network.survival.estimator: network survival estimator

In dfeehan/networkreporting: Tools for using Network Reporting Estimators

network survival estimator

Description

use an aggregate multiplicity estimator and the respondents' own network size estimates to estimate hidden population sizes

Usage

network.survival.estimator_(
  resp.data,
  attribute.data,
  attribute.names,
  known.populations,
  total.kp.size = 1,
  weights,
  attribute.weights,
  within.alter.weights = NULL,
  boot.weights = NULL,
  ego.id = NULL,
  return.boot = FALSE,
  dropmiss = FALSE,
  verbose = TRUE
)

network.survival.estimator(
  resp.data,
  attribute.data,
  attribute.names,
  known.populations,
  total.kp.size = 1,
  weights,
  attribute.weights,
  within.alter.weights = NULL,
  boot.weights = NULL,
  ego.id = NULL,
  return.boot = FALSE,
  dropmiss = NULL,
  verbose = TRUE
)

Arguments

resp.data	The dataframe that has a row for each respondent, with reported connections to the groups of known size, as well as the attributes. Note that the column names of the attributes should match their names in attribute.data
attribute.data	A dataframe with the reported attributes of hidden population members reported by survey respondents. There should be one row for each time a respondent reports a hidden population member. For example, to estimate death rates, there should be one row for each report of a death.
attribute.names	The names of the columns of attribute.data and resp.data that contain the attribute information.
known.populations	The names of the columns in resp.data that have responses to the known population questions
total.kp.size	The size of the probe alters, i.e., the sum of the known population sizes
weights	The weights or weights column for the respondent data
attribute.weights	The weights or weights column for the alter data (this should typically be the weight of the respondent who reported the alter)
within.alter.weights	The weight or weights column for within-alter weights. This could be useful if, for example, respondents only report about a subset of all of their alters. See Details.
boot.weights	Optional dataframe with bootstrap resampled weights. See Details for more info.
ego.id	If boot.weights are included, then this is the name of the column(s) we need to join the bootstrap weights onto the dataset. This is most often the id of the ego making the reports.
return.boot	If TRUE, and if boot.weights is specified, then return each bootstrap estimate
dropmiss	How to handle missingness in reported connections to known populations and number of deaths. See report.aggregator
verbose	If TRUE, print information to screen

Details

This function takes two sources of data as input: first, it requires a long-form dataframe with the attributes of the reported members of the hidden population. For example, if we are asking about emigres and we collect the age and sex of each reported emigrant, then the long form dataset might look like:

age	sex	weight
15	m	2.10
58	f	1.15
33	m	3.67

The second source of data we need is the known population responses for the respondents, along with the *same* attributes for each respondent. For example, in the situation above, we would also require a dataset like this to be passed in

age	sex	weight	hm.teachers	hm.nurses	...
20	f	2.10	4	0	...
44	m	1.65	0	2	...
60	m	2.75	1	1	...

Value

the network reporting estimate of the hidden population's size (as a prevalence) broken down by the categories defined by all combinations of attribute.names.

Technical note

This function assumes that the sampling weights are standard analysis weights and *not* relative weights. Standard analysis weights should provide an estimate for the size of the frame population when added up; relative weights, on the other hand, will sum to the number of respondents in the sample. Demographic and Health surveys typically have relative weights, which must be converted into standard sampling weights before using this function.

Returned dataframe

Currently, the dataframe that is returned has columns for the attributes that death rates were calculated for, along with:

• num.obs.deaths - number of deaths reported

• num.obs.degree - number of respondents in cell used for estimating degree

• y.F.Dcell.hat - estimated total number connections from the frame to deaths

• y.Fcell.kp.hat - estimated total number of connections from (frame intersect cell) to the groups of known size

• total.kp.size - the total size of the groups of known size

• N.Fcell.hat - estimated size of (frame intersect cell) population (based on survey weights)

• N.F.hat - estimated size of frame population (based on survey weights)

• asdr.hat - the estimated death rate

TODO

• handle missing values

• allow passing in N.F (currently, we always estimate from weights)

• write more general agg mult est fn and call that

• make unit tests

Details

If you want estimated sampling variances, you can pass in a data frame boot.weights. boot.weights is assumed to have a column that is named whatever the ego.id is, and then a series of columns named boot_weight_1, ..., boot_weight_M.

ego.id can either be a string or vector of strings, or it can be a named vector like c('a'='b'). In the second case, 'a' should be the name of the id column in resp.data, while 'b' should be the name of the id column in 'attribute.data'.

within.alter.weight is a weight for alters within a respondent. This is set to NULL by default, and many applications will not need it. However, it can be helpful if respondents are only asked to report about a subset of their alters. For example, suppose that respondents can report about 10 alters in detail. Respondent A reports connections to 3 deaths, and respondent B reports connections to 12 deaths. When aggregating Respondent A's detailed reports, we would only need to use the sampling weight. However, if we only used the sampling weight when aggregating Respondent B's reports, that would imply that Respondent B only reported connections to 10 deaths. In this case, setting within.alter.weight to (12/10) for Respondent B will make B's reports imply that she reported 12 deaths. Note that this makes the assumption that the 10 deaths Respondent B did report are a uniformly random subsample of the 12 deaths she reports being connected to.

dfeehan/networkreporting documentation built on May 17, 2023, 1:06 a.m.