The Median Polish procedure constructs an additive model to represent the contents of each cell in a two-way table based on a common value, row effects, column effects, and residuals. It estimates these terms using medians rather than means, making the estimates more resistant to outliers. The document provides an example of applying Median Polish to mortality rate data involving smoking levels and causes of death.
The Median Polish procedure constructs an additive model to represent the contents of each cell in a two-way table based on a common value, row effects, column effects, and residuals. It estimates these terms using medians rather than means, making the estimates more resistant to outliers. The document provides an example of applying Median Polish to mortality rate data involving smoking levels and causes of death.
The Median Polish procedure constructs an additive model to represent the contents of each cell in a two-way table based on a common value, row effects, column effects, and residuals. It estimates these terms using medians rather than means, making the estimates more resistant to outliers. The document provides an example of applying Median Polish to mortality rate data involving smoking levels and causes of death.
The Median Polish procedure constructs an additive model to represent the contents of each cell in a two-way table based on a common value, row effects, column effects, and residuals. It estimates these terms using medians rather than means, making the estimates more resistant to outliers. The document provides an example of applying Median Polish to mortality rate data involving smoking levels and causes of death.
The Median Polish procedure constructs a model for data contained in a two-way table. The model represents the contents of each cell in terms of:
1. a common value 2. a row effect 3. a column effect 4. a residual
Although the model used is similar to that estimated using a two-way analysis of variance, the terms in the model are estimated using medians rather than means. This makes the estimates more resistant to the possible presence of outliers.
The methodology is due to Tukey (1977). An excellent discussion is given in Velleman and Hoaglin (1981).
Sample StatFolio: medianpolish.sgp
Sample data: The file tobacco.sgd contains information on the death rates of men from various causes as a function of how much tobacco they smoked on a daily basis. The data, from Velleman and Hoaglin (1981), are shown below:
Cause None Grams 1_14 Grams 15_24 Grams 25
Lung cancer 0.07 0.47 0.86 1.66 Upper resp. cancer 0.00 0.13 0.09 0.21 Stomach cancer 0.41 0.36 0.10 0.31 Colon cancer 0.44 0.54 0.37 0.74 Prostrate caner 0.55 0.26 0.22 0.34 Other cancer 0.64 0.72 0.76 1.02 TB 0.00 0.16 0.18 0.29 Bronchitis 0.12 0.29 0.39 0.72 Other respitory 0.69 0.55 0.54 0.40 Thrombosis 4.22 4.64 4.60 5.99 Cardiovascular 2.23 2.15 2.47 2.25 Hemorrhage 2.01 1.94 1.86 2.33 Ulcer 0.00 0.14 0.16 0.22 Violence 0.42 0.82 0.45 0.90 Other 1.45 1.81 1.47 1.57
The data represent the number of male deaths per 1000.
2009 by StatPoint Technologies, Inc. Median Polish - 1
STATGRAPHICS – Rev. 7/7/2009
Data Input The data input dialog box specifies the columns containing the data to be analyzed.
Columns: two or more numeric columns containing the data in each column of the table.
Labels: optional labels corresponding to each row of the table.
Select: subset selection.
2009 by StatPoint Technologies, Inc. Median Polish - 2
STATGRAPHICS – Rev. 7/7/2009
Analysis Summary The Analysis Summary summarizes the input data and shows the original table.
Median Polish of Twoway Table
Column variables: None Grams 1_14 Grams 15_24 Grams 25 Number of rows: 15 Number of columns: 4
Original Table Cause None Grams 1_14 Grams 15_24 Grams 25 Row median Lung cancer 0.07 0.47 0.86 1.66 0.665 Upper resp. cancer 0.0 0.13 0.09 0.21 0.11 Stomach cancer 0.41 0.36 0.1 0.31 0.335 Colon cancer 0.44 0.54 0.37 0.74 0.49 Prostrate caner 0.55 0.26 0.22 0.34 0.3 Other cancer 0.64 0.72 0.76 1.02 0.74 TB 0.0 0.16 0.18 0.29 0.17 Bronchitis 0.12 0.29 0.39 0.72 0.34 Other respitory 0.69 0.55 0.54 0.4 0.545 Thrombosis 4.22 4.64 4.6 5.99 4.62 Cardiovascular 2.23 2.15 2.47 2.25 2.24 Hemorrhage 2.01 1.94 1.86 2.33 1.975 Ulcer 0.0 0.14 0.16 0.22 0.15 Violence 0.42 0.82 0.45 0.9 0.635 Other 1.45 1.81 1.47 1.57 1.52 Column median 0.44 0.54 0.45 0.74 0.54
Along the right and bottom of the table are the medians of the cells in each row and column.
Polished Table Performing a median polish on a two-way table results in a model which expresses the data in each cell of the table as:
cellij = common value + row effecti + column effectj + residualij (1)
The Polished Table pane shows the fitted model:
2009 by StatPoint Technologies, Inc. Median Polish - 3
STATGRAPHICS – Rev. 7/7/2009
Polished Table Sweeping 3 times.
Cause None Grams 1_14 Grams 15_24 Grams 25 Row effect
For example, the model for the cell in the upper left corner is:
Cell1,1 = common value + row effect1 + column effect1 + residual1,1 (2)
0.07 = 0.54625 + 0.1175 – 0.09375 – 0.5 (3)
Intuitively, the above equation expresses the rate of lung cancer among none smokers as the sum of a constant plus a small row effect (since lung cancer has a higher rate than average), minus a small column effect (since none smokers have a lower death rate on average), plus a large residual (since none smokers rarely get lung cancer).
The column effects in the above table are of special interest. The large negative effect for none smokers indicates that they have a lower risk than the other categories. The large positive effect for the heaviest smokers indicate that they have an elevated risk. The two medium smoking categories have effects that are similar to each other.
Analysis Options
Number of sweeps: the number of full iterations to use in estimating the model. Since the estimation procedure is iterative, additional sweeps may make minor changes is the estimated row effects, column effects, and common value. 2009 by StatPoint Technologies, Inc. Median Polish - 4 STATGRAPHICS – Rev. 7/7/2009
Residual Plot This plot shows the residuals plotted against the comparison values:
If an additive model adequately describes the data, there should not be a discernible relationship between the residuals and the comparison values. A strong positive correlation usually indicates a need to reexpress the values in the table using some sort of power transformation.
STATGRAPHICS fits a line to the plot of residuals versus comparison values using a resistant fit (one that is not sensitive to outliers). The slope of the fitted line b can be using to suggest a possible power p to raise the table values to, where:
p=1–b (5)
A power close to p = 0.5 would indicate a need to take the square roots of the data. p = 0 corresponds to the need to take logarithms.
In the sample data, the slope is small, suggesting that the additive model is sufficient.
2009 by StatPoint Technologies, Inc. Median Polish - 5
STATGRAPHICS – Rev. 7/7/2009
Save Results The following results can be saved to columns of the datasheet:
1. Row effects – the estimated effect of each row.
2. Column effects – the estimated effect of each column. 3. Common value – the estimated common value. 4. Residuals – the residuals for each cell of the table (the values in the polished table). 5. Comparison values – the comparison values for each cell of the table. 6. Row identifiers – the row identifier corresponding to each residual and comparison value. 7. Column identifiers – the column identifier corresponding to each residual and comparison value.
2009 by StatPoint Technologies, Inc. Median Polish - 6
