3 )2 ( 107−126.3 )2 ( 132−126.3 )2

χ= + + + + + +
126.3 126.3 126.3 126.3 126.3 126.3 126.3

(−20.3)2 33.7 2 (−3.3 )2 (−11.3 )2 14.7 2 (−19.3 )2 5.72

¿ + + + + + +
126.3 126.3 126.3 126.3 126.3 126.3 126.3

¿ 3.26+8.99+ 0.09+ 1.01+ 1.71+ 2.95+ 0.26

¿ 18.27

This gives us a χ2-statistic of 18.27. Does this indicate that our distribution of heart attacks is far away
from one that is equally distributed? To answer this, we need to know the values of typical χ2-statistics if
there is just noise in the data and not a signal. This way we can determine whether our observed chi-
square value is fairly small (just noise) or fairly large (evidence of a signal).

A simulated distribution of typical χ2-statistics (with just some noise)

We now need to determine typical values of the χ2-statistics that would occur if our null hypothesis was
true. This will let us see the noise (or variability) that these statistics have and help us determine
whether or our statistic of 18.27 fits in with this noise. It is probably a good time to remind you of the
null hypothesis, so we write it again below.

H0: Heart attacks are distributed evenly across the days of the week for employed people in the
population.

We want to take the 884 heart attacks and randomly distribute them across the 7 days of the week. To
do this we can think of what we did earlier with rolling a die 120 times. This is the same process. Imagine
a 7-sided die where each side had a day of the week on it instead of a number. For each of the 884 heart
attacks, we would roll the die to randomly determine which day it occurred. The complete process to
develop a χ2-statistic under the assumption that the null is true would be to:

1. Roll the die, note day of the week that is facing up.
2. Repeat this for a total of 884 times.
3. Develop a distribution of these 884 heart attacks (similar to our observed data back in Table 1).
4. Calculate the χ2-statistic from the randomized data.
5. Repeat this many times (like 1,000) so we can see typical values of the statistic under a true null.

We don’t have a 7-sided die and if we did, you can see that this process would be quite time consuming.
However, technology can come to our aid. We have a computer applet that can simulate this process
very quickly. We did this simulation in our applet five times and came up with the five distributions and
their χ2-statistics shown in Table 3.

Day Sun Mon Tu Wed Thur Fri Sat χ2

e
Trial 127 124 133 125 125 12 128 0.60
#1 2
Trial 132 142 124 107 128 13 121 5.55
#2 0
Trail 119 127 118 142 108 12 146 8.69
#3 4
Trial 126 131 129 118 121 13 128 1.20
#4 1
Trial 123 128 130 117 126 13 125 1.52
#5 5
 Table 3: Five simulated distributions giving how many heart attacks occurred on each day of the
week for the 884 heart attacks under the assumption that they are distributed evenly in the
population. The accompanying χ2-statistics are also given.

With our five simulated distributions and the resulting χ2-statistics we haven’t seen anything as extreme
as our observed χ2-statistic of 18.27. We should do some more repetitions, so we can better see the
noise in the resulting statistics. We had our applet do 95 more simulations for a total of 100. Each
resulting χ2-statistic is plotted in a graph as shown in Figure 1.

Figure 1: A distribution of 100 simulated χ2-statistics under the assumption that the null
hypothesis is true.

It doesn’t appear from the graph of the 100 simulated χ2-statistics shown in Figure 1 that any of the
simulated statistics are as large as our observed value of 18.27. At this point it looks like a statistic as
extreme as 18.27 would rarely happen by chance if the null hypothesis is true. We should still probably
do more simulations. We had our applet do 900 more simulations for a total of 1,000. The resulting χ2-
statistics are collected and shown in the distribution in Figure 2.
Figure 2: A distribution of 1,000 simulated χ2-statistics under the assumption that the null hypothesis is
true.

From the distribution shown in Figure 2 we can see that a χ2-statistic as extreme as 18.27 (our observed
statistic) is quite unlikely to occur. In only 6 out of the 1,000 repetitions did we get something at least as
extreme. So what does this tell you? Is 18.27 part of the noise or is it sending us a signal that something
different is going on? In other words, does 18.27 seem to be far enough out in the tail of this distribution
that you would say it is unlikely to occur by chance?

The standard that is typically used to say something is unlikely to occur by chance is a probability of less
than or equal to 0.05 or 5%. From our applet output, we can see that we estimate that a χ2-statistic at
least as extreme as 18.27 happened about 0.6% of the time. Because this is less than 5%, we would
conclude that there is strong evidence this observed statistic arose from something other than just
noise. In other words we would say we have strong evidence against the null hypothesis that the
population distribution is distributed evenly and say we have strong evidence for the alternative
hypothesis that the population distribution is not distributed evenly.

Could a statistic like 18.27 happen just by chance if the population was distributed evenly? Yes, it
happened 6 times in our 1,000 repetitions. However, a probability of 6/1,000 means it would be very
unlikely to get a χ2-statistic at least as large as 18.27 if the population is equally distributed. Therefore it
is more plausible that the population is not equally distributed.

Theory-based p-values
The probability of 0.006 that was given to us in the applet is called a p-value. More generally, a p-value
is the probability of obtaining a value of the statistic at least as extreme as the observed statistic when
the null hypothesis is true. We obtained our p-value of 0.006 through a simulation. If we repeated the
simulation again we might get a slightly different p-value. (In fact we just repeated the simulation and
obtained a p-value of 0.008). It shouldn’t be too concerning that we might get slightly different p-values,
because they all would be fairly close together and should all be telling us the same thing in terms of
strength of evidence.

Before computers could quickly do simulations like we just did, the way to get p-values had to rely on
theory-based methods. In theory-based methods the simulated distribution that we obtained in the
applet is predicted using mathematical formulas. The theory-based distribution that we would need to
use in this example is appropriately called a χ2-distribution. The applet we were using earlier can both
show a picture of the theory-based χ2-distribuiton and compute the theory-based p-value using this
distribution. All this is shown in Figure 3.

Figure 3: A theory-based χ2-distribution is overlaid on our simulated χ2-distribution and a theory-

based p-value is calculated in the applet.

Notice that we get a very similar p-value using theory-based methods. As long as the sample size is large
enough this will happen. A theory-based χ2-goodness-of-fit test (this is the name of the test we have just
used) to give a valid p-value the expected frequencies should all be at least 5. Remember that our
expected counts were all 126.3, so our sample size was definitely large enough to get a valid theory-
based p-value.

Remember back when we were developing our statistic and we decided to square the differences
between the observed and expected frequencies? If we only use simulation techniques to get a p-value,
using absolute values instead of squaring would have been fine. Simulation-based techniques can use a
wide variety of statistics to obtain a p-value. The theory-based techniques can be a bit more finicky,
however. Using absolute values could result in a simulated distribution in which the heights of the bars
go up, then down, then back up again. They don’t tend to be as smooth as when we square the
differences. This makes it hard to create a mathematical model that will fit a simulated distribution
based on absolute values. Squaring doesn’t lead to this kind of problem.

Scope of conclusion
With a p-value of 0.006 (or 0.0056) we have strong evidence against the null and hence strong evidence
that heart attacks are not distributed equally across the seven days of the week in the population. But
what exactly is this population we keep referring to?

If our data were obtained from a random sample of all heart attacks from some specific population (like
all German citizens in 1990) our population would be the group from which our sample was drawn.
However, in many studies, like this one, the sample is not a random sample. In this case, the researchers
collected data from 13 hospitals in and around Augsburg, Germany from 1985 to 1990. In particular,
they looked at all the heart attacks that occurred in that period and found which day of the week they
occurred. Furthermore, for this specific data set, they focused on heart attacks from people that were
employed during that time.

Since this is not a random sample can we generalize our results to a larger population? We can, though
we might do so with some hesitation as to exactly what that population is. Even though this wasn’t a
random sample it is probably representative of some larger populations. We could probably be
comfortable generalizing these results to all of Germany (or what was commonly called West Germany
at the time of this study) since heart attacks in Augsburg are probably very similar to those around the
rest of the country. Can we generalize to all of Europe? All industrialized nations? The entire world? Just
how far you go depends on what populations you think are quite similar (in terms of having heart
attacks) to those in and around Augsburg, Germany in the late 1980s.

What we need to emphasize here is that these heart attacks were from people that were employed. We
saw that the largest frequency of heart attacks took place on Mondays. Might this also be true for those
that are not employed? Are Mondays “special” for that population? We will have you look at this
question with another data set in just a bit.

Review
Let’s review what we just did.

 We (or the researchers) wanted to see whether or not heart attacks were distributed evenly
across the seven days of the week.
 In setting up a test of significance we first wrote out the null and alternative hypotheses:
o H0: Heart attacks are distributed evenly across the days of the week for employed
people in the population.
 o Ha: Heart attacks are not distributed evenly across the days of the week for employed
people in the population.
 The researchers collected data and from that we computed a χ2-statistic of 18.27. This statistic
measures how much the observed distribution differs from a distribution of the same sample
size that has the heart attacks distributed equally across the seven days of the week.
 We then used an applet to find a p-value of 0.006. What does this number mean? Remember
that we simulated χ2-statistics for this scenario under the assumption that the heart attacks are
distributed evenly in the population. Only 6 of the 1,000 simulations resulted in χ2-statistics that
were at least as large as the 18.27 that was observed in the study. This means that if heart
attacks were distributed evenly across the seven days of the week in the population, it is very
unlikely that we would get a χ2-statistic as large or larger than we did. (P-values less than 0.05
are considered small and ours was much smaller than 0.05.)
 We also found a theory-based p-value of 0.0056 which guides us to the same conclusion as did
the simulation-based p-value.
 We then concluded that based on a p-value of 0.006, we have strong evidence that heart attacks
in the population are not distributed evenly across the seven days of the week.

Instructions for using the Analyzing One-way Tables Applet

The applet we used can be found at: http://www.rossmanchance.com/applets/GOF.html. To find a p-

value with our heart attack data do the following.

1. Open up the applet and put the heart attack table of data like it is shown below. Then click on
Use Table and you should then see a bar graph of the data like the one shown below.
2. Below the table of data, click on the arrow in the box next to Statistic and change it to the χ2-
statistic as shown.

3. Click on the Show Sampling Options check box on the right side of the applet. If heart attacks
were distributed evenly, there should be 1/7 of them on each day. So for the hypothesized
probabilities of heart attacks you need to enter 1/7 written as a decimal 7 times separated by
commas as shown below.

4. Put 1000 in the Number of Samples box and click on the Sample button. The applet has now
created 1000 distributions, calculated its χ2-statistic of each of them, and then plotted each
statistic in a graph.
 5. Put the χ2-statistic (which you should see on the lower left side of the applet) into the box as
shown below and click on Count. You should now see a simulation-based p-value in red. It may
not match the one shown below exactly, but it should be similar.

6. To calculate the theory-based p-value simply click on the Overlay Chi-square distribution box and
you will see the theory-based p-value (which should be exactly the same as shown below) as well
as a smooth distribution over the top of the simulated distribution.
What about heart attacks for those not employed?
We’ve seen that there is strong evidence that heart attacks are not distributed evenly across the seven
days of the week for those that are employed. In particular, our data showed that Monday had the
largest number by quite a bit. Since this data set came from people that are employed, maybe Mondays
are particularly difficult for this group. What about people that are not employed? Is there the same
Monday effect? We have data on that as well.

The same researchers found that there were 1,191 heart attacks among those that were not employed.
The heart attacks in this group were distributed as shown in Table 4. Again we ask the same question as
earlier. Do we have strong evidence that heart attacks are not distributed evenly across the days of the
week in this population? To answer this, we ask you to answer the following questions.

Day Sun Mon Tue Wed Thur Fri Sat

Number of heart attacks 168 180 157 169 180 16 168
9

Table 4: A distribution of the frequency of heart attacks for 1,191 people that were not
employed at the time of their heart attack.

1. Write the null and alternative hypotheses for this test.

2. Put the table of data into the applet. Based on the distribution of heart attacks, do you think
there will be strong evidence that heart attacks are not distributed evenly across the seven
days of the week? Why or why not?
3. Calculate the χ2-statistic using the applet. How does this statistic compare with the 18.27 from
the employed data? What does this difference mean?
4. Calculate both a simulation-based p-value and theory-based p-value using the applet. Are these
the types of numbers you would expect based on what the distribution looked like and the size
of the χ2-statistic?
5. Write out a conclusion in the context of this situation. (Note: When we have a small p-value we
say we have strong evidence against the null hypothesis or strong evidence for the alternative.
When we get a large p-value we don’t have strong evidence of anything. Instead, we say we do
not have strong evidence against the null or we do not have strong evidence for the alternative
in general.)
Employed Data Table Unemployed Data Table

Day Count Day Count

Sun 106 Sun 168
Mon 160 Mon 180
Tue 123 Tue 157
Wed 115 Wed 169
Thu 141 Thu 180
Fri 107 Fri 169
Sat 132 Sat 168