Title: Quantitative Analysis of Survey Data and Other Assessments for Non-Experts
1Quantitative Analysis of Survey Data and Other
Assessments for Non-Experts
- How to do SoTL without a statistician on retainer
Gintaras Duda Creighton University
June 2, 2011
2My Background
- I am a theoretical particle physicist
- Came to SoTL (PER) as a junior faculty member
- New faculty workshop experience
- Huge roadblocks no experience with
- How educational research is conducted
- Quantitative or qualitative analysis
- Weak background in statistics
3Areas of SoTL interest
- Attitude of introductory physics students
- Particularly how it affects learning
- Online discussion behavior
- Realism in physics
- Problem-Based Learning in upper division courses
- Student note taking
- How students use the internet to learn physics
4Workshop Purpose
- As SoTL matures, publication requires more and
more rigorous measures and evidence - Sophisticated statistical tests
- Careful survey design and analysis
- Mixed method research
- Evidence, evidence, evidence!
- But, many of us are untrained in these things
5Who are you?
- Please share with the group
- Name, institution, and discipline
- Why you picked this workshop
- What you hope to gain
6Workshop Purpose continued
- Leave you with some simple tools to analyze
- Likert scale surveys
- Effects of instruction
- Survey reliability and validity
- No stats class or methods courses required
7Part I. What to do with Likert Scales
- Likert scale instruments seem ubiquitous in SoTL
work - Particularly useful in measuring students
attitudes, feelings, opinions, dispositions, etc. - Can use pre-post scheme to see changes and
growth/deterioration - Of interest in Jesuit Pedagogy (another workshop)
8Example from physics
- Attitudinal surveys
- Measure students changes in attitude towards
physics due to instruction - Instruments VASS, MPEX, C-LASS, Attitude II, and
others - These instruments all show a similar trend
- Students have more negative attitudes towards
physics after instruction
9Example Questions from Attitude II Instrument
- Physics is irrelevant to my life
- I can use physics in my everyday life
- I will did find it difficult to understand how
physics applies in the real-world - I see and understand physics in technology and
the world around me - 5 point Likert scale Strongly agree, weakly
agree, neutral, weakly disagree, strongly disagree
10One of my Likert Scale Instruments
11What do I do with Likert Scale Data?
- Two camps on analyzing Likert scale data
- Interval Approach
- Ordinal Approach
- Methods for data analysis differ between the two
methods
12Interval Data
- Basic philosophy differences between responses
are all equal - i.e. Difference to a student between strongly
disagree and weakly disagree is the same as the
difference between a neutral response and weakly
agree - Basic technique Sum the data and do some
statistics
13Ordinal Data
- Basic philosophy Differences between responses
are not equal - i.e. Students tend not to distinguish highly
between strongly and weakly statements - 3 pt Likert scale more appropriate?
- Basic technique Examine statistically the number
of students who agreed or disagreed
14Controversy over neutral response
- Good debate in the literature about the
neutral/neither agree nor disagree response - Some claim its crucial
- Some claim you should get rid of it
- Not going to discuss it here
15Analyzing Ordinal Data
- One method is to reduce the problem to a
binomial analysis - Lump all disagrees together, all agrees together,
and dont worry about neutral responses - Visual method Agree-disagree (Redish) plots
- Redish, J. Saul, and R. Steinberg, Student
expectations in introductory physics, Am. J.
Phys. 66, 212224 1998.
16Agree-Disagree Plots
- Introduced by Redish et al. in their MPEX paper -
called Redish plots
New Disagree Percentage
New Agree Percentage
Change from pre to post must be gt 2s to be
considered significant (at 5 probability level)
Standard Deviation
Redish, J. Saul, and R. Steinberg, Am. J. Phys.
66, 212224 1998.
17Example of an Agree-Disagree Plot
Duda, G., Garrett, K., Am. J. Phys. 76, 1054
(2008).
18Duda, G., Garrett, K., Am. J. Phys. 76, 1054
(2008).
19Analyzing Interval Data
- Basic idea here is assign a numerical value to
each response - Strong Disagree -2 (or 0)
- Weakly Disagree -1 (or 1)
- Neither Agree/Nor Disagree 0 (or 2)
- Weakly Agree 1 (or 3)
- Strong Agree 2 (or 4)
- Sum the responses then analyze using standard
statistical techniques
20Simple (student) t-test
- The t-test is a simple (but robust) statistical
test - Tests a hypothesis Is there a difference between
two sets of data? - Are differences statistically significant?
- 95 confidence level, i.e. only a 5 probability
the difference is due to statistical fluctuations
21Example The Gender Gap in Intro Physics
Is there a difference between male and female
students?
22Which image is random?
Sometimes our eyes can deceive us! And sometimes
we think things are true because wed like them
to be true
23The Gender Gap FMCE Gains
In the experimental group, there is no
statistically significant difference between the
two genders.
24Students t-test
- Assumptions
- Each data set follows a normal distribution
- Parameters
- One-tailed vs. two-tailed
- Types paired, two-sample equal variance, and a
two-sample unequal variance test - Can have different of data points if conducting
an unpaired test
25Demo
26Two Sample t-test
Here p lt 0.05, so the null hypothesis is
falsified statistical difference between Group
A and Group B
27Measuring Effects of Instruction
- Suppose you apply some educational innovation
- Control group and experimental group
- Or pre-test and post-test
- How do you know if its effective?
- Say you give some sort of standard assessment
- How big do the changes need to be to be
statistically significant?
28Method 1 Use a t-test
- You can always use a t-test
- Compare scores of control vs. experimental group
- or
- Compare pre vs. post tests
- More difficult due to other variables
29Method 2 Effect Size
- Effect Size (ES) is a method to quantify how
effective an educational intervention has been
relative to a control group - Extremely useful when there is no familiar scale
to judge outcomes
30A thought experiment
- Suppose we do a study to see if children learn
better in the morning or afternoon - Morning trial 15.2 average on assessment
- Afternoon trial 17.9 average on assessment
- Is this a big difference? It depends on overlap!
Robert Coe What is an Effect Size A guide for
users
31Two distributions
If the distributions of scores looked like this,
you would think the result is quite significant
Robert Coe What is an Effect Size A guide for
users
32Two distributions
But if the distributions of scores looked like
this you wouldnt be so impressed
Robert Coe What is an Effect Size A guide for
users
33Effect Size Continued
- The Effect Size
- Compares the difference between groups in light
of the variance of scores within each group - ES (mean of experimental group) (mean of
control group) - Standard Deviation
- Actually quite simple to calculate
- Robert Coe has great information online about ES
34How to Interpret Effect Size
Robert Coe What is an Effect Size A guide for
users
35How to Interpret Effect Size
Robert Coe What is an Effect Size A guide for
users
IQ differences between typical freshmen and
Ph.D.s corresponds to an effect size of 0.8
36Effect Size Example
Duda, G., Garrett, K., Am. J. Phys. 76, 1054
(2008).
37Making a better survey
- In my experience surveys and assessment
instruments are difficult to write - How do you know your instrument is
- Reliable
- Valid
- Are there alternatives to writing your own
instruments?
38Reliability Cronbach Alpha
- Cronbach Alpha measure of how closely items in a
group are related - Cronbach Alpha is often used for instruments
which are not marked right or wrong - Think Likert Scale
- Measures if students responses are the same for
similar types of questions
39How to Cronbach Alpha
- You could calculate it by hand
- or you buy SPSS and figure out how to use it
- or you could download an excel spreadsheet which
is programmed to do this http//www.gifted.uconn.
edu/siegle/research/Instrument Reliability and
validity/reliabilitycalculator2.xls
40Cronbach Alpha Values
- Typically a Cronbach Alpha (a) gt 0.8 is
considered good - At this level survey is reliable
- However, there are exceptions
- Different types of surveys/instruments may have
different natural levels of reliability - Experimental instruments may be still useful even
if a0.6
41Warning! Common Mistakes with Cronbach Alpha
- Paper Calculating, Interpreting, and Reporting
Cronbachs Alpha Reliability Coefficient for
Likert-Type Scales by Joseph A. Gliem and
Rosemary R. Gliem - Lesson
- Use Cronbach Alpha for Likert scale surveys
- Draw conclusions based on clusters of items
- Single item reliability is generally very low
42Instrument Validity
- Validity is never universal
- Valid for a certain population and for a
specific purpose - Three general categories of validity
- Content validity
- Predictive validity
- Concurrent validity
43Ideas for Establishing Validity
- Establish content or face validity
- Correlate with other independent measures such as
exam scores, course grades, other assessment
instruments - Predictive validity
- Longitudinal studies and student tracking are
needed here - Concurrent validity
- Compare with other assessment instruments or
calibrate with the proper groups
44Survey/Assessment Creation Tips
- Build in measures to show reliability
- e.g. multiple questions within a survey on the
same topic (both positive and negative) - Questions that establish that students are taking
the survey seriously - For content driven assessments, research student
difficulties - Beta-version open ended questions
- Correlations can help show validity
45An Example of evidence for Validity
Duda, G., Garrett, K., Am. J. Phys. 76, 1054
(2008).
46Buros Institute of Mental Measurement
- By providing professional assistance, expertise,
and information to users of commercially
published tests, the Institute promotes
meaningful and appropriate test selection,
utilization, and practice.
http//www.unl.edu/buros/bimm
47Conclusion
- Some simple statistical tests can provide
rigorous evidence of - Student learning
- Instructional effectiveness
- Improvements in attitude
- All of these methods are extremely effective when
coupled with qualitative methods - Stats involved can be done with little or no
training
48My SoTL advice
- Plan a throw-away semester in any SoTL study
- trial period to tinker with your study design
- Flexibility to alter your study design when you
find it doesnt work - Involving students in SoTL work can be very
effective - Try to publish in discipline specific journals
- When in doubt, ask your students!
49Good References
- Analysis of Likert Scales (and attitudinal data
in general) CLASS survey - http//www.colorado.edu/sei/class/
- Effect Size
- What is an Effect Size A guide for users by
Robert Coe (easily found by google) - Coe also has an excel spreadsheet online to
calculate effect size
50Good references
- Reliability and Validity
- http//www.gifted.uconn.edu/siegle/research/Instru
ment20Reliability20and20Validity/Reliability.ht
m - http//www.gifted.uconn.edu/siegle/research/Instru
ment20Reliability20and20Validity/validity.htm - T-test
- Step by step video on excel http//www.youtube.co
m/watch?vJlfLnx8sh-o
51Good References
- The FLAG Field-Tested Learning Assessment Guide
- www.flaguide.org
- Contains broadly applicable, self-contained
modula classroom assessment techniques (CATs) and
discipline-specific tools for STEM instructors
52Good References
John Creswells books (and courses) have been
highly recommended to me