Transcript Null hypothesis - People Server at UNCW

Chapter 4
Hypothesis Testing,
Power, and Control: A
Review of the Basics
From Question to Hypothesis



Finding the TRUTH starts with asking a question that
comes from
– Curiosity
– Necessity
– Past Research
As scientists we PREDICT the answer from
– Theory
– Past Research
– Common Sense
That prediction is EDUCATED not random
– An educated prediction is a HYPOTHESIS

To ANSWER the question we TEST the HYPOTHESIS
Three levels of hypotheses

Conceptual hypotheses
– State expected relationships among concepts.

Research hypotheses
– Concepts are operationalized so that they are
measurable.

Statistical hypotheses
– State the expected relationship between or
among summary values of populations, called
parameters.
 Null hypothesis (H0)
 Alternative hypothesis (H1)
Example
Question – What is the role of
neurotransmitters in memory?
 Conceptual – Increasing certain
neurotransmitter will increase memory
 Research – Smoking 1 crack rock before
testing will increase performance on a
standard test of memory compared to
placebo control
 Statistical – HO: MT = MC HA: MT >MC

Testing the null hypothesis

Null hypothesis
– The hypothesis being statistically tested when
you use inferential statistics.
– The researcher hopes to show that the null is
not likely to be true (i.e.. hopes to nullify it).

Alternative hypothesis
– The hypothesis the researcher postulated at
the outset of the study.
– If the researcher can show that the null is not
supported by the data, then he or she is able
to accept the alternative hypothesis.
Testing the null hypothesis

Steps in testing a research hypothesis:
1. State the null and the alternative.
2. Collect the data and conduct the appropriate
statistical analysis.
3. Reject the null and accept the alternative or
fail to reject the null.
4. State your inferential conclusion.
Statistical significance

Statistical difference
– The probability that the groups are the same
is very low.

Significance levels (α)
– Alpha (α) is the level of significance chosen by
the researcher to evaluate the null
hypothesis.
– 5% or 1%
Inferential Errors:
Type I and Type II

Type I Error
– Rejecting a true null.
– Probability is equal to alpha (α).

Type II Error
– Failing to reject a false null.
– Probability is beta (β).

Power – our ability to reject false nulls.
Inferential Errors:
Type I and Type II
True State of Affairs
Our decision
Null is true
Reject
the null
Null is false
Correct inference
Type I error (α)
(power)
Fail to
reject
Correct inference Type II error (β)
the null
Why Power is Important
A powerful test of the null is more likely to
lead us to reject false nulls than a less
powerful test.
 Powerful tests are more sensitive than less
powerful tests to differences between the
actual outcome (what you found) and the
expected outcome (null hypothesis).
 Power, or the probability of rejecting a
false null, is 1 – β.

Power and How to Increase it

How one measures variables
– Interval or ratio scales are better
 In testing the effects of alcohol intoxication on
aggression…
– Intoxication – BAC better than # of drinks
– Aggression – Level of shock (1-10) as opposed to shock
or no shock
Power and How to Increase it

Use more powerful statistical analyses

Parametric vs. Nonparametric

ANOVA vs. Chi-Square
Power and How to Increase it

Use designs that provide good control over
extraneous variables.

Remove unintended variation


Experimental vs. Correlational Designs
Laboratory vs. Field
Power and How to Increase it

Restrict your sample to a specific group of
individuals.

Use selection procedures to reduce nuisance variables
Power and How to Increase it

Increase your sample size  reduces error
variance
Power and How to Increase it

Maximize treatment manipulation


Precision
Separation
Effect size

Effect size – a measure of the strength of
the relationship between/among variables.
 Effect size helps us determine if
differences are not only statistically
significant, but also whether they are
important.
 Powerful tests should be considered to be
tests that detect large effects.
Effect size

Ways to calculate effect size:
– Cohen’s d – use with t-tests.
– Coefficient of determination (r2) – use with
correlations.
– eta-squared (η2) – use with ANOVAs.
– Cramer’s v – use with Chi-square analyses.
Power and the role of
replication in research

Power increases when we replicate
findings in a new study with different
participants in a different setting.
External and internal validity

External validity
– When the findings of a study can be
generalized to other populations and settings.

Internal validity
– Refers to the validity of the measures within
the study.
– The internal validity of an experiment is
directly related to the researcher’s control of
extraneous variables.
Confounding and
extraneous variables

Extraneous variable
– A variable that may affect the outcome of a study
but was not manipulated by the researcher.

Confounding variable
– A variable that is systematically related to the
independent and dependent variable.

Spurious effect
– An outcome that was influenced not by the
independent variable itself but rather by a
variable that was confounded with the
independent variable.
Confounding and
extraneous variables

Controlled variable
– A variable that the researcher takes into
account when designing the research study or
experiment.

Nuisance variables
– Variables that contribute variance to our
dependent measures and cloud the results.
Controlling extraneous variables

Elimination
– Get rid of the extraneous variables completely
(e.g.. by conducting research in a lab).

Constancy
– Keep the various parts of the experiment
constant (e.g.. instructions, measuring
instruments, questions).

Secondary variable as an IV
– Make variables other than the primary IV
secondary variables to study (e.g.. gender).
Controlling extraneous variables

Randomization: Random assignment of
participants to groups
– Randomly assigning participants to each of
the treatment conditions so that we can
assume the groups are initially equivalent.

Repeated measures
– Use the same participants in all conditions.

Statistical control
– Treat the extraneous variable as a covariate
and use statistical procedures to remove it
from the analysis.