Some Key Comparisons between Statistics and Mathematics

and Why Teachers Should Care

Allan Rossman1, Beth Chance1, Elsa Medina2

Cal Poly – San Luis Obispo

1
 Statistics is a mathematical science. That sentence is likely to be the shortest in this entire

article, but we want to draw your attention to several things about it:

We use the singular “is” and not the plural “are” to emphasize that statistics is a field of

study, not just a bunch of numbers.

We use “mathematical” as an adjective, because while statistics certainly makes use of

much mathematics, it is a separate discipline and not a branch of mathematics.

We use the noun “science,” for statistics is the science of gaining insight from data.

In this article we highlight some of the differences between statistics and mathematics, and

we suggest some implications of these differences for teachers and students. We realize that

these distinctions may not be universal but hope our broad strokes can help highlight some

fundamental distinctions in the disciplines. Our aim is not to provide a philosophical discussion

of these issues but rather to present ideas that will inform classroom practice. Toward this end

we motivate and illustrate our points with concrete examples.

Crucial Role of Context

A primary difference between the two disciplines is that in statistics, context is crucial.

Mathematics is an abstract field of study; it exists independently of context. Mathematicians

often strive to “strip away” the context that can get in the way of studying the underlying

structure. For example, one can study linear functions for their own mathematical properties,

without considering their applications. Indeed, one could argue that worrying about the

complications of real data diverts students’ attention away from the underlying mathematical

ideas. But in statistics, one cannot ignore the context when analyzing data. Consider the

following dotplot:

2
This plot reveals virtually nothing, it’s just a bunch of dots! Granted, we can see that one dot

appears far to the left of the others, and there is a cluster of six dots less far to the left that seems

to be separated from the majority of the dots. But we can’t interpret the plot or draw any

conclusions from it. Let’s include a scale:

This is better, and we now know that the outlier is at 120, and the lower cluster is between 150

and 160. But we still can not gain any insights from this display. Let’s include the units:

Now we know that this graph reveals weights in pounds, but we still do not know whether we’re

analyzing weights of statistics students or of horses or furniture or …

Finally, we’ll tell you the context: these are the weights of the rowers on the 2004 U.S. men’s

Olympic rowing team. Armed with that knowledge, we can summarize what the dotplot reveals

and even suggest some explanations for its apparent anomalies. It makes sense that the data

include an outlier who weighs much less than the others: he’s the coxswain, the team member

who calls out instructions to keep the rowers in synch but does not put an oar in the water

himself. He needs to be light so as not to add much weight to the boat. The cluster in the 150’s

also has an explanation: those six rowers participate in “lightweight” events with strict weight

3
limitations, two in a pairs event and the rest in a fours event. As for the rest, the majority of the

rowers are big, strong athletes in the 180-230 pound range.

For another example of how context is paramount in statistics, consider the following

scatterplot of data from a study about whether there is a relationship between the age (in months)

at which a child first speaks and his/her score on a Gesell aptitude test taken later in childhood

(found in Moore and McCabe, 1993). We have drawn the least squares regression line on the

plot:

This scatterplot and line reveal a negative association between the variables, indicating that a

large value of one variable tends to appear with a small value of the other. Moreover, the slope

of the line is statistically significant (p-value = .002, R2 = .410). But on closer inspection, we see

that this apparent negative association is driven largely by the two extreme cases in the bottom

right of the plot. What do we make of this? Should we conclude that there’s a negative

association here or not? Should we discount the outliers or not? Of course, to answer this

question, we must first consider the context. We see that those two outliers are exceptional

children who take a very long time to speak (3.5 years and a bit longer than 2 years) and who

also have very low aptitude as measured by Gesell. To get a sense for whether the negative

4
association between speaking age and aptitude score holds for more “typical” children, we can

delete the outliers and re-fit the line:

This scatterplot and line reveal essentially no association between the variables. The slope

coefficient is no longer statistically significant (p-value = .890, R2 = .001).

So, our conclusion here contains two parts: Children who take an exceptionally long time to

speak tend to have low aptitude, but otherwise there is virtually no relationship between when a

child speaks and his/her aptitude score. (We should learn more about how these data were

collected before deciding whether these results generalize to a larger group, and we could gather

more data to examine whether those two exceptional children are indicative of a larger pattern.)

Once again, the context drives our analysis and conclusions. Fitting a line to these data without

considering the context would have blinded us to much of what the data reveal about the

underlying question of speaking and intelligence.

Issues of Measurement

Another important issue that distinguishes statistics from mathematics is that measurement

issues play a large role in statistics. Measurement is also important in mathematics; in fact, it is

one of the standards in the NCTM Principles and Standards for School Mathematics (NCTM,

2000), but the focus is different. In mathematics, measurement includes getting students to learn

5
about appropriate units to measure attributes of an object such as length, area, and volume and to

use formulas to measure those attributes. In statistics, drawing conclusions from data depends

critically on taking valid measurements of the properties being studied. Measuring a rower’s

weight is quite straight-forward, but measuring a child’s aptitude is quite challenging. Many

other properties of interest in statistical studies of human beings are hard to measure accurately;

examples include unemployment, intelligence, memory ability, and teaching effectiveness.

Another example of a hard-to-measure property comes from a fascinating recent study

involving a city’s pace of life. Researchers studied whether a city’s “pace of life” is associated

with its heart disease rate (Levene, 1990, as found in Ramsey and Schafer, 1997). They

measured a city’s pace of life in three different ways:

Average walking speed of pedestrians over a distance of 60 feet during business hours on

a clear summer day along a main downtown street;

Average time a sample of bank clerks take to make change for two $20 bills or to give

$20 bills for change;

Average ratio of total syllables to time of response when asking a sample of postal clerks

to explain the difference between regular, certified, and insured mail.

The following scatterplots reveal that there is a slight positive association between heart rate and

these three measures of a city’s pace of life:

6
Importance of Data Collection

One can study and do mathematics without analyzing data, but even when mathematicians

examine data they typically focus on detecting and analyzing patterns in the data. How the data

were collected is not relevant to purely mathematical analyses, but this is a crucial consideration

in statistics. The design of the data collection strategy determines the scope of conclusions that

can be drawn. Can you generalize a study’s results to a larger population? It depends on

whether the sample was randomly selected. Can you draw a cause-and-effect conclusion from a

study? It depends on whether it was a controlled, randomized experiment.

For example, consider two studies that asked women: “Do you give more emotional support

to your husband or boyfriend than you receive in return?” In study A, 96% of a sample of 4500

women answered “yes,” but in study B 44% of a sample of 767 women answered “yes.” How do

we reconcile these results? In which study do we place more confidence for representing the

beliefs of all American women? The answers depend on how the data were collected. Study A

was conducted by sociologist Shere Hite, who distributed over 100,000 questionnaires through

women’s groups (Hite, 1987). Study B was conducted on a random sample of women,

sponsored by ABC News and the Washington Post (Moore, 1992). Even with the smaller

sample, study B provides more credible data because it involved a random sample.

For another example, consider two studies A and B that involve comparing “success” rates

between two groups. The data are summarized in the table, including the p-value for comparing

the two success rates with each study:

Study Group Successes Trials Success rate p-value

7
The methods used for calculating these p-values are identical, and their values are very close.

The small p-values indicate strong evidence of a statistically significant difference in success

rates between the two groups. But does that mean that we draw identical conclusions from the

two studies? The answer depends on how the data were collected.

Study A is a social experiment in which three- and four-year-old children from poverty-level

families were randomly assigned to either receive pre-school instruction or not, with a response

of whether they were arrested for a crime by the time they were 19 years old (as found in

Ramsey and Schafer, 1997). Because the children were randomly assigned to a treatment, and

because the p-value turned out to be so small, we can legitimately draw a cause-and-effect

conclusion between the pre-school instruction and the higher success rate. On the other hand,

study B is an observational study in which researchers examined court records of people who

had been abused or not as children, comparing their rates of committing a violent crime as an

adult (Widom, 1989, as found in Ramsey and Schafer, 1997). Because this was not a

randomized experiment, no cause-and-effect conclusion between the child abuse and the violent

crime rate can be drawn, despite the very small p-value. So while the calculations in two studies

can be identical, the conclusions drawn can differ substantially, depending on how the data were

collected.

Lack of Definitive Conclusions

Statistics and mathematics ask different types of questions and therefore reach different kinds

of conclusions, with different thought processes to link questions to conclusions. Mathematics

involves rigorous deductive reasoning, proving results that follow logically from axioms and

definitions. The quality of a solution is determined by its correctness and succinctness, and there

is often an irrefutable correct answer.

8
 In contrast, statistics involves inductive reasoning and uncertain conclusions. Statisticians

often come to different but reasonable conclusions when analyzing the same data. In fact, within

these types of judgments lies the art of data analysis. All of statistical inference requires one to

use inductive reasoning, as informed inferences are made from observed results to defensible,

but ultimately uncertain, conclusions. In statistics we summarize conclusions with phrases such

as “We have strong evidence that…” and “The data strongly suggest that…” but steadfastly

resist saying things like “The data prove that…”. The quality of conclusions lies in the analysts’

ability to support and defend their arguments.

For example, we often ask students to collect data for comparing prices between two

different grocery stores. This project raises many practical issues of measurement and data

collection, such as whether students should record sale prices or regular prices, and how students

should obtain a random sample of grocery items. Consider some sample data collected by our

students of price differences between the two stores on 37 items:

How should we analyze these data? One reasonable approach is to calculate the mean of

these differences and test whether it differs significantly from zero; a t-test yields a p-value of

.308. Thus, the mean price difference in this sample does not differ significantly from zero. But

we have not established that the average price is the same between the two stores; we have only

concluded that the sample data do not provide compelling evidence to reject that hypothesis.

Not only is this conclusion uncertain, but we could have selected a different analysis

altogether. We could instead perform a sign test of whether the median price differs significantly

9
from zero, which is equivalent to asking whether the proportion of items costing less in one

particular store differs significantly from one-half. Twenty-one of the items are cheaper in one

store, with only eight items cheaper in the other store (and eight “ties”). The p-value for this test

turns out to be .024, which does suggest a statistically significant difference. But even this

conclusion is uncertain, for the p-value reveals that sample data this extreme could have arisen

by chance even if there was no difference between the median prices in these stores.

So, do the stores’ prices differ significantly, as the sign test suggests, or not, as the t-test

suggests? Which of these two conflicting conclusions is correct? Which is reasonable? Well,

neither and both. It depends on what question we want to ask (about the mean or median), and

even then neither conclusion is certain. To complicate matters still further, statisticians might

disagree on whether a normal model for the price differences is reasonable enough to justify

applying the t-test, and statisticians might also disagree on how the “ties” should be handled

when conducting the sign test. This lack of definitive conclusions, and even the lack of a single

appropriate method of analysis, is common in statistical investigations.

Communicating Statistical Knowledge

Terminology is essential in mathematics as well as statistics, but one difference is that many

common terms from everyday language have technical meanings in statistics. Examples include

words such as bias, sample, statistic, accuracy, precision, confound, correlation, random, normal,

confident, and significant. Students are very tempted to use these words loosely, without

considering their technical meanings. Rather, studying statistics is akin to studying foreign

language, for students need lots of practice to become comfortable using these terms correctly,

and they often stumble at first before acquiring enough familiarity to use the language well.

Although there are English terms such as “multiple” and “factor” with a technical meaning in

10
mathematics, it seems that students expect to see more technical terms in mathematics than in

statistics.

Communication is essential in both disciplines, but more so in statistics because it is a

consulting enterprise. Statisticians routinely must interact with clients whose technical skills

vary greatly, from eliciting a clear statement of the problem from those clients through

communicating to them the results and conclusions of the analysis. While introductory students

are far from professional statisticians, the ability to communicate statistical ideas in layperson’s

terms is essential and an important component of many courses. Communication is important in

mathematics also, but that communication is more often done symbolically in mathematics.

Why Should Teachers Care?

We have argued that statistics is a different discipline than mathematics, that it involves a

different type of reasoning and different intellectual skills. Even if you find our case persuasive,

the question remains: Why should classroom teachers care? We see two primary reasons:

A different type of instructional preparation is needed for teaching statistics.

Students react differently to statistics than to mathematics.

In order to help students see the relevance of context, measurement issues, and data

collection strategies in statistics, it’s imperative that teachers present real data, in meaningful

contexts, from genuine studies. Fortunately, there are a plethora of resources available now to

help teachers with this, from books to CD-ROMs to websites (Moore, 2000).

Instructors also need to help students learn to relate their comments to the context and to

always consider data collection issues when stating their conclusions. While this is sometimes

done in mathematics courses, it’s not nearly as prevalent or as essential as it is with statistics.

Indeed, this need for different types of assessments is a key difference between teaching statistics

11
and mathematics. Many students do not initially expect this type of focus and teachers need to

be prepared for students’ discomfort. Students also need to be reminded that there are multiple

correct approaches, and that they will also be evaluated on how well they explain and support

their conclusions.

Another difference in terms of instructional preparation is that many teachers do not have

ample opportunities to develop their own statistical skills and understanding of statistical

concepts before teaching them to students. This challenge is especially acute because few

programs in mathematic teacher preparation offer much instruction in statistics, and much of the

instruction that is provided concentrates on the mathematical aspects of statistics. The recent

Mathematical Education of Teachers report makes these points quite forcefully (Conference

Board of the Mathematical Sciences, 2001). Helping students to develop their communication

skills and statistical judgment, so crucial in the practice of statistics, is also very challenging and

an area in which mathematics teachers are provided with little instruction.

The experiences and reactions of students to studying statistics are different from studying

mathematics. Educational research shows that students (and others) have tremendous difficulties

with reasoning under uncertainty (Garfield, 1995; Shaughnessy, 1992; Garfield & Ahlgren,

1988). Also, many students (and others) are very uncomfortable with uncertainty, with the lack

of definitive conclusions, and with the need for detailed interpretations and explanations that are

integral to studying statistics. Helping students to develop a healthy skepticism about numerical

arguments, without allowing them to slip to the extremes of cynicism or naïve acceptance, is a

great challenge.

Because of the differences between statistics and mathematics, teachers should expect that

some mathematically strong students may be frustrated while studying statistics. But on the

12
bright side, many students who may not be initially excited by mathematics will be intrigued and

empowered by their experiences with statistics.

References

Conference Board of the Mathematical Sciences. The Mathematical Education of Teachers.

Providence, RI and Washington, DC: American Mathematical Society and Mathematical

Association of America, 2001.

Garfield, Joan. “How Students Learn Statistics.” International Statistical Review 63 (1995): 25-

34.

Garfield, Joan, and Andrew Ahlgren. “Difficulties in Learning Basic Concepts in Statistics:

Implications for Research.” Journal for Research in Mathematics Education 19 (1988): 44-

63.

Hite, Shere. Women and Love: A Cultural Revolution in Progress. New York: Alfred A. Knopf,

1987.

Levene, R. V. “The Pace of Life.” American Scientist 78 (1990): 450-459.

Moore, David S. and George P. McCabe. Introduction to the Practice of Statistics (2nd ed.).

New York: W.H. Freeman and Company, 1993.

Moore, David W. The Super Pollsters. New York: Four Walls Eight Windows, 1992.

Moore, Thomas L., ed. Teaching Statistics. Washington: Mathematical Association of America,

2000.

National Council of Teachers of Mathematics. Principles and Standards for School

Mathematics. Reston, VA: National Council of Teachers of Mathematics, 2000.

Ramsey, Fred and Daniel Schafer. The Statistical Sleuth: A Course in Methods of Data Analysis.

Pacific Grove, CA: Duxbury, 1997.

13
Shaughnessy, Michael. “Research in Probability and Statistics: Reflections and Directions.” In

Handbook of Research in Mathematics Teaching and Learning, edited by Douglas Grouws,

pp. 465-494. New York: Macmillan, 1992.

Widom, C.S. “The Cycle of Violence.” Science 244 (1989):160.

14