National Centre for Research Methods: NCRM EPrints Repository

Hannemann, Tina

Shlomo, Natalie

Compensating for Non-response
in Biosocial Research: 
Simulation Study from a Cross-sectional Analysis
Dr Tina Hannemann 
Biomarkers in Longitudinal Studies in the UK
• Understanding Society
• Millennium Cohort Study
• British Cohort Study
• Child Development Study
• English Longitudinal Study of Ageing
English Longitudinal Study of Ageing (ELSA)
• Data from 11,391 core sample members of men and 
women over 50 years living in private households  
• Multistage stratified probability sample from the Health 
Survey of England (HSE) with refreshment samples 
• Further health examinations including blood samples 
(W2, W4 & W6) and hair samples (W6)
Missing Information in ELSA
• Potential non-response at three different stages: 
• Main interview (CAPI)
• Health examination (Nurse visit)
• Blood/Hair sample collection 
Patterns of Missing information
• Missing completely at random (MCAR)
• Missing at random (MAR)
• Missing not at random (MNAR) 
Patterns of Missing information
• Missing completely at random
X
Y M
X  explanatory variable
Y  depend variable
M  missingness of Y
Patterns of Missing information
• Missing at random
X
Y M
X  explanatory variable
Y  depend variable
M  missingness of Y
Patterns of Missing information
• Missing not at random
X
Y M
X  explanatory variable
Y  depend variable
M  missingness of Y
Accounting for missing data bias
• Several methods for missing data compensation
• Confusion over effectiveness of methods
• Problem of missing data is ignored 
(assumption of MAR)
Objective
• Test and evaluate several analytical approaches
• Simulation study based on data from ELSA (wave 2)
• Comparison of five different approaches
• Evaluation over all three missingness patterns 
Strategy
Step 1
Choose 
substantive 
model
Step 2
Select data in 
ELSA, expand 
sample size
Step 3
Run baseline 
model on “true” 
values
Step 4
Create missing 
data for MCAR, 
MAR and 
MNAR
Step 5
Test five 
different 
compensation 
approaches
Step 6
Compare with 
“true” model 
and evaluate 
effectiveness
Step 1 – Substantive model
Impact of socio-economic status on biomarker level
• SES 
• Education, occupational class and wealth quintile
• Biomarker
• C-reactive protein (CRP)
• Confounding variables
• Age, sex, physical activity, BMI, Longstanding 
illness, pain, general health, BP, in paid work
Step 2 – ELSA missing information
8.780 
Wave 2 core member
Step 2 – ELSA missing information
2.881
Non-response and missing
5.899   Eligible blood samples
32.8 %
67.2%
Step 2 – ELSA missing information
2881
Non-response and missing
5.899   Eligible blood samples
332
839
596
1.114
Unusable blood sample
No blood sample
Refuse blood sample
Refuse of nurse visit
3.8 %
9.6 %
6.8 %
12.7 %
67.2%
Step 3 – Model on expanded full dataset
• Logistic regression on logged CRP values 
(10.000 sample)
• Coefficients for all three SES variables as baseline for 
simulation study
Baseline Model
Education
(ref. high)
med. low
Occupation
(ref. professional)
Wealth quintiles
(ref. lowest)
In
te
rm
e
d
ia
te
L
o
w
e
r 
s
u
p
. 
a
n
d
 t
e
c
h
S
e
m
i-
ro
u
ti
n
e
 a
n
d
 o
th
e
r
N
o
 i
n
fo
2nd 5th3rd 4th
Step 4 – Creating missing data patterns
• MCAR: Missingness is completely random
• MAR: Missingness dependent on age, sex, wealth 
and health
• MNAR: Missingness dependent also on level of CRP
• 100 simulation datasets for each missing patterns
Step 5 – Five compensation methods
• Complete Case (CC)
• Inverse Propensity Weighting (IPW)
• Selection model using Mills Ratio (MR)
• Multiple Imputation (MI)
• Multiple Imputation with Mills Ratio (MI+MR)
Step 6 – Run and evaluate simulations
CC IPW MR MI MI+MR
MCAR
MAR
MNAR
Expectation
CC IPW MR MI MI+MR
MCAR
MAR
MNAR
Complete Case –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
IPW –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
MR –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
MI –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
MI+MR –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
MI+MR –Wealth Quintiles
MCAR MAR MNAR
2nd quintile (-0.09)
3rd quintile (-0.07)
4th quintile (-0.23)
highest quintile (-0.31)
Expectation
CC IPW MR MI MI+MR
MCAR
MAR
MNAR
Observation
CC IPW MR MI MI+MR
MCAR
MAR
MNAR
Conclusions
• No clear patterns
• Multiple Imputation and Mills Ratio method has no 
advantage
• MCAR and MAR rather robust
• MNAR bias not compensated with these methods
Future Research
• Expand the simulation study to larger iteration number
• Simplify substantial model
• Test and evaluate effectiveness of methods depending 
on the proportion of missingness in dataset
• Longitudinal analysis including wave 4 and wave 6 using 
growth curve modelling
32
Thank you for your attention
Tina.Hannemann@manchester.ac.uk


Compensating for Non-response in Biosocial Research: Simulation Study from a Cross-sectional Analysis

Olsen, Wendy

Using Mixed-Methods 
Evaluation Methods Taking 
Into Account Gender/Class 
Realities:  Using QCA and 
NVIVO 
 
Wendy Olsen 
Funded by British Academy: 
Innovation in Global Labour 
Research Using Deep Linkage and 
Mixed Methods  
 
Applications to: 
□ Theorising Bangladesh Indebtedness  
□ Mediated by involvement in an NGO  
□ All NGOs are different; and 
□ Indian Women’s Work 
□ Mediated by involvement in  
 either the self-help groups, an MFI, 
   Or an NGO, or the Employment 
Guarantee Scheme 
Steps for a Mixed-Methods 
Evaluation Approach 
□ Step 1:  a complex theory of the ontic 
realities, ie the types of things  
□ Step 2:  fieldwork 
□ Step 3:  analysing early, & linking results 
□ Step 4:  keyness, discourses  
□ Step 5:  perhaps QCA analysis 
□ Step 6:  transparency:  database 
□ Step 7:  draw conclusions 
Key Sampling Themes 
□ Representativeness at some level 
 
□ Idea of the replication of entities 
across a geographic space 
 
□ Generalisation to known sub-
populations and concrete spaces 
Step 1:  a complex theory of the ontic 
realities, ie the types of things 
□ The ontic reality is treated by 
statisticians as Structured 
□ Outcome = result of structures, events. 
Logic is 
□ Y = results arise from S, I, E, C, random error 
I = institutions, local entities 
C = context 
 A non-statistical approach. 
 
Discussion of Key Sampling 
Themes 
  
ADVICE 
□ You may triangulate a national dataset onto your local data 
□ Match questions on demographics, take a random sample not 
non-random! 
□ Randomness at some, not all levels is, overall, non-random 
□ But generalisation can be made at the level-to-which 
randomness was applied, e.g. by geographic transect walks. 
□ E.g. a village.  Or all the Slums of Dhaka if the Slums were stage 1 and 
□ The choice of households was stage 2 
□ And the choice of individuals (KISH) was stage 3 
□ So be very professional about selection of cases. 
□ Multi-stage quota sampling vs. Multi-stage RANDOM sampling: The 
difference is in the degree of REPLACEMENT of non-response cases. 
 
ADVICE 
□ You may triangulate a national dataset onto your local data 
□ Match questions on demographics, take a random sample not 
non-random! 
□ Randomness at some, not all levels is, overall, non-random 
□ But generalisation can be made at the level-to-which 
randomness was applied, e.g. by geographic transect walks. 
□ E.g. a village.  Or all the Slums of Dhaka if the Slums were stage 1 and 
□ The choice of households was stage 2 
□ And the choice of individuals (KISH) was stage 3 
□ So be very professional about selection of cases. 
□ Multi-stage quota sampling vs. Multi-stage RANDOM sampling: The 
difference is in the degree of REPLACEMENT of non-response cases. 
 
Never sample 
on the outcome 
variabl . 
 For xample on income 
levels, if you want to 
explain the change in 
income over time.   
NVIVO Keyness Analysis of 
Discourses in Large Dataset ( With 
Example of Matrix Results from 
South India) 
Wendy Olsen 2016 
Gender Norms Project 
Acknowledging funding of ESRC DFID 
Pov. Allev. Fund and British Academy 
9 
Next Steps:  
□ Step 2:  fieldwork 
□ Step 3:  analysing early, & linking results 
□ Step 4:  keyness 
How to conduct a Keyness Analysis for a 
Social Science Research Project.   
1. Pool all the transcripts 
 
2. Find out the keyness 
of words 
 
3. Code up the 
concordances 
 
4. Group the words into 
discourses 
 
5. Interpret selected 
discourses only 
 
6. Treat each one of 
those very carefully:  
the dominant 
discourse must be 
discerned, then the 
marginalised, deviant 
and innovative 
(intertextual) ones. 
7. Trace key arguments 
through these. (Mixed 
Methods) 
11 
How the Keyness Analysis is Done 
1. Keyness of words 
2. Discourses too 
3. Interpretations: 
dominant discourse; 
Marginalised & 
intertextual ones. 
4. Trace key arguments 
through these. (Mixed 
Methods) 
 
 
 
Key References: 
□ Touri, M., and N. Koteyko 
(2014) “Using Corpus 
Linguistic Software”, 
International Journal of 
Social Research 
Methodology 
□ Fairclough, Norman various, 
books on Discourse and 
Power. 
12 
Part One: The Keyness of 
Words 
(Touri and Koteyko 2014) 
□ Keyness is the relative prevalence of 
words in one corpus of material over 
another. 
□ Specifically, count S words in corpus, 
vs. N words in the British National 
Corpus of English Language. 
□ Use the formula provided here. 
13 
Formula for Keyness 
□ 𝐾𝑒𝑦𝑛𝑒𝑠𝑠 = 𝑜𝑑𝑑𝑠 𝑟𝑎𝑡𝑖𝑜  
□ The odds of a word appearing in the fieldwork based corpus 
vs. the odds of it appearing in the national corpus 
□ K = 
𝑠𝑖
S−𝑆𝑖
𝑛𝑖
N−𝑛𝑖
  For each word I 
Counting words using NVIVO then 
Matching words using STATA or SPSS 
Report output as a word list, RANKED. 
Word Count Query in NVIVO 
15 
Excel Spreadsheet – Highest 
Keyness 
Delemmatised   (mention) of mentions     Ratio 
Word Length Count Percent 
BNC 
Prevalence 
BNC % of BNC 
Odds 
brickfields 11 2 0% 2 0% 4738.06 
laws’ 5 2 0% 3 0% 3158.71 
purdah’ 7 1 0% 2 0% 2369.03 
coops 5 2 0% 11 0% 861.47 
passbook 8 1 0% 6 0% 789.68 
betel 5 3 0% 23 0% 618.01 
mindset 7 1 0% 11 0% 430.73 
parishad 8 2 0% 25 0% 379.04 
stipends 8 2 0% 38 0% 249.37 
negatively 10 2 0% 52 0% 182.23 
sons’ 5 1 0% 28 0% 169.22 
educate 7 12 1% 365 0% 155.77 
workloads 9 1 0% 43 0% 110.19 
rears 5 1 0% 45 0% 105.29 
chores 6 6 1% 275 0% 103.38 
robbers 7 5 0% 244 0% 97.09 
tailoring 9 3 0% 147 0% 96.70 
dhaka 5 1 0% 55 0% 86.15 
16 
Illustration of Keyness by Odds 
Ratio. 
We created a spreadsheet to 
rank words. 
17 
In this example from South 
India, 39 interviews with 
couples. 
□ 39 Interviews 
□ 47,000 Words 
□ We reduced these to 233 key words.  
Extremely concise summary. 
□ Then as an expert I examined these to 
group them into discourse topics. 
□ Next I study these discourse topics to 
identify discursive patterns. 
18 
Interim Product Conforms to Miles 
& Huberman’s Advised “one-
page summary” 
19 
20 
Here’s an example  
(a small South Indian project) 
□ Annotate and summarise the Key Terms. 
□ Group them into dominant discourses. 
□ This is also like thematic analysis, initially. 
□ Discourses are sets of rules which are 
coherent but which are held to only via 
normed practices, and which can be 
broken, at a certain price. 
□ Example of patriarchal talk about 
marriage as an exchange of assets. 
□ Next:  Locate the marginalised discourses 
21 
Steps for a Mixed-Methods 
Evaluation Approach 
□ Step 1:  a complex theory of the ontic realities, ie 
the types of things  
□ Step 2:  fieldwork 
□ Step 3:  analysing early, & linking results 
□ Step 4:  keyness, discourses  
□ Step 5:  perhaps QCA analysis 
□ Step 6:  transparency:  database 
□ Step 7:  draw conclusions 
Discourses we found (South 
India; North India) 
□ Dominant ones: 
 
□ Agriculture as 
production 
□ Family as duty, 
obligations ( 
disciplining) 
□ Moneylending as 
a solution 
□ Marginal ones: 
 
□ Agriculture as a 
burden the older 
generation carry, 
disliked 
□ Family as conflict 
□ Moneylending and 
debt as a 
problem 
23 
SCALE of the DATABASE:  A Small 
Research Project in Bangladesh  
1 interview 673 raw 
words of 5+ 
letters 
396 “words” 
i.e. word-
roots, in one 
interview, if 
you stem the 
words 
By hand 
11 interviews 1666 words 1249 after 
stemming 
By NVIVO 
32 interviews 2798 words 2066 word-
roots, after  
stemming 
By NVIVO 
24 
Using the Words with Highest 
Keyness  
□ We set a cutoff level for keyness 
(the odds ratio) e.g. 4, or 9. 
□ Collect the concordances using 
NVIVO 
□ You now have extensive 
quotations to compare and 
contrast. 
□ Link the survey data to this database. 
25 
REMINDER:  My Keyness Method 
1. Pool all the transcripts 
 using NVIVO. 
1. Find out the keyness of 
words 
2. Code up the 
concordances 
3. Group the words into 
discourses 
4. Interpret selected 
discourses only 
5. Treat each one of those 
very carefully 
6. Trace key arguments 
through them. 
 
26 
COMPARATIVE NVIVO 
Results for two discourses (family talk and money talk) 
[india 1 and bangla 1 combined]  Mentioned within 
30 words of each other, in combination. 
A : 
Tightness 
node 
B : money 
C : 
problems 
D : spend E : works 
1 : 
Family 
20 18 11 9 22 
2 : 
children 
12 10 6 5 11 
3 : 
daughter 
15 6 3 0 13 
4 : 
husband 
10 6 8 1 11 
5 : 
mother 
11 6 1 4 11 
27 
Steps for a Mixed-Methods 
Evaluation Approach 
□ Step 1:  a complex theory of the ontic 
realities, ie the types of things  
□ Step 2:  fieldwork 
□ Step 3:  analysing early, & linking results 
□ Step 4:  keyness, discourses  
□ Step 5:  perhaps QCA analysis 
□ Step 6:  transparency:  database 
□ Step 7:  draw conclusions 
Qualitative Comparative 
Analysis 
Logic is 
□ Y = results arise from S, I, E, C, random error 
I = institutions, local entities 
C = context 
 A non-statistical approach. 
 
Is event E necessary, or sufficient for Y? 
 
Aims and Means of QCA 
Aims 
□ To focus on one 
outcome. 
□ How does the effect of 
X or T or E on that 
outcome change 
depending upon the 
contexts? 
□ Circumstances matter 
□ Measure to what 
extent it was the case. 
Means 
□ Insert a survey matrix 
into fsQCA freeware 
□ Produce tests of 
necessity of EACH 
condition for Y 
□ Then test for sufficient 
PATHWAYS. 
□ Test the results using a 
measure, or an F Test 
□ See my GITHUB 
freeware. 
Details of the QCA F-Tests 
1 We first define our terms and 
conceptual framework (S, I, E, X, Y, C) 
2 Empirical measure of Csuff 
(consistency for sufficiency of X for Y) 
3 Empirical measure of Goodness-of-fit 
(F-tests) for each pathway to Y 
See 
https://github.com/WendyOlsen/fsgof 
 
Amending the QCA for treatments, 
impacts of interventions 
□ In logic add ‘T’ as a new event 
□ Allow it to work as a ‘necessary’ cause 
(test) of higher levels of Y 
□ Allow it to be considered as a sufficient 
pathway for higher levels of Y 
□ Allow it to be considered as part of 
sufficient combination pathways for 
higher levels of Y 
Practical Example 
 
Applications to: 
□ Theorising Bangladesh Indebtedness  
□ Mediated by involvement in an NGO  
□ All NGOs are different; and 
□ Indian Women’s Work 
□ Mediated by involvement in  
 either the self-help groups, an MFI, 
   Or an NGO, or the Employment 
Guarantee Scheme 
Sample of Raw Debt Data – 
Bangladesh  
 
Results from QCA Part for India 
 
Reminder: Mixed Mode Data 
□ Step 1:  ontic exploration, list the types of things, name the key 
processes,  
□ SAMPLING:  Get samples which have CONTRASTS on BOTH 
X and Y 
□ AND ON T, the treatment event (low/high!) or (Yes/No) 
□ And on contextual factors (see leaflet) 
□ Make sure the qualitative cases are chosen from among 
the pre- and post-intervention sample cases. 
□ Step 2:  fieldwork 
□ Step 3:  analysing early, & linking results 
□ Step 4:  keyness, discourses  
□ Step 5:  perhaps QCA analysis 
□ Step 6:  transparency:  database 
□ Step 7:  draw conclusions 
 
Discussion 
  
Critiques and Responses 
□ RCT critique 
 
□ Unobserved heterogeneity critique 
 
□ Responses:  Complex differentiation of 
how causal mechanisms work 
Critique 2 
□ Endogeneity critique 
□ (it says that the key factors in your model 
can’t be distinguished from the irrelevant 
ones you have included because you’ve 
included too many factors) 
□ Responses: 
□ Complex interactions  do not ignore 
possible pathway reversal phenomena! 
□ That’s why statistics is weaker. 
□ Furthermore, be parsimonious in setting up the 
QCA explanatory model. 
 
Conclusions 
□ Ontic complexity 
□ Teamwork  
□ Combining the keyness stage with a selective 
interpretation stage; and  
□ Add A QCA or Fuzzy Set QCA Stage. 
□ Models and results are debated in an ongoing, 
open-ended way. 
□ We try to make the interpretation match, 
complement or contradict the original 
Research Question. 
□ Be rigorous and transparent. 
41 
Acknowledgements- 
Collaborators John McLoughlin at Univ of 
Manchester, Samantha Watson at 
Flowminder, University of Southampton 
□ John has programmed in Python to break up 
the British National Corpus into parts and put 
them into NVIVO.  Counting the word 
frequencies in Baby BNC in NVIVO, he then 
compared these with the word frequencies 
in each qualitative data set. 
□ See GITHUB for the programme, searching on 
either John McLoughlin or Wendy Olsen. 
42 
See Also:   
□ See also a calibration example at:  
https://www.facebook.com/groups/mi
xednetwork/ 
□ Integrated Mixed Methods Network 
□ And many examples of QCA and 
Fuzzy Set Analysis of Cases at 
www.compasss.org   (sic) 
□ And JISCMAIL  QUAL-COMPARE (190 
members) email list.  Free to join. 
Key References 
 
□ Heale, R., and D. Forbes (2013), “Research Made Simple: Understanding 
triangulation in research”, Evid Based Nurs 16:4 98 doi:10.1136/eb-2013-101494.  
□ Morgan, J., and W.K. Olsen (2011) “Aspiration Problems for the Indian Rural Poor: 
Research on Self-Help Groups and Micro-Finance”, Capital and Class, June, 35:2,  
189-212, doi: 10.1177/0309816811402646;  
□ O’Cathain, A., E. Murphy, and J. Nicholl (2010)  “Three Techniques for Integrating 
Data in Mixed Methods Studies”, British Medical Journal,341: , doi: 
10.1136/bmj.c4587 
□ Olsen, W.K., and J. Morgan (2010) “Institutional Change From Within the Informal 
Sector in Indian Rural Labour Relations”, International Review of Sociology, 20:3, 
535-553, on “The Regulation of Indian Informal Capitalism”, 
DOI:10.1080/03906701.2010.51190.) 
□ Touri, M., and N. Koteyko (2014) “Using Corpus Linguistic Software in the Extraction 
of News Frames: Towards a dynamic process of frame analysis in journalistic texts”, 
International Journal of Social Research Methodology, DOI: 
10.1080/13645579.2014.929878 
□   
 
44 


Introducing Keyness and Discourse Analysis in Mixed Methods Research

Desai, Tanvi

Disseminate data not access

Lienert, Jeffrey

Analyzing healthcare 
sequences in social networks
Jeff Lienert, M.S.
ESRC Research Methods Festival
3 July, 2018
Background
• When a person becomes infected and notices 
symptoms, they can take multiple actions
• Depending on the outcome, they can follow-up 
with other actions
• Which actions they take are dependent on many 
things
• Actions they take affect their outcome
• May also affect others
Background
Step 1
•Self-care / rest
•At home
•3 days
•Called relative to 
talk about 
discomfort
Step 2
• Informal healer
•At home
•8 days
•Paracetamol (1 
daily, for 3 days)
Step 3
•Public county 
hospital
• > 2 hours from 
home
• 14 days
• Called taxi
Step 4
•Private doctor
•10 minutes from 
home
•1 day
•Phenoxymethyl-
penicillin (half 
dose daily, for 1 
day)
Step 5
•Self-care / rest
•At home
•7 days
•Phenoxymethyl-
penicillin (half 
dose daily, for 7 
days)
Process Data
Symptoms | Diagnosis | Self-perceived severity | Total duration | Start date | Current status | People involved
Background
• Actions can be analyzed in a number of ways
• Counts
• Event sequence
• Network
• Each gives different information
• Here, we focus on the network approach, and what 
it offers over the others
Participant demographics
Making the network
Data Network
Making the network
Data Network
Making the network
Data Network
1
Making the network
Data Network
1
Making the network
Data Network
1
2
Making the network
Data Network
1
2
Making the network
Data Network
1
2
Making the network
Data Network
1
2
3
Making the network
Data Network
1
2
3
Making the network
Data Network
1
2
3
Making the network
Data Network
1
2
3
Making the network
Data Network
1
2
3
All steps
Step 1
Step 2
Step 3
Step 4
Step 5
Step 6
Step 7
Step 8
Step 9
Centrality measures
Centrality measures
• Common first steps
• Doing nothing
• Pharmacist
• Primary care
• Common last steps
• Self-care
• Care from family and friends
• Common intermediate steps
• Self-care
• Care from family and friends
Correlation of centrality and stage
Conclusions
• Network analysis provides important information 
about event history of infection healthcare-seeking 
behaviors
• Network centrality parameters independent of evet 
sequence parameters
• The small number of those seeking traditional 
healers have unique trajectory
Next steps
• Dynamic network modeling (SAOMs)
• Multilevel network modeling to understand how 
social network influences event sequence and vice-
versa
Acknowledgements
PI: Marco J Haenssgen a, b, c
Proochista Ariana a
Caroline OH Jones f
Rachel Greer e
Yoel Lubell c
Mayfong Mayxay f
Paul N Newton f, g
Felix Reed-Tsochas b, h, i
Heiman FL Wertheim j, k 
Giacomo Zanello l, m
a. Centre for Tropical Medicine and Global Health, Nuffield Department of Medicine, University of Oxford, Oxford, UK; b. CABDyN
Complexity Centre, Saïd Business School, University of Oxford, Oxford, UK; c. Economics and Translational Research Group
(ETRG), Mahidol Oxford Tropical Medicine Research Unit (MORU), Faculty of Tropical Medicine, Mahidol University, Bangkok,
Thailand; d. Department of Health System and Research Ethics, KEMRI Wellcome Trust Research Programme, Kilifi, Kenya; e.
Chiangrai Clinical Research Unit, Chiangrai Prachanukroh Hospital, Chiang Rai, Thailand; f. Lao Oxford Mahosot Wellcome Trust
Research Unit (LOMWRU), Mahidol Oxford Tropical Medicine Research Unit (MORU), Faculty of Tropical Medicine, Mahidol
University, Bangkok, Thailand; g. Faculty of Postgraduate Studies, University of Health Sciences, Vientiane, Lao PDR; h. Institute
for New Economic Thinking, Oxford Martin School, University of Oxford, Oxford, UK; i. Department of Sociology, University of
Oxford, Oxford, UK; j. Oxford University Clinical Research Unit (OUCRU), Ho Chi Minh City, Viet Nam; k. Medical Microbiology
Department, Radboudumc, Nijmegen, Netherlands; l. School of Agriculture, Policy and Development, University of Reading,
Reading, UK; m. Leverhulme Centre for Integrative Research on Agriculture and Health, London, UK
References
• Allison, P. D. (2014). Event History and Survival 
Analysis. Thousand Oaks, CA: Sage.
• DuBois, C., Butts, C. T., McFarland, D., & Smyth, P. 
(2013). Hierarchical models for relational event 
sequences. Journal of Mathematical Psychology, 
57(6), 297-309. doi: 10.1016/j.jmp.2013.04.001
• Haenssgen, M. J., & Ariana, P. (2017). Healthcare 
access: a sequence-sensitive approach. SSM -
Population Health, 3, 37-47. doi: 
10.1016/j.ssmph.2016.11.008


Analysing healthcare sequences in social networks

Ding, Xuejie

Department of Sociology
A Lifecourse Approach to 
Genetic Endowment and 
Cognitive Decline
Xuejie Ding
July 5th, 2018
Department of Sociology
The Puzzle
• Cognitive status decline with age, but some people experience 
faster decline than others (Stern,2009)
• Educational attainment as cognitive reserve is related to better level 
of cognitive status
• Improve problem solving, stress coping abilities
• Improve effectiveness in using brain network
• Mixed evidence on education and rate of cognitive decline
• Education and cognitive ability has shared genetic bases (Okbay et 
al., 2016; Lee et al., in press)
• Current research seldom account for genetic markers
Department of Sociology
Research Questions
• How does the association between genetic endowment 
of education and rate of cognitive decline change with 
age among older adults (50+) in the US?
• After accounting for genetic endowments, what are the 
effects of sociodemographic and behavioral factors?
Department of Sociology
Three competing hypotheses
• Active reserve model:
• Higher education, slower rate of 
decline (divergence)
• Passive reserve model:
• Higher education has no effect on rate 
of cognitive decline (stable)
• Compensation model:
• Higher education, faster rate of decline 
(Convergence)
Department of Sociology
Sample
• Health and Retirement Survey 1998-2012
• a biennial, longitudinal survey of a nationally-
representative sample of individuals and their 
spouses aged 50 and above. 
• Genetic information is collected in 2006 and 2008
• 5,005 European and non-Hispanic respondents 
with at least 4 cognitive interviews 
Department of Sociology
DV – Cognitive Measure
• Episodic Memory (EM)
• Encodes and retrieves personal experienced events occurred at 
specific time and place
• Delayed word recall 
• Working Memory (EM)
• Remember and use memory while in the middle of an activity
• Serial 7s subtraction test 
• Mental status (MS)
• Knowledge-related cognitive ability
• naming the date, month, year and day of the week, backwards 
counting from 20, object naming, and naming the current president 
and vice president of the U.S. 
• Vocabulary (Vocab)
• Define 5 words
• Global cognition (GC)
• Average corrected percentage score: (score/maximum score)*100
Department of Sociology
IV
• Polygenic score for education (Okbay et al., 2016), 
standadised
• Inverse probability weight (Domingue et al., 2017)
• Years of schooling
• Gender, social interaction, health behavior, chronic 
diseases
• Analytical Strategy: Linear mixed models
𝐶𝑜𝑔𝑛𝑖𝑡𝑖𝑜𝑛'( = 𝛽+ + 𝛽- ∗ 	𝐴𝑔𝑒( + 	𝛽2 ∗ 𝑃𝐺𝑆𝑒𝑑𝑢' +
𝛽8 ∗ 𝑃𝐺𝑆𝑒𝑑𝑢' ∗ 𝐴𝑔𝑒( +	𝛽9 ∗ 𝛸'( 	+ 	𝛽; ∗
𝛸'( ∗ 𝐴𝑔𝑒( 	+	𝜇'( +		𝜇'( ∗ 	𝐴𝑔𝑒( + 𝜀>
PGS for individual 𝑖 are calculated as the 
sum of the allele counts 𝑎𝑖𝑗	(0, 1, or 2) 
for each SNP	𝑗 = 1,…𝑀, weighted by 
association strength 𝑝𝑗:
𝑃𝑜𝑙𝑦𝑔𝑒𝑛𝑖𝑐	𝑠𝑐𝑜𝑟𝑒 = 	∑ 𝑝𝑗𝑎𝑖𝑗𝑀𝑗=1
Department of Sociology
Results
Department of Sociology
Results – Linear Mixed Models
• Higher education polygenic score is 
associated with better cognitive performance 
(β Range from 0.08 – 2.44) 
• PGS education on rate of cognitive decline 
varies across domains
• Strength of association declines for EM, 
GC (Compensation); 
• Magnifies for MS (Active Reserve); 
• Remains constant for WM, and Vocab 
(Passive)
Department of Sociology
Results
Department of Sociology
Results
• Women performs better in GC, EM, Vocab, worse in 
WM, no difference in MS. Women also decline faster 
in GC, EM, WM.
• Years of schooling positively related to cognitive 
ability, but is associated with faster decline in EM, 
MS, and slower decline in WM.
• Social interaction: slower decline in GC, EM and MS.
• Smoking: faster decline in MS.
• Multi-morbidity in chronic diseases: faster decline in 
all measures.
Department of Sociology
Discussion
• Is the result due to insufficient control? 
(Keller, 2014)
• Limitations: floor and ceiling effect, 
mechanisms from a biological angle, other 
race and ethnicity.


A lifecourse approach to genetic endowment and cognitive decline

This presentation demonstrates and discusses the development of sensor-generated data, and how these can be visualised and analysed for a range of purposes

Jiang, Jie

Sensor-generated data: visualisations and analytic options | Dr Jie Jiang

Millard, Stephanie

Nind, Melanie

MethodsNews 2018: 2

In a special mini-series of our podcast, we discuss the secondary analysis of large-scale qualitative data as a new research method. Here Ros Edwards and Sarah Lewthwaite from the University of Southampton talk about how the method might benefit and challenge teachers and students of research methods

Edwards, Rosalind

Lewthwaite, Sarah

NCRM podcast: Teaching Big Qual: benefits and challenges for students and teachers

Susan Banducci and Iulia Cioroianu talks at the NCRM Autumn School 2017
https://www.ncrm.ac.uk/autumn

Banducci, Susan

Cioroianu, Iulia

Online data sources: linking old and new, big and small

Professor Patrick Sturgis in the fifth (of six) part of the ‘Structural Equation Modelling (SEM): What it is and what it isn't’ online course. This video is part of the online learning resources from the National Centre for Research Methods (NCRM). To access all videos and the supporting materials (presentation slides, datasets, recommended reading, links to related publications and resources) visit hhttps://www.ncrm.ac.uk/resources/online/SEM2016

National Centre for Research Methods

Not a member yet