Spatial Smoothing


In the spatial analysis of attributes measured for areal units, it is often necessary to transform an extensive variable, such as number of disease cases per census tract, into an intensive variable that takes into account the underlying population at risk. Raw rates, counts divided by population values, are a common standardization in the literature, yet these tend to have unequal reliability due to different population sizes across the spatial units. This problem becomes severe for areas with small population values, since the raw rates for those areas tend to have higher variance.

A variety of spatial smoothing methods have been suggested to address this problem by aggregating the counts and population values for the areas neighboring an observation and using these new measurements for its rate computation. PySAL provides a range of smoothing techniques that exploit different types of moving windows and non-parametric weighting schemes as well as the Empirical Bayesian principle. In addition, PySAL offers several methods for calculating age-standardized rates, since age standardization is critical in estimating rates of some events where the probability of an event occurrence is different across different age groups.

In what follows, we overview the methods for age standardization and spatial smoothing and describe their implementations in PySAL. [1]

Age Standardization in PySAL

Raw rates, counts divided by populations values, are based on an implicit assumption that the risk of an event is constant over all age/sex categories in the population. For many phenomena, however, the risk is not uniform and often highly correlated with age. To take this into account explicitly, the risks for individual age categories can be estimated separately and averaged to produce a representative value for an area.

PySAL supports three approaches to this age standardization: crude, direct, and indirect standardization.

Crude Age Standardization

In this approach, the rate for an area is simply the sum of age-specific rates weighted by the ratios of each age group in the total population.

To obtain the rates based on this approach, we first need to create two variables that correspond to event counts and population values, respectively.

>>> import numpy as np
>>> e = np.array([30, 25, 25, 15, 33, 21, 30, 20])
>>> b = np.array([100, 100, 110, 90, 100, 90, 110, 90])

Each set of numbers should include n by h elements where n and h are the number of areal units and the number of age groups. In the above example there are two regions with 4 age groups. Age groups are identical across regions. The first four elements in b represent the populations of 4 age groups in the first region, and the last four elements the populations of the same age groups in the second region.

To apply the crude age standardization, we need to make the following function call:

>>> from pysal.esda import smoothing as sm
>>> sm.crude_age_standardization(e, b, 2)
array([ 0.2375    ,  0.26666667])

In the function call above, the last argument indicates the number of area units. The outcome in the second line shows that the age-standardized rates for two areas are about 0.24 and 0.27, respectively.

Direct Age Standardization

Direct age standardization is a variation of the crude age standardization. While crude age standardization uses the ratios of each age group in the observed population, direct age standardization weights age-specific rates by the ratios of each age group in a reference population. This reference population, the so-called standard million, is another required argument in the PySAL implementation of direct age standardization:

>>> s = np.array([100, 90, 100, 90, 100, 90, 100, 90])
>>> rate = sm.direct_age_standardization(e, b, s, 2, alpha=0.05)
>>> np.array(rate).round(6)
array([[ 0.23744 ,  0.192049,  0.290485],
       [ 0.266507,  0.217714,  0.323051]])

The outcome of direct age standardization includes a set of standardized rates and their confidence intervals. The confidence intervals can vary according to the value for the last argument, alpha.

Indirect Age Standardization

While direct age standardization effectively addresses the variety in the risks across age groups, its indirect counterpart is better suited to handle the potential imprecision of age-specific rates due to the small population size. This method uses age-specific rates from the standard million instead of the observed population. It then weights the rates by the ratios of each age group in the observed population. To compute the age-specific rates from the standard million, the PySAL implementation of indirect age standardization requires another argument that contains the counts of the events occurred in the standard million.

>>> s_e = np.array([10, 15, 12, 10, 5, 3, 20, 8])
>>> rate = sm.indirect_age_standardization(e, b, s_e, s, 2, alpha=0.05)
>>> np.array(rate).round(6)
array([[ 0.208055,  0.170156,  0.254395],
       [ 0.298892,  0.246631,  0.362228]])

The outcome of indirect age standardization is the same as that of its direct counterpart.

Spatial Smoothing in PySAL

Mean and Median Based Smoothing

A simple approach to rate smoothing is to find a local average or median from the rates of each observation and its neighbors. The first method adopting this approach is the so-called locally weighted averages or disk smoother. In this method a rate for each observation is replaced by an average of rates for its neighbors. A spatial weights object is used to specify the neighborhood relationships among observations. To obtain locally weighted averages of the homicide rates in the counties surrounding St. Louis during 1979-84, we first read the corresponding data table and extract data values for the homicide counts (the 11th column) and total population (the 13th column):

>>> import pysal
>>> stl ='../pysal/examples/stl_hom.csv', 'r')
>>> e, b = np.array(stl[:,10]), np.array(stl[:,13])

We then read the spatial weights file defining neighborhood relationships among the counties and ensure that the order of observations in the weights object is the same as that in the data table.

>>> w ='../pysal/examples/', 'r').read()
>>> if not w.id_order_set: w.id_order = range(1,len(stl) + 1)

Now we calculate locally weighted averages of the homicide rates.

>>> rate = sm.Disk_Smoother(e, b, w)
>>> rate.r
array([  4.56502262e-05,   3.44027685e-05,   3.38280487e-05,
         4.78530468e-05,   3.12278573e-05,   2.22596997e-05,
         5.29577710e-05,   5.51034691e-05,   4.65160450e-05,
         5.32513363e-05,   3.86199097e-05,   1.92952422e-05])

A variation of locally weighted averages is to use median instead of mean. In other words, the rate for an observation can be replaced by the median of the rates of its neighbors. This method is called locally weighted median and can be applied in the following way:

>>> rate = sm.Spatial_Median_Rate(e, b, w)
>>> rate.r
array([  3.96047383e-05,   3.55386859e-05,   3.28308921e-05,
         4.30731238e-05,   3.12453969e-05,   1.97300409e-05,
         6.10668237e-05,   5.86355507e-05,   3.67396656e-05,
         4.82535850e-05,   5.51831429e-05,   2.99877050e-05])

In this method the procedure to find local medians can be iterated until no further change occurs. The resulting local medians are called iteratively resmoothed medians.

>>> rate = sm.Spatial_Median_Rate(e, b, w, iteration=10)
>>> rate.r
array([  3.10194715e-05,   2.98419439e-05,   3.10194715e-05,
         3.10159267e-05,   2.99214885e-05,   2.80530524e-05,
         3.81364519e-05,   4.72176972e-05,   3.75320135e-05,
         3.76863269e-05,   4.72176972e-05,   3.75320135e-05])

The pure local medians can also be replaced by a weighted median. To obtain weighted medians, we need to create an array of weights. For example, we can use the total population of the counties as auxiliary weights:

>>> rate = sm.Spatial_Median_Rate(e, b, w, aw=b)
>>> rate.r
array([  5.77412020e-05,   4.46449551e-05,   5.77412020e-05,
         5.77412020e-05,   4.46449551e-05,   3.61363528e-05,
         5.49703305e-05,   5.86355507e-05,   3.67396656e-05,
         3.67396656e-05,   4.72176972e-05,   2.99877050e-05])

When obtaining locally weighted medians, we can consider only a specific subset of neighbors rather than all of them. A representative method following this approach is the headbanging smoother. In this method all areal units are represented by their geometric centroids. Among the neighbors of each observation, only near collinear points are considered for median search. Then, triples of points are selected from the near collinear points, and local medians are computed from the triples’ rates. [2] We apply this headbanging smoother to the rates of the deaths from Sudden Infant Death Syndrome (SIDS) for North Carolina counties during 1974-78. We first need to read the source data and extract the event counts (the 9th column) and population values (the 9th column). In this example the population values correspond to the numbers of live births during 1974-78.

>>> sids_db ='../pysal/examples/sids2.dbf', 'r')
>>> e, b = np.array(sids_db[:,9]), np.array(sids_db[:,8])

Now we need to find triples for each observation. To support the search of triples, PySAL provides a class called Headbanging_Triples. This class requires an array of point observations, a spatial weights object, and the number of triples as its arguments:

>>> from pysal import knnW
>>> sids ='../pysal/examples/sids2.shp', 'r')
>>> sids_d = np.array([i.centroid for i in sids])
>>> sids_w = knnW(sids_d,k=5)
>>> if not sids_w.id_order_set: sids_w.id_order = sids_w.id_order
>>> triples = sm.Headbanging_Triples(sids_d,sids_w,k=5)

The second line in the above example shows how to extract centroids of polygons. In this example we define 5 neighbors for each observation by using nearest neighbors criteria. In the last line we define the maximum number of triples to be found as 5.

Now we use the triples to compute the headbanging median rates:

>>> rate = sm.Headbanging_Median_Rate(e,b,triples)
>>> rate.r
array([ 0.00075586,  0.        ,  0.0008285 ,  0.0018315 ,  0.00498891,
        0.00482094,  0.00133156,  0.0018315 ,  0.00413223,  0.00142116,
        0.00221541,  0.00354767,  0.00259903,  0.00392952,  0.00207125,
        0.00392952,  0.00229253,  0.00392952,  0.00229253,  0.00229253])

As in the locally weighted medians, we can use a set of auxiliary weights and resmooth the medians iteratively.

Non-parametric Smoothing

Non-parametric smoothing methods compute rates without making any assumptions of distributional properties of rate estimates. A representative method in this approach is spatial filtering. PySAL provides the most simplistic form of spatial filtering where a user-specified grid is imposed on the data set and a moving window withi a fixed or adaptive radius visits each vertex of the grid to compute the rate at the vertex. Using the previous SIDS example, we can use Spatial_Filtering class:

>>> bbox = [sids.bbox[:2], sids.bbox[2:]]
>>> rate = sm.Spatial_Filtering(bbox, sids_d, e, b, 10, 10, r=1.5)
>>> rate.r
array([ 0.00152555,  0.00079271,  0.00161253,  0.00161253,  0.00139513,
        0.00139513,  0.00139513,  0.00139513,  0.00139513,  0.00156348,
        0.00240216,  0.00237389,  0.00240641,  0.00242211,  0.0024854 ,
        0.00255477,  0.00266573,  0.00288918,  0.0028991 ,  0.00293492])

The first and second arguments of the Spatial_Filtering class are a minimum bounding box containing the observations and a set of centroids representing the observations. Be careful that the bounding box is NOT the bounding box of the centroids. The fifth and sixth arguments are to specify the numbers of grid cells along x and y axes. The last argument, r, is to define the radius of the moving window. When this parameter is set, a fixed radius is applied to all grid vertices. To make the size of moving window variable, we can specify the minimum number of population in the moving window without specifying r:

>>> rate = sm.Spatial_Filtering(bbox, sids_d, e, b, 10, 10, pop=10000)
>>> rate.r
array([ 0.00157398,  0.00157398,  0.00157398,  0.00157398,  0.00166885,
        0.00166885,  0.00166885,  0.00166885,  0.00166885,  0.00166885,
        0.00202977,  0.00215322,  0.00207378,  0.00207378,  0.00217173,
        0.00232408,  0.00222717,  0.00245399,  0.00267857,  0.00267857])

The spatial rate smoother is another non-parametric smoothing method that PySAL supports. This smoother is very similar to the locally weighted averages. In this method, however, the weighted sum is applied to event counts and population values separately. The resulting weighted sum of event counts is then divided by the counterpart of population values. To obtain neighbor information, we need to use a spatial weights matrix as before.

>>> rate = sm.Spatial_Rate(e, b, sids_w)
>>> rate.r
array([ 0.00114976,  0.00104622,  0.00110001,  0.00153257,  0.00399662,
        0.00361428,  0.00146807,  0.00238521,  0.00288871,  0.00145228,
        0.00240839,  0.00376101,  0.00244941,  0.0028813 ,  0.00240839,
        0.00261705,  0.00226554,  0.0031575 ,  0.00254536,  0.0029003 ])

Another variation of spatial rate smoother is kernel smoother. PySAL supports kernel smoothing by using a kernel spatial weights instance in place of a general spatial weights object.

>>> from pysal import Kernel
>>> kw = Kernel(sids_d)
>>> if not kw.id_order_set: kw.id_order = range(0,len(sids_d))
>>> rate = sm.Kernel_Smoother(e, b, kw)
>>> rate.r
array([ 0.0009831 ,  0.00104298,  0.00137113,  0.00166406,  0.00556741,
        0.00442273,  0.00158202,  0.00243354,  0.00282158,  0.00099243,
        0.00221017,  0.00328485,  0.00257988,  0.00370461,  0.0020566 ,
        0.00378135,  0.00240358,  0.00432019,  0.00227857,  0.00251648])

Age-adjusted rate smoother is another non-parametric smoother that PySAL provides. This smoother applies direct age standardization while computing spatial rates. To illustrate the age-adjusted rate smoother, we create a new set of event counts and population values as well as a new kernel weights object.

>>> e = np.array([10, 8, 1, 4, 3, 5, 4, 3, 2, 1, 5, 3])
>>> b = np.array([100, 90, 15, 30, 25, 20, 30, 20, 80, 80, 90, 60])
>>> s = np.array([98, 88, 15, 29, 20, 23, 33, 25, 76, 80, 89, 66])
>>> points=[(10, 10), (20, 10), (40, 10), (15, 20), (30, 20), (30, 30)]
>>> kw=Kernel(points)
>>> if not kw.id_order_set: kw.id_order = range(0,len(points))

In the above example we created 6 observations each of which has two age groups. To apply age-adjusted rate smoothing, we use the Age_Adjusted_Smoother class as follows:

>>> rate = sm.Age_Adjusted_Smoother(e, b, kw, s)
>>> rate.r
array([ 0.10519625,  0.08494318,  0.06440072,  0.06898604,  0.06952076,

Empirical Bayes Smoothers

The last group of smoothing methods that PySAL supports is based upon the Bayesian principle. These methods adjust a raw rate by taking into account information in the other raw rates. As a reference PySAL provides a method for a-spatial Empirical Bayes smoothing:

>>> e, b = sm.sum_by_n(e, np.ones(12), 6), sm.sum_by_n(b, np.ones(12), 6)
>>> rate = sm.Empirical_Bayes(e, b)
>>> rate.r
array([ 0.09080775,  0.09252352,  0.12332267,  0.10753624,  0.03301368,

In the first line of the above example we aggregate the event counts and population values by observation. Next we applied the Empirical_Bayes class to the aggregated counts and population values.

A spatial Empirical Bayes smoother is also implemented in PySAL. This method requires an additional argument, i.e., a spatial weights object. We continue to reuse the kernel spatial weights object we built before.

>>> rate = sm.Spatial_Empirical_Bayes(e, b, kw)
>>> rate.r
array([ 0.10105263,  0.10165261,  0.16104362,  0.11642038,  0.0226908 ,

Excess Risk

Besides a variety of spatial smoothing methods, PySAL provides a class for estimating excess risk from event counts and population values. Excess risks are the ratios of observed event counts over expected event counts. An example for the class usage is as follows:

>>> risk = sm.Excess_Risk(e, b)
>>> risk.r
array([ 1.23737916,  1.45124717,  2.32199546,  1.82857143,  0.24489796,

Further Information

For further details see the Smoothing API.


[1]Although this tutorial provides an introduction to the PySAL implementations for spatial smoothing, it is not exhaustive. Complete documentation for the implementations can be found by accessing the help from within a Python interpreter.
[2]For the details of triple selection and headbanging smoothing please refer to Anselin, L., Lozano, N., and Koschinsky, J. (2006). “Rate Transformations and Smoothing”. GeoDa Center Research Report.