session 12: egra data preparation and analysis...november 12, 2015 session 12: egra data preparation...
TRANSCRIPT
November 12, 2015
Session 12:
EGRA Data Preparation and Analysis
Outline of Presentation
This presentation will cover the following topics:
• EGRA Data Preparation
• EGRA Data Analysis
• Expectations for Reporting
3
Data Processing: [If done correctly]
• Technically, most difficult aspect of the study.
– Requires extensive understanding of hierarchical data structure
and complex sample statistics.
– Requires a lot of previous hands-on experience.
– Requires knowledge of how the specific sample was drawn and
how the data were collected.
– Requires only 1 person to do it. 1-2 people to check it.
– Takes a relatively short period of time to finish once data
collection is complete.
– But should not be rushed. [Should be checked over several
times by 1-2 people]
• Once finished and checked, the hard part is over.
– Several people can analyze the data and help write the report. 4
Data Processing: [If done incorrectly]
5
Activity #1
• What two components make up an EGRA data set?
• [Hint] Think Rows and Columns…
6
Data Composition: Sample + Questionnaires/Assessment
7
Questionnaires Variables Columns
Sam
ple
Ob
serv
ati
on
s
Ro
ws
school Id age female Item 1 … Score
school1 Student1 10 1 0 0 33.33
school1 Student2 9 0 1 1 25
school2 Student3 10 1 0 0 50
school2 Student4 12 0 0 1 75
EGRA Data Preparation – Observation [Row] Checklist
• Game of Clue: Who-when-where.
– Check all other assessments.
• Who: Conducted by a real assessor.
• When: Not on a weekend or holiday.
• Where: A school that was randomly sampled.
• How Many: Actual counts in the data match with the expected
number of students
• Review incomplete assessments
9
Day # date
Day of week School code School name
CompletedEGRA
counts
Enumerator
counts enumerators1 10/21/13 Mon PS0102-088 OLASITI PR. SCHOOL 11 4 AROLE(2)//JMTOW(3)//KMPAN(3)//SELIA(3)1 10/21/13 Mon PS0105-018 KIMOSONU PR. SCHOOL 10 4 CMATO(3)//FHENJ(3)//MASU(1)//RUSAM(3)1 10/21/13 Mon PS0205-038 WISDOM PR. SCHOOL 12 4 FNYAK(3)//HKILO(3)//HMDUN(3)//TKIRI(3)1 10/21/13 Mon PS0405-047 KASANGA PR. SCHOOL 12 4 IKIWA(3)//MMLAA(3)//MNOL(3)//MPAZI(3)1 10/21/13 Mon PS0406-105 MATIGANJOLA PR. SCHOOL 12 4 JURAS(3)//LSIMBE(3)//MKYEJ(3)//NKIBO(3)1 10/21/13 Mon PS0503-030 HAPPY PR. SCHOOL 11 4 GHAMI(2)//GJOSE(3)//JMANY(3)//SYMSE(3)
1 10/21/13 Mon PS0505-187JIPE MOYO ENGLISH
MEDIUM 12 4 HSELE(3)//LBIJO(3)//PANDE(3)//RBUSY(3)
Activity #2:
• Circle 4 things in the table below that should be
checked for data quality. What could have caused
these inconsistencies?
10
Day
# date
Day of
week School code School name
Completed
EGRA
counts
Assessor
counts Assessor User Names
1 10/19/13 Sat PS0102-088
OLASITI PR.
SCHOOL 12 4 AROLE(3)//JMTOW(3)//KMPAN(3)//SELIA(3)
2
10/21/13 Mon PS0105-018
KIMOSONU
PR. SCHOOL 10 4 CMATO(3)//FHENJ(3)//MASU(1)//RUSAM(3)
2
10/21/13 Mon PS0105-031
MUGABE PR.
SCHOOL 2 1 MASU(2)
2
10/21/13 Mon PS0205-038
WISDOM PR.
SCHOOL 9 3 FNYAK(3)//HKILO(3)//TKIRI(3))
3
10/22/13 Tue PS0205-038
WISDOM PR.
SCHOOL 3 1 HMDUN(3)
3 10/22/13 Tue Practice Practice 12 4 JURAS(3)//LSIMBE(3)//MKYEJ(3)//NKIBO(3)
3
10/22/13 Tue PS0505-187
JIPE MOYO
ENGLISH
MEDIUM 12 4 HSELE(3)//LBIJO(3)//PANDE(3)//RBUSY(3)
EGRA Data Preparation
12
Name of
subtask
variable
Label for subtask
variable
Name of
subtask
timed
variable Label for subtask timed variable
letterLetter Idenfication (Names) clpm Correct Letter Names per Minute
letter_soundLetter Identification (Sounds) clspm Correct Letter Sounds per Minute
fam_word Familiar Word Reading cwpm Correct Words per Minute
invent_word Nonword Reading cnonwpm Correct Nonwords per Minute
oral_read Oral Reading Fluency orf Oral Reading Fluency
read_comp Reading Comprehension
list_compListening Comprehension
syll_soundSyllable Identification (Sounds) csspm Correct Syllable Sounds per Minute
oral_vocab Oral Vocabulary
vocab Vocabulary
maze Maze
dict Dictation
EGRA subtask variable nomenclature andnames of the timed score variables
EGRA Data Preparation
13
Item and score variables
Suffix Variable suffix label Possible values
1-# Item #
0 "Incorrect" 1 "Correct" . <missing> "Not asked/didn't attempt"
_score Raw Score 0 - # Items in Subtask
_attempted Total items attempted 0 - # Items in Subtask
_score_pcnt Percent Correct 0-100
_score_zero Zero score indicator0 "Score>0" 1 "Score=0"
_attempted_pcnt Percent Correct of Attempted 0-100
Activity #3: Calculate Oral Reading Fluency
14
Reading Passage
[Oral Reading Fluency]
Reading
Comprehension
One day, Juma lost his hat.6
What did Juma lose?
[his hat]
He was not happy. It was very cold. He looked into
his desk and on his chair. 23
Where did Juma look for his hat?
[his desk, chair/seat, classroom, under the big
tree, playground]
The hat was not there. He ran to the playground.33
Where did Juma run?
[the playground]
He looked under the big tree. It was not there. He
told the teacher he had lost his hat. The teacher
pointed to Juma’s head.
58
Where was Juma’s hat?
[on Juma’s head]
Juma laughed.
60
Why did Juma laugh?
[one or more of because the hat was on Juma’s
head / he felt silly / embarrassed ]
___ Number Incorrect ____ Number Correct
___ Time Taken
*Grade2 Tanzania National Cross-Section Study
English EGRA Reading Passage
EGRA Data Preparation – Timed Subtasks
15
The intention of a timed subtask in an EGRA or EGMA
instrument is to calculate the items per minute rate.
Subtask_per_minute=𝑠𝑢𝑏𝑡𝑎𝑠𝑘_𝑠𝑐𝑜𝑟𝑒
𝑠𝑒𝑐𝑜𝑛𝑑𝑠 𝑡𝑎𝑘𝑒𝑛 𝑡𝑜 𝑐𝑜𝑚𝑝𝑙𝑒𝑡𝑒 𝑠𝑢𝑏𝑡𝑎𝑠𝑘× 60
𝑠𝑒𝑐𝑜𝑛𝑑𝑠
𝑚𝑖𝑛𝑢𝑡𝑒
EGRA Data Preparation – Timed Subtasks Score Checks
• Check and Edit Potential Extreme Values
16
0
100
200
300
400
500
600
700
800
0 5 10 15 20 25 30 35 40 45 50 55 60
Timed Oral Reading Fluency Score
Tim
edO
ral R
ead
ing
Sco
re [
wp
m]
Time Taken (seconds)
Outline of Presentation
This presentation will cover the following topics:
• EGRA Data Preparation
• EGRA Data Analysis
• Expectations for Reporting
18
EGRA Data Analysis - Definitions
Descriptive Analysis (non-inferential statistics)
Simply describes specifics of the sample
Inferential Analysis
Incorporates sample weights and the cluster effect to project the
sample estimates to the population from which the sample was
drawn.
19
Sample
Sample
Clu
ster
Eff
ect
Descriptive Analysis - Recommendation
• Describe the sample according to the sub-group level to
be reported in the main results section.
• Commonly report the basic demographic information.
– Gender, grade, age, region ect.
20
Tanzania 2013 National G2 EGRA Sample Counts by School’s Performance Band*
PerformanceBand
CountsStudents % of Sample
High 465 20.5
Medium 840 37.1
Low 961 42.4Total 2266 100.0
*Performance band based on Standard 7 PSLE for 2012
Inferential Statistics - Recommendation
• Describe the Population according to the sub-group
level to be reported in the main results section.
– If possible include the estimated population counts for the
subgroups.
21
Tanzania 2013 National G2 EGRA Sample Counts by School’s Performance Band*
PerformanceBand
CountsStudents
% of Sample
Estimated Population Count
% of Population
High 465 20.5 14,762 1.4Medium 840 37.1 121,666 11.4
Low 961 42.4 931,773 87.2Total 2266 100.0 1,068,201 100.0
*Performance band based on Standard 7 PSLE for 2012
Inferential Statistics - Goal
Samples will always have some uncertainty in the estimates
when projecting them to the population.
• Goal:
– Obtain unbiased means/percentages estimates that truly reflect
(representative of) the population. Sample Weights
– Obtain an appropriate level of precision for these estimates
based on the subgroup level you would like to report Sample
Size, Cluster Effect, Sample Methodology
• An appropriate 95%Confidence band width
• Be able to detect statistical significant (and appropriate) difference
between sub-groups.
22
Activity #4 True or False
Samples are drawn to answer the main research questions.
Samples are typically proportionally representative of the population.
Sample weights make the sample representative of the population.
23
True
False
True
Sample Weights
• Samples are drawn the answer the main research questions
• Typically NOT proportionally representative of the
population.
• This causes: Over/Under sampling of sub-groups makes the
sample NOT representative of the population.
• Sample weights make the sample representative of the
population
• Unweighted analysis produces biased estimates that will not
be representative of the population. 24
Sample Weights Representative
• Example: Tanzania Grade 3 2013
• Level of Report: School Performance Band
25
Tanzania 2013 National G2 EGRA Sample Counts by School’s Performance Band*
PerformanceBand
CountsStudents
% of Sample
% of Population
High 465 20.5 1.4
Medium 840 37.1 11.4
Low 961 42.4 87.2Total 2266 100.0 100.0
*Performance band based on Standard 7 PSLE for 2012
Notice any over sampling?
Any under sampling?
How could this affect the
National estimates?
National Estimates. Kiswahili English
Sample [Not weighted]
Population [Weighted]
25.7
17.8
23.19.4
Estimates should be reflective of the population
Outline of Presentation
This presentation will cover the following topics:
• EGRA Data Preparation
• EGRA Data Analysis
• Expectations for Reporting
26
SampleC
lust
er
Eff
ect
Reporting
Before writing the EGRA research report, consideration needs to be given to
who the audience is.
Primary Audience
• USAID
• Country Ministry of Education personnel
Secondary Audience
• Other stakeholders, e.g. District Education Officers
• Education researchers
Understanding the audience and their needs enables the writer to structure
and write the report appropriately.
27
Reporting
• The report must explicitly define the population of interest in the
study and clearly explain the sample methodology.
• The report must clearly state the objectives of the study and its
limitations.
• The main findings should be presented in clear, concise, and
non-technical language.
• The main report should present summary findings of inferential
data analysis including:
– Means and percentages
– Standard errors and/or 95% Confidence Intervals
– Distributions
– Formal statically tests when appropriate (i.e. comparing difference in
means between subgroups)
28
Tanzania 2013 “Snap Shot” Survey
• Population
– All P2 pupils attending public schools on mainland Tanzania
(estimated ~1.1 million pupils) during the 2012-2013 school
year.
• Objective
– Obtain a national estimate for the P2 pupil reading ability and to
see if P2 pupils attending high performing schools read
significantly better than those attending medium, and low
performing schools.
Notes:
1. National estimate did not include Zanzibar and Pemba.
2. Performance Band: High, Medium, low are based on the
Standard 7 Primary School Leaving Exam (PSLE) in 2012 29
Example Of Report Table
– Means and percentages
– Standard errors or 95% Confidence Intervals
– Distributions
– Formal statically tests when appropriate (i.e. comparing
difference in means between subgroups)
30
Tanzania 2013 National G2 Kiswahili EGRA Mean Kiswahili Results by School Performance Band
PerformanceBand Mean
95% Confidence Interval
Difference in Mean
High* 40.8 (36.6, 44.9) 12.2
Medium^ 28.6 (24.6, 32.7) -Low* 16.1 (12.8, 19.4) 12.5
Overall 17.0 (14.5, 21.3) n/a`Performance band based on Standard 7 PSLE for 2012
^Reference subgroup from which other groups were compared*P-value < 0.05 in t-test difference in means
Should these estimates
be reflective of the sample
or the population?
Examples of Graph Distributions:
Probability Density & Cumulative
31
0
10
20
30
40
50
60
70
80
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Tanzania 2013 Grade 2 National Kiswahili Oral Reading Fluency
Ora
l Rea
din
g Fl
uen
cy
[Wo
rds/
min
]
Percentile of Grade 2 Students
27.7
13.0
19.4
15.714.3
5.72.9
0.9 0.60
5
10
15
20
25
30
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
Tanzania 2013 Grade 2 National Kiswahili Oral Reading Fluency
Oral Reading Fluency Score [Words/min]
% o
fG
rade
2 S
tude
nts
Probability Density by Performance Band
32
05
101520253035
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
High Band
% o
f G
rade
2 S
tude
nts
Oral Reading Fluency Score [words/min]
-5
5
15
25
35
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
Medium Band
% o
f G
rad
e2
Stu
de
nts
Oral Reading Fluency Score [words/min]
0
5
10
15
20
25
30
35
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
Low Band%
of G
rade
2 S
tude
nts
Oral Reading Fluency Score [words/min]
Cumulative Distribution by Performance Band
33
0
10
20
30
40
50
60
70
80
90
0 10 20 30 40 50 60 70 80 90 100
Tanzania 2013 Grade 2 National Kiswahili Oral Reading Fluency by Performance band
High* Medium^ Low*
Ora
l R
eadin
g F
luency [
Word
s/m
in]
Percentile of Students
Reporting – Data Visualization
Data visualization should be used to facilitate understanding of the findings by
general audiences. Visualizations should be ‘standalone’, such that the visual
is interpretable without the audience needing to read extra text.
34
0
10
20
30
40
50
60
70
80
90
0 10 20 30 40 50 60 70 80 90 100
Tanzania 2013 Grade 2 National Kiswahili Oral Reading Fluency by Performance band
High* Medium^ Low*
Ora
l R
eadin
g F
luency [
Word
s/m
in]
Percentile of Students
Performance band based on 2012 Standard 7 PSLE examination
^Reference subgroup from which other groups were compared.
* P-values <0.05 t-test difference in means
PerformanceBand n
Wt%
MeanORF 95%CI
High* 465 1.4 40.8 (36.6, 44.9)
Medium^ 840 11.4 28.6 (24.6, 32.7)
Low* 961 87.2 16.1 (12.8, 19.4)
Over all 2266 100 17.9 (14.5, 21.3)
Reporting – Regression
• Linear Regression: – Continuous variable such as the Oral Reading Fluency score.
– Allows reports to conclude:
35
Tanzania 2013 National G2 Kiswahili EGRAMean Kiswahili Oral Reading Fluency Results by School Performance Band
PerformanceBand Mean
95% Confidence Interval
Difference in Mean
High* 40.8 (36.6, 44.9) 24.7
Medium* 28.6 (24.6, 32.7) 12.5
Low^ 16.1 (12.8, 19.4) -
`Performance band based on Standard 7 PSLE for 2012^Reference subgroup from which other groups were compared
*P-value < 0.05 in T-test difference in means
“Students attending medium performing schools read on average 12.5 words/min faster
than students attending low performing schools. Students attending high performing
schools read on average 24.7 words/min faster than students attending low performing
schools”.
Reporting
• The results should be reported by common
demographic variables and other variables of interest
as appropriate to the research questions.
• Results should be reported even if they are not
statistically significant.
• Difference between Actual Significant and Statistical
Significant
36
Activity #6: Circle the “Difference in Mean” that show actual/real
significant difference.
Tanzania 2013 National G2 Kiswahili EGRA
Mean Kiswahili Oral Reading Fluency Results by various demographics
Characteristic Sub-Group Mean
Difference in
Mean
95% Confidence
Interval
Performance
Band
High* 40.8 24.7 (36.6, 44.9)
Medium* 28.6 12.5 (24.6, 32.7)
Low^ 16.1 - (12.8, 19.4)
School Type’’Public^ 17.8 - (16.8, 18.5)
Private* 18.5 1.2 (17.8, 20,2)
GenderMale^ 16.2 - (12.8,19.6)
Female* 19.6 3.4 (15.8, 23,2)
PreschoolAttended 19.4 - (16.0, 22.6)
Did not
attend* 13.9 5.5 (10.6, 17.2)`Performance band based on Standard 7 PSLE for 2012
^Reference subgroup from which other groups were compared
*P-value < 0.05 in T-test difference in means
`` Estimates were fabricated
37
Activity #6: Circle the “Difference in Mean” that show actual/real
significant difference.
National G2 Kiswahili EGRA
Mean Kiswahili Oral Reading Fluency Results
by years
Year`` Mean
Difference
in Mean
95% Confidence
Interval
2013^ 17.0 - (14.5, 21.3)
2015 15.7 -1.3 (13.6, 18.2)
2017* 23.5 6.5 (18.3, 25.8^Reference subgroup from which other groups were
compared
*P-value < 0.05 in T-test difference in means
`` Estimates were fabricated
38
Examples of Non-Normal Distributions: Ghana 2015
40
0
10
20
30
40
50
60
70
80
90
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
Ghana 2013 Grade2 Mother-TongueNzema Oral Reading Fluency
% o
fG
rade
2 S
tude
nts
Oral Reading Fluency Score [Words/min]
Mean 3.0
Median 0
Mode 0
Logistic Regression
43
27.7
13.0
19.4
15.714.3
5.7
2.90.9 0.6
0
5
10
15
20
25
30
0 1-10 11-20 21-30 31-40 41-50 51-60 61-70 71+
Tanzania 2013 Grade 2 National Kiswahili Oral Reading Fluency
Oral Reading Fluency Score [Words/min]
% o
fG
rade
2 S
tude
nts
Mean 17.9
Median 16
Mode 0
• Use the data, find an appropriate Cut-Point
– Dichotomous “No/Yes”, (0,1) variables
– Report percentage by subgroup
Top ORF Reader
[Cut at 40 words/min]
0 1
Reporting – Logistic Regression
44
• Logistic Regression: – Report the percentages [rather than means]
– Report Odds-Ratios [rather than difference in means]
Tanzania 2013 National G2 Kiswahili EGRA Compare Top Kiswahili Readers' with Non-Top Readers by Performance Band’’
Performance Band
Not Top Orf Sampled
Top Orf Sampled
% TopReaders
Odds Ratio
High* 239 226 49.7 11.4Medium^ 670 170 21.1 3.1
Low* 866 95 8.0 -Overall 1775 491 10.0
`Top Kiswahili Readers defined as grade 2 students reading 40+ words/min‘’Performance band based on Standard 7 PSLE for 2012^Reference Subgroup from which other groups were compared*P-value < 0.05
“Students attending high performing schools are 11.4 times more likely to be top readers than students
attending middle performing schools. And students attending middle performance schools are 3.1 times
more likely to be top readers than students attending low performing schools”.
Reporting
• The following must accompany all reported inferential
estimates (including but not limited to means, median,
mode and proportions):
– Precision – either as 95% confidence interval for estimates, or
a t-score and p-value for comparisons in addition to standard
errors.
– Sample size
45
Note about Baseline Scores [Impact Evaluation Study]
46
Baseline
Mean T-Test P-Value
Effect
Size
Control 30.91.1 0.313 0.10
Treatment 32.3
Check for Control/Treatment Balance.
The difference is measured in standard deviations (called the ‘effect size’)
and should ideally be less than 0.25.
Baseline
Mean T-Test P-Value
Effect
Size
Control 30.95.3 <0.001 0.42
Treatment 22.3
Reporting: Not on Impact Evaluation Study
• Whenever results of comparisons of data across groups
at different times (baseline and end-line), effect size of
the difference needs to be reported.
48
Table 1: Example of Difference-in-Difference AnalysisBaseline Endline
Treatment
Mean
fluency
(wpm)
Standard
error
Number
of
sampled
students
t-statp-
value
Mean
fluency
(wpm)
Standard
error
Number
of
sampled
students
t-statp-
value
Differ-
ence–in–
difference
p-
value
(DiD)
Effect
size
Control 4.5 0.6 656 – – 9.5 1.6 475 – – – – –
Intervention 5.2 1.2 349 0.510 0.611 11.7 1.1 4801.18
90.236 1.5 0.490 0.12
Difference-in-Difference: (Mean endline treatment – mean baseline treatment) – (mean endline control – mean baseline control)
Effect Size (Cohen’s d): Difference-in-difference / pooled standard deviation
Reporting - Annexes
Researchers should also include details of the
methodology and results of the analysis is the
annexes, which can be quite lengthy when written
in a technical language. The following should be
included in the annexes:
49
Reporting - Annexes
50
1. Details of the methodology, methods and data
collection:
a. Study objectives
b. Design
c. Data collection methods and process
d. Data collection instruments
e. Method and results of equating if different tools were used
at different study points
f. Sampling parameters and attrition (for longitudinal studies)
g. Details on weighting
h. Limitations
i. Results of test reliability analysis (Cronbach’s alpha; item-
total correlations)
j. Intra-class correlation coefficient (ICC)
Reporting - Annexes
51
2. Details of analyses that were not included in the main
report:
a. Sample description
b. Details of descriptive analyses
c. Details of bivariate/multivariate analyses
Take-Home Suggestions
• EGRA Data Preparation– Have experienced statisticians do it right so that the data can be
analyze to infer about the population
• EGRA Data Analysis– Inferential analyze to project the estimates to the population.
• Expectations for Reporting– Report findings that answer the research questions, not just
statistically significant findings.
52
Sample
Clu
ster
Eff
ect
More Information
Chris Cummiskey
RTI International
Sample
Clu
ster
Eff
ect