New Investigators Network

1
New Investigators Network

• The Longitudinal Study of Australian Children

2
Overview

• Introduction to the LSAC Datasets
• Data analysis issues
• Variable naming
• Data user resources

3
Introduction to the LSAC Datasets
4
Datasets

• At last count 23 datasets
• Dont panic!
• 4 main datasets
• 1 for each age at which weve interviewed
  children
• lsacconf0, lsacconf2, lsacconf4, lsacconf6

5
Main LSAC datasets

• Contain everything not excluded for good reason
• Identifiers
• Question variables
• Derived items
• Linked data
• Between waves questionnaire data

6
Supplementary Datasets

• Time Use Diary (TUD)
• Medicare Australia linked data
• Household composition data
• Wave 1 original format datasets

7
TUD datasets

• Cleaned datasets
• Imputations to improve data quality
• Poor quality cases deleted
• Poor quality cases
• Imputations to improve quality made
• Uncleaned datasets
• No improvements or deletions

8
Household Composition

• Detailed information on every person living with
  study child at time of interview
• Less detailed information on those who lived with
  study child between waves
• Parent 1, Parent 2, Mother, Father variables
• Derived items
• Subset of this information also on the main
  dataset (P1, P2, SC, Mother, Father, derived
  items)

9
Medicare Linked Data

• Medicare Benefits Scheme (MBS)
• Each record represents a claim
• Pharmaceutical Benefits Scheme (PBS)
• Each record represents a claim
• Australian Childhood Immunisation Register (ACIR)
• Each record represents an immunisation
• Note data is just for child, no data for other
  members of the household

10
Wave 1 formatted data

• Changes made between Wave 1 and Wave 2 data
  releases
• Renamed variables
• Merging of centre and home-based carer
  questionnaire data
• Included to help those with old code
• Will be phased out by Wave 3

11
General Data Analysis Issues
12
Data analysis general issues

• Confidentialisation
• Weighting
• Clustering

13
Confidentialisation

• In order to protect the anonymity of respondents
  some changes have been made to the dataset
• Removal of items
• Transformation of items
• Top-coding (ie recoding outliers)
• Aggregation

14
Removal of items

• Qualitative data provided by responses
• Other specify
• Other open ended text questions
• Linked census data for childcare/education
  arrangements

15
Transformation of items

• Postcode -gt postcode ID numbers
• Date left hospital after birth -gt time in
  hospital after birth

16
Topcoding

• Income
• Housing costs
• Child support
• Physical measurements
• Hours in childcare

17
Aggregation

• Occupation ASCO4 -gt ASCO2
• SEIFA rounded to the nearest 10
• Country of birth coded 0 if fewer than 5 cases
• LOTE coded 0 if fewer than 5 cases
• Date of birth changed to month of birth
• (Also religion, but hasnt been necessary so far)

18
Weighting
19
Weighting

• Particularly acknowledge the contribution of
  David Lawrence of Institute for Child Health
  Research
• See also
• LSAC Technical Paper 3 Wave 1 Weighting and
  Non-response
• LSAC Technical Paper 5 Wave 2 Weighting and
  Non-response

20
Why weight?

• To cut down on error due to
• Sampling
• Not using a simple random sample
• Non-response
• Unlikely to be random

21
Steps

• Design weights
• Wave 1 Non-response analysis
• Final wave 1 weights
• Wave 2 non-response analysis
• Final Wave 2 weights

22
Design Weights

• Aim for sample design to introduce as little bias
  as possible
• Differential probability of selection is a
  known response bias
• Can be adjusted for perfectly (or close to)

23
Large postcodes - probability

• For most purposes design weight is calculated as
  the inverse of the probability of selection
• Pc Psi x Ppi
• Where
• Pc is the probability a child will be selected
• Psi is the probability a childs postcode will be
  selected from all postcodes in the stratum
• Pci is the probability a child will be selected
  from all the children in their postcode

24
Implications

• Weights become extreme if
• Large discrepancy between postcode population at
  time of postcode selection and child selection
• Large discrepancy between number of B and
  K-cohort children

25
Small postcodes - why not probability?

• Extra children selected from postcodes with gt20
  kids to represent postcodes with lt20 kids
• Using the probability would lead to
  underweighting
• Set number of postcodes selected based on average
  size
• Probability doesnt account for differences
  between final number of selections and target
  number

26
What did we use instead?

• D2 Nr
• ns
• Where
• D2 is the design weight for children in size 2
  strata
• Nr is the total number of children in the stratum
  minus those in remote postcodes
• ns is the number of children selected in the
  stratum

27
Implications

• Assumes kids from ineligible postcodes are
  similar to those from eligible ones
• Weights become extreme when select more of the
  larger or smaller postcodes within the stratum
  (basically random)

28
Wave 1 Non-response

• What do we know about non-respondents?
• We know where they live
• We know how many of their neighbours responded to
  the study
• We have census data on their neighbourhoods

29
Poisson regression

• Models the chance that something will happen a
  given number of times
• For our purposes response rate within a postcode
• Identifies a minimum set of predictors
• Variables must be measured by the census and LSAC
  in a consistent way
• Must have variation by postcode (e.g. gender of
  child inappropriate)

30
B-cohort response rate regression
31
K-cohort response rate regression
32
Population Benchmarks

• ABS ERP 31 March 2004 - State and gender
• ABS ERP 30 June 2003 - Part of State
• Medicare database - Adjusted for exclusion of
  some remote postcodes
• 2001 Census - Proportion of mothers who had
  completed Year 12 and spoke a LOTE (at state X
  part of state level)

33
Standard weighting

• Classify cases into cells based on all variables
  weighted for
• Obtain frequencies for each cell and compare with
  population figures
• WeightPopulation/Sample for each cell
• Requires cross classified benchmarks
• Unreliable when sample sizes get small

34
Calibration of weights

• Generalised raking procedure (Deville and
  Sårndal, 1992)
• Adjusts design weights
• Searches for a solution that
• Weights correctly to each population benchmark
• Has smallest distance between original weights
  and final weights

35
Advantages

• Only requires benchmarks for each weighting
  variable separately
• Range of weights can be constrained

36
Wave 2 non-response

• What do we know about Wave 2 non-respondents?
• Quite a bit
• 2,000 odd variables in the main LSAC Wave 1
  dataset
• However, information is 2 years old

37
Wave 2 adjustment

• Adjust weights based on probability of response
  in Wave 2
• Probability based on logistic regression
• Over 50 variables used as predictors selected
  based on
• Little or no missing data
• Likelihood of prediction of non-response
• Coverage of topics included in the survey

38
Weights provided

• LSAC provides 3 weights per wave per cohort
• Sample weight
• Weights add to sample size
• Useful when using N to calculate test of
  significance
• Population weight
• Weights add to population totals
• Useful when need frequencies to produce
  population estimates
• Day weight
• Sample weight adjusted for day-of-the-week of
  diary completion

39
Clustering
40
Clustering

• Usual tests of significance assume independence
  of observations
• Not the case for LSAC sample design as children
  selected from postcodes
• Therefore standard errors need to be adjusted

41
Design effect

• Is the standard error if design accounted for
  divided by standard error is random sampling
  assumed
• Will be small if the variable is unrelated to
  postcode (e.g. study child gender)
• Will be larger if the variable is related to
  postcode (e.g. neighbourhood characteristics,
  housing costs, income)

42
SAS

• Use survey procedures (eg proc surveymeans,
  proc surveyreg)
• proc surveyreg dataltfilenamegt totalltlibrarygt.str
  atum
• stratum stratum
• cluster pcodes
• model ltstandard SAS model detailsgt
• weight lta/b/c/dweightsgt
• run

43
SPSS

• Need complex samples add-on
• Set up an analysis plan with an analysis plan
  wizard
• Stratastratum
• Clusterpcodes
• Weighta/b/c/dweights
• Use plan file to run analyses

44
Variable Naming
45
Variable Naming

• Standard input variables
• Derived items
• Age invariant indicators
• Household composition items

46
Standard input variables

• The variable names follow the standard format in
  most cases.
• A tt xxxxx
• A child age indicator
• tt topic indicator
• xxxxx specific question identifier

47
Child Age Indicator

• a indicates the child is aged 0-1 years
• b indicates the child is aged 2-3 years
• c indicates the child is aged 4-5 years
• d indicates the child is aged 6-7 years
• z indicates that the data item is permanent and
  will not change (e.g. date of birth)

48
Some examples

• zhs03a is the birthweight of the study child for
  both B and K cohorts
• apa01aParent 1 rating of parent Self-efficacy at
  Wave 1 for the B-cohort
• bpa01aThe same thing at Wave 2 for the B-cohort
• cpa01aThe same thing at Wave 1 for the K-cohort

49
Topic Indicator

• 2-character abbreviation
• Eg. fdFamily Demographics
• Links in with topic field in the Data Dictionary
• Complete list of topics available in the Data
  User Guide

50
Specific question identifier

• The last 5 characters of a variable name
• Contain whatever information is necessary to
  uniquely identify each item.
• Generally has an arbitrary two-digit question
  number
• Items of related content have been grouped
  together as much as possible.

51
Subject/informant - general

• The 6th character in the variable name can also
  indicate the subject/informant
• aParent 1
• bParent 2
• cStudy Child
• pPLE
• mMother
• fFather
• tTeacher/carer
• iBetween waves mail-out respondent

52
Some examples

• bhs13a is Parent 1s rating of their own overall
  health status
• bhs13b is Parent 2s rating of their own overall
  health status
• bhs13c is Parent 1s rating of the Study Childs
  overall health status
• bhs13p is the PLEs rating of their own overall
  health status
• bhs13m is the Mothers rating of their own
  overall health status
• bhs13f is Fathers rating of their own overall
  health status

53
Subject/Informant Education/Child Care
54
Example

• Grouping of like items
• bhs12a is whether Parent 1 is concerned about the
  childs weight
• bhs12b is whether Parent 1 considers the child to
  be underweight, normal weight, somewhat
  overweight or very overweight

55
Item grouping

• Grouping of items (with subject indicator)
• bhs23c1 is the study childs height
• bhs23c2 is the study childs weight
• bhs23c3 is the study childs waist measurement

56
Item grouping

• Scales and subscales grouped together
• SDQ Conduct reported by Parent 1 at Wave
  1cse03a4a, cse03a4b, cse03a4c, cse03a4d,
  cse03a4e
• SDQ Conduct reported by Parent 1 at Wave
  2dse03a4a, dse03a4b, dse03a4c, dse03a4f,
  dse03a4g
• Note that items d and e at Wave 1 have been
  replaced by items f and g at Wave 2

57
Item grouping

• Recoded items
• Changes between cohorts and waves

58
Derived items

• A(S)m
• A Age indicator
• (S) optional subject/informant indicator
• m up to 6 character mnemonic

59
Examples

• bvocab MCDI Vocabulary measure (B-cohort Wave
  2)
• aaemp Parent 1 employment status (B-cohort Wave
  1)
• abemp Parent 2 employment status (B-cohort Wave
  1)
• bbemp Parent 2 employment status (B-cohort Wave 2)

60
Age invariant indicators

• A small number of age invariant indicators also
  on the file
• hicid unique identifier assigned when child was
  selected by Medicare Australia
• cohort
• wave
• stratum stratum from selection process
• pcodes postcode at the time of selection

61
Household Composition Items

• A f xmmm
• Where
• A Child age indicator
• f f (for family)
• Question number (numeric)
• x Sub-question indicator (optional)
• mmm person identifier

62
User Resources
63
User Resources

• Labelled Questionnaires
• Data Dictionary
• Frequencies
• Are located on the data CD and on the website
• www.aifs.gov.au/growingup/pubs/instruments.html

64
Labelled Questionnaires

• The interview for wave 1 was conducted using
  paper and pencil.
• The interview for wave 2 was conducted using
  Computer Assisted Interview (CAI) and paper and
  pencil.
• Therefore in wave 2 a CAI guide had to be created
• This guide does not include the Household
  variables. For these variables refer to the data
  dictionary

65
Labelled Questionnaires

• Survey question and its associated variable
• E.g. Question - Was child ever breastfed?
  Variable - zhb05a

66
Data Dictionary

• Information about every variable on the main
  datasets can be found here.
• There is 2 versions of the data dictionary one in
  Excel and one online.
• This allows users to use functions associated
  with Excel when navigating the data dictionary.

67
Frequencies

• Frequencies have been created for each of the
  datasets
• These frequencies can be found on the data CD
• Frequencies on all variables have been prepared
• If the variable is a continuous variable the mean
  and the standard deviation was obtained in order
  to reduce the number of pages.