Academics 1

Developing hypotheses and determining their

testability and plausibility
Peter Zoogman, Ph.D.
Last modified: 9/24/20

Learning about the world

Science is the process of trying to understand how the world works via observation and
experimentation. While there are currently many widely-accepted theories describing the world,
one of the core tenets of science is that we never reach the absolute truth. There is always
more to learn or modifications needed to be made to existing theories. A hypothesis sits on the
boundary between the known and the unknown: an informed conjecture about a process that
drives physical, natural, or social phenomena. You can use hypotheses to direct research into
topics as diverse as: the spread of coronavirus, judicial biases, unequal impacts of climate
change, and sporting success.

A good hypothesis proposes a causal mechanism of how the world works, a mechanism that we
don’t yet know is true, but we have reason to believe that it may be true. This hypothesis needs
to be testable by leading to specific predictions of measurable variables whose values will
support or refute the hypothesis.

Many scientists, including myself, became scientists because we love solving problems and
puzzles. Developing a hypothesis and the research to test it has many similarities to the
problem solving process with which you are already familiar. We start from an initial state of
existing theory and available data. We will then have to identify gaps in the knowledge to help
our creative process come up with new hypotheses. And to refine our hypotheses and propose
research to test them, we will consider technical and ethical constraints and use heuristics.

However, it can be particularly challenging to define a concrete goal state for a research project.
This is because research should be probing the boundary between the known and the unknown,
and we don’t know what it is that we are going to learn by the end of the process. What we can
define in place of a goal state is what potential results could result from our proposed research,
and what each of these potential alternatives would mean for our hypothesis — thus telling us
what we can expect to learn about how the world works.
Academics 2

Hypothesis Development

A hypothesis, at its core, describes a process or mechanism. Really any process: it could be
physical, biological, psychological, political, or something else. One key feature is that it has
some mechanistic component, which distinguishes a hypothesis from a prediction. Another key
feature is that we are not sure if a hypothesis is true even though there exists some supporting
evidence or logic.

If more people wear masks, the spread of COVID-19 will be reduced, because the novel
coronavirus is primarily transmitted by respired droplets. The ‘because’ part of this sentence is
the core of the hypothesis: ‘the novel coronavirus is primarily transmitted by respired droplets’.
This is a statement of why we think the world works the way it does. The ‘if-then’ part is one
prediction that comes from the hypothesis — this will be key in understanding the testability of
the hypothesis.

Here’s another example: the ‘hot hand’. Originally stemming from sports, those who believe in
the ‘hot hand’ state that after someone achieves multiple successes (such as making shots in
basketball) they are more likely to succeed in following attempts. You might already be
evaluating this statement in terms of how plausible you find it based on your previous
experience, but we’ll get to plausibility and testability in the next section. First, is this statement
a hypothesis? Notice that it is actually a prediction — there’s not yet a mention of the proposed
mechanism behind this statement. The implied hypothesis is something like: ‘people can enter a
psychological state where they have an increased chance of success’. Whether or not we think
this a strong hypothesis is something we can examine later, the first step is to make sure our
hypothesis statement is truly a conjecture about a process and not merely a prediction.

A hypothesis has to propose a causal mechanism about which there is uncertainty — coming up
with a new hypothesis takes creativity! Still, there are two main pathways to a new hypothesis:
generalizing from patterns in data or building off of existing theories.

Whenever you look at a data visualization you want to understand the patterns in the data.
Some of these patterns may be easy to explain based on what you already know about existing
theory; this is nice, but not particularly interesting. More interesting is when you find a pattern in
the data that doesn’t quite fit with what you know or expect. Aha! Can you now come up with a
potential explanation for why the data looks like this? Perhaps a mechanism that isn’t yet fully
understood or accepted: a new hypothesis!
Academics 3

Look at Figure 1. As always, start by identifying the variables being plotted and what type of
visualization this is. Then, what are the patterns? Are there any that surprise you? Take a
minute to answer these questions before you continue reading.

Figure 1: Proportion of rulings in favor of the prisoners by ordinal position. Circled points
indicate the first decision in each of the three decision sessions; tick marks on the x axis denote
every third case; dotted line denotes food break. Because unequal session lengths resulted in a
low number of cases for some of the later ordinal positions, the graph is based on the first 95%
of the data from each session. Reproduced from Danziger et al. (2011).

When researchers saw this data, they observed that judges appeared to give more favorable
decisions soon after a food break or at the beginning of the day (Danziger et al., 2011). One
hypothesis they proposed was: ‘eating restores the judges’ mental resources’. Note that this
hypothesis is more specific than stating, for example, that the rulings are impacted by
physiological factors not related to specific cases. You want to make your hypothesis as specific
as possible. Once this hypothesis about the impact of food on human decision making was
formed from this initial data set, it was then used to make subsequent predictions and tested by
other researchers (Glöckner, 2016; Kerry et al., 2019). Not all of the subsequent studies yielded
results that supported the initial hypothesis nor did they uniformly reject it: this is still an open
research question.

The other method of developing a hypothesis (as opposed to examining patterns in data) is
deduction from existing theories. Since existing theories are well-accepted, if we can make
logical deductions from what scientists already accept, we should be able to come up with new
statements that also describe the world. However, many of the steps in this process require
Academics 4
additional assumptions, so even though the process is deductive, the end result is still a
conjecture that needs to be tested because we don’t know the truth value of the assumptions.

For example: ‘In the United States, communities of color are more vulnerable to climate change
due to neighborhood characteristics’ (Lakhani, 2019). This hypothesis was developed
deductively from existing theories and facts, including: both mean and extreme temperatures
will increase over time due to the greenhouse effect, communities of color are already
disproportionately affected by heat waves, and communities of color have fewer resources to
adapt to changes in temperature. Going from the accepted theories to the new hypothesis
requires certain assumptions: for one, we assume that the resources available to these
communities will continue to be limited. Evaluating the strength of assumptions that go into a
hypothesis is a key measure of its plausibility. By combining what we already accept (theory)
with certain reasonable assumptions, we can develop new hypotheses.

Whichever path you follow to reach your hypothesis, that is only the first step. You next need to
evaluate it for testability and plausibility; once you have a hypothesis that you believe is strong,
you can then propose research to investigate whether it really does explain how the world
works.

Table 1 - Examples of hypotheses from different fields, using different written formats.
Example 1: Novel coronavirus transmission

Hypothesis If more people wear masks, then the spread of COVID-19 will be reduced,
because the novel coronavirus is primarily transmitted by respired droplets.
Note: This hypothesis follows an ‘If-Then-Because’ format.

Example 2: Hot hand

Hypothesis Athletes get a ‘hot hand’ (i.e. go on a streak of consecutive successes)

because the previous success produces a psychological state that improves
subsequent performance.
Note: This hypothesis is not in the ‘If-Then-Because’ format, but qualifies as
a good hypothesis because it proposes an explanation. Increasing the
specificity of ‘psychological state’ could potentially improve this, but for our
current purposes it is sufficient.

Example 3: Judge’s parole decisions (drawn from Danziger et al., 2011)

Hypothesis Eating restores judges’ mental resources.

Note: Other parts of the paper clarified and specified the term ‘mental
Academics 5

resources’.

Example 4: Impacts of climate change (drawn from Lakhani, 2019)

Hypothesis In the United States, communities of color are more vulnerable to climate
change due to neighborhood characteristics.
Note: This does not follow the ‘If-Then-Because’ format, nor does it use the
word ‘because’ to indicate the explanation.

Hypothesis Evaluation

After you’ve developed a hypothesis, it is critical to evaluate its testability and plausibility. If you
find it is weak in one or both of these areas, it is likely you will need to iterate on your initial
hypothesis in order to strengthen it. This way, you maximize the opportunity for research
investigating your hypothesis to contribute to overall knowledge. And you can use these tools to
evaluate others’ hypotheses as well as your own.

Testability

Scientific hypotheses must be testable. Since a hypothesis is a potential process underlying the
way the world works, you can almost never directly examine the validity of a hypothesis.
Instead, scientists deduce predictions that would be true if the hypothesis was true. These
predictions are often contained in an ‘if-then’ statement associated with the ‘because’ of the
hypothesis: “If more people wear masks, the spread of COVID-19 will be reduced” is a
prediction that followed from our hypothesis about the novel coronavirus transmission pathway.

Once you have a prediction, you can use it to test the hypothesis. To do this, you need to be
able to measure the variables that are part of the prediction. “If more people wear masks, the
spread of COVID-19 will be reduced”: the variables here are the prevalence of masks
(quantified at the percentage of the population wearing them) and the spread of COVID-19
(quantified as the number of new cases per day). Note that to demonstrate that you can test
your hypothesis, you need to concretely describe the relevant variables: it is measurements of
these variables that will go on to tell you what you’ve actually learned about your hypothesis at
the end of a research study. You can imagine multiple alternative outcomes; alternatives where
your prediction holds true and alternatives where it doesn’t. If the outcome of your study is
consistent with the prediction, that supports your hypothesis (but doesn’t prove it, that would be
Academics 6
affirming the consequent). If the outcome of your study isn’t consistent with the prediction, that
contradicts your hypothesis (by Modus Tollens).

A good hypothesis should generate multiple predictions — being able to test one hypothesis
with many lines of empirical study is how scientists start to gain more clarity about what is
actually going on. Here’s another prediction: “If we have some hamsters wear (tiny) surgical
masks and have some hamsters not wear masks and expose the hamsters to the novel
coronavirus, the ones wearing masks will have a lower rate of infection.” Notice that the
variables here are related to the ones above, but are easier to measure! Investigating this
prediction in a lab (c.f. Chan et al., 2020) would provide another piece of empirical evidence that
either supports or contradicts our initial hypothesis.

What if one or more of the variables in your prediction aren’t measurable? This could be due to
constraints of logistics, technology, or ethics. Maybe you’ve made a prediction about the future
but left your time machine at home, or your prediction would involve potential harm to a human
subject. But even if constraints prevent actually testing a prediction, the process of making
predictions from a hypothesis is critical. If a hypothesis leads to predictions that have
alternatives that can either support or contradict the hypothesis, then the hypothesis is testable
in principle. If not, then the hypothesis isn’t testable at all! At this point, we could never learn
about our hypothesis using empirical data, and we’re not really doing science anymore. But if
we can make predictions that can be falsified, then there exists a path for gaining knowledge
about whether the hypothesis might actually be true. In the case where we can measure the
variables in the prediction then the hypothesis is testable in practice. (Notice that a hypothesis
that is testable in practice must also be testable in principle.)

Plausibility

A good hypothesis should be plausible as well as testable. At its core, this requirement means
that a hypothesis should ‘make sense’. More concretely, a hypothesis is plausible when it is
consistent with existing theories, when it requires a limited number of assumptions, and when
those assumptions are well-founded. In addition, a hypothesis is also more likely to be true
when there are few or no competing hypotheses. This last criterion can be considered part of
the contextual plausibility of the hypothesis.

Let’s return to the example of physiological impacts on judicial rulings: judges’ rulings are
impacted by hunger and eating. Remember that this hypothesis was generated inductively from
preliminary data. This is consistent with existing theories of biases; we’ll even return to this
example next semester when we learn about bias identification. This hypothesis assumes that
judges’ are not able to successfully mitigate their own biases, an assumption that is perhaps
Academics 7
worrisome but also supported by previous research (Guthrie et al., 2007). Danziger et al. (2011)
also proposed a competing hypothesis that the mere act of taking a break was the cause of the
judges’ change in behavior. They discarded other competing hypotheses, including that the
rulings in this dataset were driven by prisoner characteristics, but found this was implausible as
it was not consistent with the existing data. Another potential alternative was that this pattern
was due to random chance; a statistical analysis could tell us how likely that is to be the case.

If you analyze a hypothesis and find that it is implausible, this is an opportunity to go back to the
hypothesis and revise it to make it more plausible. Scientists do this all the time. The
requirement of plausibility might appear to stifle exploration beyond the bounds of what is
already known, but that shouldn’t be the case. What is key is to recognize how plausible a given
hypothesis is, because the less plausible it is, the stronger the evidence must be (and the
greater the number of separate predictions and tests needed) to support it. In the words of Carl
Sagan: “Extraordinary claims require extraordinary evidence”.

Table 2 - Hypotheses from different fields with examples of predictions and assumptions.
Example 1: Novel coronavirus transmission

Hypothesis If more people wear masks, then the spread of COVID-19 will be reduced,
because the novel coronavirus is primarily transmitted by respired droplets.
Note: This hypothesis follows an ‘If-Then-Because’ format.

Predictions 1. Predict that countries with higher levels of mask use will have
lower/slower spread of COVID-19.
2. If we have some hamsters wear (tiny) surgical masks and have some
hamsters not wear masks and expose the hamsters to the novel
coronavirus, the ones wearing masks will have a lower rate of
infection.
Note: There are many more predictions you could come up with, and
detailing those will strengthen the #testability of the hypothesis.

Assumptions 1. Assumes that masks reduce respired droplets being either exhaled by
infected people or inhaled by uninfected people.
2. Assumes viral particles are in respired droplets of infected people.
3. The second prediction assumes hamsters get infected by the novel
coronavirus.
Note: Some assumptions are generally associated with the hypothesis,
Academics 8

whereas others may be more specific to a testable prediction (e.g.,

hamsters).

Example 2: Hot hand

Hypothesis Athletes get a ‘hot hand’ (i.e. go on a streak of consecutive successes)

because the previous success produces a psychological state that improves
subsequent performance.
Note: This hypothesis is not in the ‘If-Then-Because’ format, but qualifies as
a good hypothesis because it proposes an explanation. Increasing the
specificity of ‘psychological state’ could potentially improve this, but for our
current purposes it is sufficient.

Predictions 1. After someone makes several shots in basketball, they are more likely
to succeed in following attempts.
2. If we manipulate the time between basketball shots, then we expect
the shots a player takes sooner after making a basket will have a
greater rate of success than shots a player takes long after the shot
they made.

Assumptions 1. Assumes success changes the psychological state of an individual.

2. Assumes the psychological state of an individual is not fixed and
dissipates over time

Example 3: Judge’s parole decisions (drawn from Danziger et al., 2011)

Hypothesis Eating restores judges’ mental resources.

Note: Other parts of the paper clarified and specified the term ‘mental
resources’.

Predictions 1. Parole judges will reach more harsh decisions following breaks where
they can’t eat than after breaks when they eat.
2. If parole judges eat consistently through the day instead of only
during the breaks, their positive parole decisions will not exhibit the
substantial decrease prior to their break.

Assumptions 1. This assumes the mental state of judges (e.g., such as biases) affect
their decision making.
2. This assumes that eating or breaks changes the mental state of
judges.

Example 4: Impacts of climate change (drawn from Lakhani, 2019)

Academics 9

Hypothesis In the United States, communities of color are more vulnerable to climate
change due to neighborhood characteristics.

Note: This does not follow the ‘If-Then-Because’ format, nor does it use the
word ‘because’ to indicate the explanation.

Predictions 1. When hurricanes hit cities, damage will have a greater negative
impact (e.g., reduced food insecurity) on neighborhoods in which a
greater proportion of people of color reside.
2. Predict that neighborhoods in which a greater proportion of people of
color reside have fewer factors that ameliorate heat (e.g., shade trees
or access to air conditioning).
3. Predict that when heat waves hit, people from communities of color
will have a greater incidence of heat-related health problems (e.g.,
heat stroke).

Assumptions 1. Assumes neighborhoods in the US are at least somewhat racially

segregated.
2. Assumes communities of color live in neighborhoods with common
characteristics
3. Assumes hurricane frequency and severity are a result of climate
change.
4. Assumes that trees or air conditioning reduce negative impacts of
extreme heat events.

How can we learn?

These skills that relate to hypotheses will allow you to follow the same path that researchers use
to investigate the world around us. These skills not only let you follow this path, but also
evaluate the strength of others’ research and where more work needs to be done before we
have results that should be accepted.

After this process is completed — developing a hypothesis, evaluating its testability and
plausibility, and examining new empirical evidence — how likely is it that the hypothesis is true?
The likelihood of a hypothesis being true is a combination of how plausible it is, how probable it
would be to get this evidence given your hypothesis (whether it lines up with predictions), and
how directly the evidence bears on your hypothesis.
Academics 10
However, remember when making a hypothesis the goal isn’t necessarily to make one that is
going to be supported by future results. The goal is to create a hypothesis that might be true
and might be false, but either way leads to understanding the world — whichever facet of the
world you care about — a little bit better.

References

Chan, J. F. W., Yuan, S., Zhang, A. J., Poon, V. K. M., Chan, C. C. S, Lee, A. C. Y., et al. (2020).
Surgical mask partition reduces the risk of noncontact transmission in a Golden Syrian Hamster
model for coronavirus disease 2019 (COVID-19). Clinical Infectious Diseases, ciaa644.

Danziger, S., Lehav, J., and Avniam-Pesso, L. (2011). Extraneous factors in judicial decisions.
Proceedings of the National Academies of Sciences, 108, 6889-6892.

Glöckner, A. (2016). The irrational hungry judge effect revisited: Simulations reveal that the
magnitude of the effect is overestimated. Judgment and Decision Making, 11, 601-610.

Kerry, N., Loria, R. N., Murray, D. R. (2019). Gluttons for punishment? Experimentally induced
hunger unexpectedly reduces harshness of suggested punishments. Adaptive Human Behavior
and Physiology, 5. 352-370.

Lakhani, N. (2019, October 21). ‘Racism dictates who gets dumped on’: how environmental
injustice divides the world. The Guardian.
https://www.theguardian.com/environment/2019/oct/21/what-is-environmental-injustice-and-why
-is-the-guardian-covering-it