OBJECTIVE
The objective in this study was to explore data
on employment and commuting from different
sources, using statistical analytic techniques
together with geographical experts to obtain
information to be provided to modelers in order
to help them improve the employment and
commuting component of their models,
determine potential issues related to these data,
and identify problem areas where further
investigation is needed.
BACKGROUND
Evidence suggests that transmission within the
workplace contributes significantly to the
magnitude of a pandemic flu epidemic. A
significant number of large organizations have a
pandemic plan in place which may help in
controlling this manner of transmission. These
plans typically include telecommuting and other
measures to reduce the need to physically
commute to the workplace. Good data are
needed in order to obtain valid results from
simulation models and to be able to assess the
effect of reductions in commuting.
In general, the commuting population of workers
lives in one tract and work in a second tract.
There are over 65,000 Census tracts in the US.
These tracts are very heterogeneous in their
demographic, geographic and economic makeup.
A realistic commuting model component should
correctly account for this heterogeneity.
Currently the MIDAS models mixing
subpopulations of workers without regard to
occupation, income etc., and many sources of
heterogeneity in this population of tracts that are
not fully accounted for, which raises the question
whether these population traits are important in
predicting transmission.
There are alternative sources of data that have
not been incorporated into the model component
of commuting behavior.
With an explicit commuting model component in
place a number of important questions could be
addressed such as: What will be the impact on
disease spread if business air travel is restricted?
What will the impact be of more liberal
telecommuting policies? What will be the impact
of business practices such as replacing face-toface meeting with teleconferencing?
METHODS
We analyzed commuter data from different
sources. All data were aggregated to the Census
tract level. First, univariate outliers were
identified and a random sample of these was
drawn for examination by geography experts.
This was followed by principal components
analysis (PCA) leading to a classification of the
tracts into a number of strata. A random sample
of multivariate outliers (based on PCA scores)
was also examined manually by geography
experts. Regression analysis was carried out on
the remaining data in order to identify
explanations for differences in employment data
between data sets. Regression residual analysis
revealed some additional outliers which were
also examined manually.
RESULTS
There is significant heterogeneity in the tracts.
The sometimes large discrepancy in the
employment data between the data sets we
analyzed, is associated with certain
characteristics of these tracts. When these were
manually examined, it was found out that many
of these were tracts where airports were located,
prisons, large campuses, lakes and parks, heavy
industry, etc.
CONCLUSIONS
In order to make valid inference on the impact of
commuting on the dynamics of a pandemic in the
U.S, and assess the effects on changes in its
patterns, valid data are needed. Using data from
multiple sources, together with manual
examination of statistical outliers can be a good
step in this direction.