A two mode personal network method
for creating categories of knowing
Christopher McCarty
H. Russell Bernard
University of Florida
Dimitri Fazito
Universidade Federal de Minas Gerais
Sunbelt XXXI, St. Pete Beach, FL February 8-13 2011
NSUM
• The network scale-up method (NSUM) is based on a four part
equation.
•
m/c=e/t
• t is the total size of a population
• e is the size of a sub-population in E that we want to estimate
• m is the average number of people in e that each member of our
sample knows
• c is the average personal network size.
• Each person’s network (c) reflects, with some deviations the
distribution of various populations, e’s, in the total population t,
and the deviations average out if we study a large, representative
sample of people.
Testing NSUM
• We tested NSUM in the U.S. in seven surveys, using two
methods to estimate c:
– 1. The known population method: Asking people how
many people they know in 29 populations of known size
and estimating c using a maximum-likelihood method. (See
Killworth et al. 1998)
– 2. The summation method: Asking people how many
people they know in each of 17 relation categories –
people in their immediate family, people who are coworkers, etc. – and summing to find c. (See McCarty et al.)
• Both methods produced an average network size of 290 (sd
232, median 231).
•
•
•
1998 P. D. Killworth, C. McCarty, H. R. Bernard, G. A. Shelley, and E. C. Johnsen. Estimation of Seroprevalence, Rape and Homelessness in the U.S. Using a Social Network Approach.
Evaluation Review 22:289–308.
McCarty, C., P. D. Killworth, H. R. Bernard, E. Johnsen, and G. A. Shelley. Comparing Two Methods for Estimating Network Size. Human Organization 60:38–39
H Russell Bernard, Tim Hallett, Alexandrina Iovita, Eugene C Johnsen, Rob Lyerla, Christopher McCarty, Mary Mahy, Matthew J Salganik, Tetiana Saliuk, Otilia Scutelniciuc, Gene A Shelley,
Petchsri Sirinirund, Sharon Weir, Donna F Stroup (2010) “Counting hard-to-count populations: the network scale-up method for public health” Sexually Transmitted Infections, 2010 86:
ii11-ii15.
Applications of NSUM
• NSUM was developed to understand who people
know and how they know each other.
– It is used today to estimate the size of hard-to-count
populations, like populations at risk for HIV/AIDS and illegal
migrants.
– Where good, trackable statistics are available for many
populations of known size, the known population method
is preferred.
– In countries where good statistics on populations of known
size are lacking we rely on the summation method for
estimating c.
Finding relation categories
• In the U.S., the categories for the summation
method were derived from ethnography and
from experiments we did on how people know
one another.
• The result was the reliable estimate reported
above of 290 for c, across seven surveys.
The Challenge
• NSUM is of most interest
to public health officials
in countries with nonIndo-European languages.
• How do we insure that
the summation categories
are mutually exclusive
(alters are not doublecounted) and exhaustive
(everyone in the network
is included)?
The Solution
• We need a method that can discover categories
of knowing that are mutually exclusive and
exhaustive.
• The method must be able to discover categories
in the language of the respondent without preconceived categories as cues.
• We apply two methods developed in cognitive
science: free listing and frame substitution.
•
•
Frake, C. O. 1964. Notes on queries in anthropology. In Transcultural studies in cognition, A. K. Romney and R. G. D’Andrade, eds. American Anthropologist 66, Part II.
Rosch, Elizabeth 1975. Cognitive representations of semantic categories. Journal of Experimental Psychology 104:192–233.)
One-Mode Personal Network
• Typically, we use name
generators to elicit the
names of alters.
• Respondents provide
information on their
alters, including the ties
between them.
• The result is a one-mode
network of ties between
actors.
Two-Mode Personal Network
• In contrast, two mode networks represent ties
between actors and situations.
• For a personal two-mode network we elicit
alter names from respondents using a name
generator.
• Respondents then answer whether each alter
corresponds to some event.
Reasonable and unreasonable twomode event questions
• Some questions would not be reasonable as
we would not expect respondents to be
accurate in reporting about all of their alters:
– Attendance at meetings
– Places they shop
• Respondents can report accurately on the way
they perceive their alters
– How you know them
Method – Step 1
• Twenty one participants
at an NSUM workshop in
Thailand free-listed the
words in Thai that
describe how people
know each other.
• The Thai terms were
ordered by frequency.
• We cut off the terms at
those that were
mentioned at least three
times resulting in 26
categories.
Method – Step 2
• From each of the 21
respondents we elicited a
network of 30 alters using the
following name generator
• “You know them and they
know you by sight or by name,
you have had some contact in
the past two years and you
could contact them now.”
• For each alter the respondent
then evaluated if each of the
26 categories applied to them
or not.
colleague
ปอนด์
นุช
เพ็ ญ
พี่ยู
หมี
อาจารย์นิ
อาจารย์อมรา
พี่นด
ิ
มด
พี่จม
ุ๋
พี่ภา
พี่จวิ่
น ้าช่วย
อาจารย์มานพ
วรา
โจ ้
สุทป
ี
พี่ยาว
พี่เกด
ส ้ม
เกด
พี่เหว่า
เอ๋ย
ปิ ง
เล็ก
น ้าม่อน
ป้ าขวด
นุ ้ย
household
0
1
0
0
1
1
1
1
1
0
1
1
1
0
0
0
1
0
0
0
1
1
0
0
0
0
0
0
neighbour
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
sport club/ park
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
meeting
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
1
1
0
1
1
1
1
0
1
0
0
0
1
0
0
0
0
1
1
0
0
0
0
relatives temple/ church
same community
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Method – Step 3
• We created 21 category by
category matrices and
summed these one- mode
matrices into a single one
mode matrix .
• The numbers in the cells
indicate the number of
times the two terms were
used for the same alter.
• High numbers indicate high
overlap between terms; low
numbers indicate low
overlap.
Results - Objective
• Ultimately we want a set of categories that are
culturally salient in the language of the
respondent (in this case, Thai).
• Categories that have high overlap can then be
collapsed.
• This will produce a set of mutually exclusive
and – we hope – culturally salient categories.
Results – Unconstrained graph
• We treat the affiliation matrix as
a network and use the spring
embedder program in NetDraw
(available in UCINET) to visualize
the connections in the matrix
• This visualization shows the
overlap between categories. A
line exists if even one alter is a
member of the two categories
• There is one large component
and four isolates
• These isolates are candidates for
mutually exclusive categories
• We need to identify which
categories are functionally
overlapping so they can be
consolidated
Unconstrained
Greater than 1
Greater than 2
Greater than 3
Greater than 4
Greater than 5
Greater than 6
Greater than 7
Greater than 8
Greater than 9
Greater than 10
Greater than 96
Distribution of ties between categories
•
•
•
•
Mode – 0
Median – 0
Mean – 2.38
How much overlap
should we tolerate?
• Look at gaps between
overlap values
Gaps between overlap values
• This graph shows the
distribution of the gap
between overlap
values.
• A Very large number of
overlaps do not occur
until values 67 and 96.
• Smaller but noticeable
increases occur at
number 17 and number
31.
Visual using constrained tie definitions
Tie for overlap >= 17
Tie for overlap>=41
To avoid duplicate alter nominations with the summation
method we would look for categories (or sets) to collapse
Tie for overlap >= 17
Tie for overlap>=41
Future Directions
• Develop other quantitative methods to decide
where the tolerance for overlap should be.
• For example: present 20 native speakers of Thai
with a pack of 26 cards, each with the name of
one category.
• Free pile sort these cards and run consensus
analysis (using the informal model) to see if there
is agreement about the way the categories
should be sorted.
– If there is agreement, we ask the same or other native
speakers to name the piles.
Descargar

A two mode personal network method for creating …