Поделиться через


Crosstabs using RevoScaleR

Important

This content is being retired and may not be updated in the future. The support for Machine Learning Server will end on July 1, 2022. For more information, see What's happening to Machine Learning Server?

Crosstabs, also known as contingency tables or crosstabulations, are a convenient way to summarize cross-classified categorical data—that is, data that can be tabulated according to multiple levels of two or more factors. If only two factors are involved, the table is sometimes called a two-way table. If three factors are involved, the table is sometimes called a three-way table.

For large data sets, cross-tabulations of binned numeric data, that is, data that has been converted to a factor where the levels represent ranges of values, can be a fast way to get insight into the relationships among variables. In RevoScaleR, the rxCube function is the primary tool to create contingency tables.

For example, the built-in data set UCBAdmissions includes information on admissions by gender to various departments at the University of California at Berkeley. We can look at the contingency table as follows:

UCBADF <- as.data.frame(UCBAdmissions)
z <- rxCube(Freq ~ Gender:Admit, data = UCBADF)

(Because cross-tabulations are explicitly about exploring interactions between variables, multiple predictors must always be specified using the interaction operator ":", and not the terms operator "+".)

Typing z yields the following output:

Call:
rxCube(formula = Freq ~ Gender:Admit, data = UCBADF)

Cube Results for: Freq ~ Gender:Admit
Data: UCBADF
Dependent variable(s): Freq
Number of valid observations: 24
Number of missing observations: 0 
Statistic: Freq means 

  Gender    Admit      Freq Counts
1   Male Admitted 199.66667      6
2 Female Admitted  92.83333      6
3   Male Rejected 248.83333      6
4 Female Rejected 213.00000      6

This data set is widely used in statistics texts because it illustrates Simpson’s paradox, which is that in some cases a comparison that holds true in a number of groups is reversed when those groups are aggregated to form a single group. From the preceding table, in which admissions data is aggregated across all departments, it would appear that males are admitted at a higher rate than women. However, if we look at the more granular analysis by department, we find that in four of the six departments, women are admitted at a higher rate than men:

z2 <- rxCube(Freq ~ Gender:Admit:Dept, data = UCBADF)
z2

This yields the following output:

Call:
rxCube(formula = Freq ~ Gender:Admit:Dept, data = UCBADF)

Cube Results for: Freq ~ Gender:Admit:Dept
Data: UCBADF
Dependent variable(s): Freq
Number of valid observations: 24
Number of missing observations: 0 
Statistic: Freq means 

   Gender    Admit Dept Freq Counts
1    Male Admitted    A  512      1
2  Female Admitted    A   89      1
3    Male Rejected    A  313      1
4  Female Rejected    A   19      1
5    Male Admitted    B  353      1
6  Female Admitted    B   17      1
7    Male Rejected    B  207      1
8  Female Rejected    B    8      1
9    Male Admitted    C  120      1
10 Female Admitted    C  202      1
11   Male Rejected    C  205      1
12 Female Rejected    C  391      1
13   Male Admitted    D  138      1
14 Female Admitted    D  131      1
15   Male Rejected    D  279      1
16 Female Rejected    D  244      1
17   Male Admitted    E   53      1
18 Female Admitted    E   94      1
19   Male Rejected    E  138      1
20 Female Rejected    E  299      1
21   Male Admitted    F   22      1
22 Female Admitted    F   24      1
23   Male Rejected    F  351      1
24 Female Rejected    F  317      1

Letting the Data Speak Example 1: Analyzing U.S. 2000 Census Data

The CensusWorkers.xdf data set contains a subset of the U.S. 2000 5% Census for individuals aged 20 to 65 who worked at least 20 weeks during the year from three states. Let’s examine the relationship between wage income (represented in the data set by the variable incwage) and age.

A useful way to observe the relationship between numeric variables is to bin the predictor variable (in our case, age), and then plot the mean of the response for each bin. The simplest way to bin age is to use the F() wrapper within our initial formula; it creates a separate bin for each distinct value of age. (More precisely, it creates a bin of length one from the low value of age to the high value of age—if some ages are missing in the original data set, bins are created for them anyway.)

We create our original model as follows:

#  Letting the data speak: Example 1

readPath <- rxGetOption("sampleDataDir")
censusWorkers <- file.path(readPath, "CensusWorkers.xdf")
censusWorkersCube <- rxCube(incwage ~ F(age), data=censusWorkers)

We first look at the results in tabular form by typing the returned object name, censusWorkersCube, which yields the following output:

Call:
rxCube(formula = incwage ~ F(age), data = censusWorkers)

Cube Results for: incwage ~ F(age)
File name:
    C:\Program Files\Microsoft\MRO-for-RRE\8.0\R-3.2.2\library\RevoScaleR\SampleData\CensusWorkers.xdf
Dependent variable(s): incwage
Number of valid observations: 351121
Number of missing observations: 0 
Statistic: incwage means 

   F_age incwage  Counts
1     20 12669.94   6500
2     21 14114.23   6479
3     22 15982.00   6676
4     23 18503.92   6884
5     24 20672.06   6931
6     25 23856.25   7273
7     26 25938.17   7116
8     27 26902.97   7584
9     28 28531.59   8184
10    29 30153.10   8889
11    30 30691.10   9055
12    31 31647.06   8670
13    32 33459.31   8459
14    33 34208.33   8574
15    34 34364.06   9058
16    35 35739.92   9743
17    36 36945.24   9888
18    37 36970.63   9860
19    38 37331.39  10211
20    39 38899.67  10378
21    40 38279.34  10756
22    41 39678.52  10503
23    42 40748.10  10511
24    43 39910.90  10296
25    44 40524.19  10122
26    45 41450.27  10074
27    46 40521.07   9703
28    47 41371.40   9527
29    48 42061.04   9093
30    49 41618.36   8776
31    50 42789.36   8868
32    51 41912.11   8506
33    52 43169.23   8690
34    53 41864.13   8362
35    54 42920.45   6275
36    55 42939.81   6171
37    56 41157.10   5915
38    57 40984.69   5881
39    58 40553.04   5047
40    59 38738.45   4512
41    60 37200.02   3775
42    61 35978.18   3704
43    62 35000.53   3206
44    63 34098.00   2563
45    64 32964.57   2248
46    65 31698.98   1625

As we wanted, the table contains average values of incwage for each level of age. If we want to create a plot of the results, we can use the rxResultsDF function to conveniently convert the output into a data frame. The F_age factor variable will automatically be converted back to an integer age variable. Then we can plot the data using rxLinePlot:

censusWorkersCubeDF <- rxResultsDF(censusWorkersCube)
rxLinePlot(incwage ~ age, data=censusWorkersCubeDF, 
    title="Relationship of Income and Age")

The resulting plot shows clearly the relationship of income on age:

Chart showing relationship of income and age

Transforming Data

Because crosstabs require categorical data for the predictors, you have to do some work to crosstabulate continuous data. In the previous section, we saw that the F() wrapper can do a transformation within a formula. The transforms argument to rxCrossTabs can be used to give you greater control over such transformations.

For example, the kyphosis data from the rpart package consists of one categorical variable, Kyphosis, and three continuous variables Age, Number, and Start. The Start variable indicates the topmost vertebra involved in a certain type of spinal surgery, and has a range of 1 to 18. Since there are 7 cervical vertebrae and 12 thoracic vertebrae, we can specify a transform that classifies the start variable as either cervical or thoracic as follows:

#  Transforming Data

cut(Start, breaks=c(0, 7.5, 19.5), labels=c("cervical", "thoracic"))

Similarly, we can create a factorized Age variable as follows (in the original data, age is given in months; with our set of breaks, we cut the data into ranges of years):

cut(Age, breaks=c(0, 12, 60, 119, 180, 220), labels=c("<1", "1-4", 
    "5-9", "10-15", ">15"))

We can now crosstabulate the data using the preceding transforms and it is instructive to start by looking at the three two-way tables formed by tabulating Kyphosis with the three predictor variables:

library(rpart)
rxCube(~ Kyphosis:Age, data = kyphosis, 
    transforms=list(Age = cut(Age, breaks=c(0, 12, 60, 119, 
    180, 220), labels=c("<1", "1-4", "5-9", "10-15", ">15"))))
Call:
rxCube(formula = ~Kyphosis:Age, data = kyphosis, transforms = list(Age = cut(Age, 
    breaks = c(0, 12, 60, 119, 180, 220), labels = c("<1", "1-4", 
	"5-9", "10-15", ">15"))))

Cube Results for: ~Kyphosis:Age
Data: kyphosis
Number of valid observations: 81
Number of missing observations: 0 

   Kyphosis   Age Counts
1    absent    <1     13
2   present    <1      0
3    absent   1-4     13
4   present   1-4      4
5    absent   5-9     17
6   present   5-9      6
7    absent 10-15     19
8   present 10-15      7
9    absent   >15      2
10  present   >15      0
rxCube(~ Kyphosis:F(Number), data = kyphosis)
Call:
rxCube(formula = ~Kyphosis:F(Number), data = kyphosis)

Cube Results for: ~Kyphosis:F(Number)
Data: kyphosis
Number of valid observations: 81
Number of missing observations: 0 

   Kyphosis F_Number Counts
1    absent        2     12
2   present        2      0
3    absent        3     19
4   present        3      4
5    absent        4     16
6   present        4      2
7    absent        5     12
8   present        5      5
9    absent        6      2
10  present        6      2
11   absent        7      2
12  present        7      3
13   absent        8      0
14  present        8      0
15   absent        9      1
16  present        9      0
17   absent       10      0
18  present       10      1

rxCube(~ Kyphosis:Start, data = kyphosis, 
    transforms=list(Start = cut(Start, breaks=c(0, 7.5, 19.5), 
    labels=c("cervical", "thoracic"))))
Call:
rxCube(formula = ~Kyphosis:Start, data = kyphosis, transforms = list(Start = cut(Start, 
    breaks = c(0, 7.5, 19.5), labels = c("cervical", "thoracic"))))

Cube Results for: ~Kyphosis:Start
Data: kyphosis
Number of valid observations: 81
Number of missing observations: 0 

  Kyphosis    Start Counts
1   absent cervical      8
2  present cervical      9
3   absent thoracic     56
4  present thoracic      8

From these, we see that the probability of the post-operative complication Kyphosis seems to be greater if the Start is a cervical vertebra and as more vertebrae are involved in the surgery. Similarly, it appears that the dependence on age is non-linear: it first increases with age, peaks in the range 5-9, and then decreases again.

Cross-Tabulation with rxCrossTabs

The rxCrossTabs function is an alternative to the rxCube function, which performs the same calculations, but displays its results in format similar to the standard R xtabs function. For some purposes, this format can be more informative than the matrix-like display of rxCube, and in some situations can be more compact as well.

As an example, consider again the admission data example:

#  Cross-Tabulation with rxCrossTabs

z3 <- rxCrossTabs(Freq ~ Gender:Admit:Dept, data = UCBADF)
z3
Call:
rxCrossTabs(formula = Freq ~ Gender:Admit:Dept, data = UCBADF)

Cross Tabulation Results for: Freq ~ Gender:Admit:Dept
Data: UCBADF
Dependent variable(s): Freq
Number of valid observations: 24
Number of missing observations: 0 
Statistic: sums 

Freq, Dept = A (sums):
	Admit
Gender   Admitted Rejected
  Male        512      313
  Female       89       19

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = B (sums):
	Admit
Gender   Admitted Rejected
  Male        353      207
  Female       17        8

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = C (sums):
	Admit
Gender   Admitted Rejected
  Male        120      205
  Female      202      391

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = D (sums):
	Admit
Gender   Admitted Rejected
  Male        138      279
  Female      131      244

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = E (sums):
	Admit
Gender   Admitted Rejected
  Male         53      138
  Female       94      299

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = F (sums):
	Admit
Gender   Admitted Rejected
  Male         22      351
  Female       24      317

You can see the row, column, and total percentages by calling the summary function on the rxCrossTabs object:

summary(z3)
Call:
rxCrossTabs(formula = Freq ~ Gender:Admit:Dept, data = UCBADF)

Cross Tabulation Results for: Freq ~ Gender:Admit:Dept
Data: UCBADF
Dependent variable(s): Freq
Number of valid observations: 24
Number of missing observations: 0 
Statistic: sums 

Freq, Dept = A (sums):
	      Admitted   Rejected Row Total
Male        512.000000 313.000000 825.00000
   Row%      62.060606  37.939394          
   Col%      85.191348  94.277108          
   Tot%      54.876742  33.547696  88.42444
Female       89.000000  19.000000 108.00000
   Row%      82.407407  17.592593          
   Col%      14.808652   5.722892          
   Tot%       9.539121   2.036442  11.57556
Col Total   601.000000 332.000000          
Grand Total 933.000000                     

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = B (sums):
	      Admitted   Rejected  Row Total
Male        353.000000 207.000000 560.000000
   Row%      63.035714  36.964286           
   Col%      95.405405  96.279070           
   Tot%      60.341880  35.384615  95.726496
Female       17.000000   8.000000  25.000000
   Row%      68.000000  32.000000           
   Col%       4.594595   3.720930           
   Tot%       2.905983   1.367521   4.273504
Col Total   370.000000 215.000000           
Grand Total 585.000000                      

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = C (sums):
	     Admitted  Rejected Row Total
Male        120.00000 205.00000 325.00000
   Row%      36.92308  63.07692          
   Col%      37.26708  34.39597          
   Tot%      13.07190  22.33115  35.40305
Female      202.00000 391.00000 593.00000
   Row%      34.06408  65.93592          
   Col%      62.73292  65.60403          
   Tot%      22.00436  42.59259  64.59695
Col Total   322.00000 596.00000          
Grand Total 918.00000                    

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = D (sums):
	     Admitted  Rejected Row Total
Male        138.00000 279.00000 417.00000
   Row%      33.09353  66.90647          
   Col%      51.30112  53.34608          
   Tot%      17.42424  35.22727  52.65152
Female      131.00000 244.00000 375.00000
   Row%      34.93333  65.06667          
   Col%      48.69888  46.65392          
   Tot%      16.54040  30.80808  47.34848
Col Total   269.00000 523.00000          
Grand Total 792.00000                    

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = E (sums):
	      Admitted  Rejected Row Total
Male         53.000000 138.00000 191.00000
   Row%      27.748691  72.25131          
   Col%      36.054422  31.57895          
   Tot%       9.075342  23.63014  32.70548
Female       94.000000 299.00000 393.00000
   Row%      23.918575  76.08142          
   Col%      63.945578  68.42105          
   Tot%      16.095890  51.19863  67.29452
Col Total   147.000000 437.00000          
Grand Total 584.000000                    

::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::

Freq, Dept = F (sums):
	      Admitted  Rejected Row Total
Male         22.000000 351.00000  373.0000
   Row%       5.898123  94.10188          
   Col%      47.826087  52.54491          
   Tot%       3.081232  49.15966   52.2409
Female       24.000000 317.00000  341.0000
   Row%       7.038123  92.96188          
   Col%      52.173913  47.45509          
   Tot%       3.361345  44.39776   47.7591
Col Total    46.000000 668.00000          
Grand Total 714.000000                    

You can see, for example, that in Department A, 62 percent of male applicants are admitted, but 82 percent of female applicants are admitted, and in Department B, 63 percent of male applicants are admitted, while 68 percent of female applicants are admitted.

A Large Data Example

The power of rxCrossTabs is most evident when you need to tabulate a data set that won’t fit into memory. For example, in the large airline data set AirOnTime87to12.xdf, you can obtain the mean arrival delay by carrier and day of week as follows (if you have downloaded the data set, modify the first line as follows to reflect your local path):

The blocksPerRead argument is ignored if run locally using R Client. Learn more...

#  A Large Data Example

bigDataDir <- "C:/MRS/Data"
bigAirData <- file.path(bigDataDir, "AirOnTime87to12/AirOnTime87to12.xdf")
arrDelayXT <- rxCrossTabs(ArrDelay ~ UniqueCarrier:DayOfWeek, 
    data = bigAirData, blocksPerRead = 30)
print(arrDelayXT)

Gives the following output:

Call:
rxCrossTabs(formula = ArrDelay ~ UniqueCarrier:DayOfWeek, data = bigAirData, 
    blocksPerRead = 30)

Cross Tabulation Results for: ArrDelay ~ UniqueCarrier:DayOfWeek
File name: C:\MRS\Data\AirOnTime87to12\AirOnTime87to12.xdf
Dependent variable(s): ArrDelay
Number of valid observations: 145576737
Number of missing observations: 3042918 
Statistic: sums 

ArrDelay (sums):
	     DayOfWeek
UniqueCarrier      Mon     Tues      Wed     Thur      Fri      Sat      Sun
       AA     15956852 13367087 16498840 20554660 20714146  9359729 14577582
       US     11797366 11688903 14065606 17379113 19541862  5865427 10264583
       AS      3144578  2677858  3182356  3980209  4415144  2433581  3039129
       CO      8464507  7966834  9537366 11901028 11749616  3553719  6562487
       DL     18146092 15962559 19474389 24077435 24933864 10483280 16414060
       EA       782103   832332   796811  1152825  1405399   638911   670924
       HP      3577460  3170343  3700890  4734543  5015896  2864314  3985741
       NW      7750970  7818040  9256994 11199718 10294116  3726129  6504924
       PA (1)   191137   235924   225260   290844   345238   174284   229677
       PI      1164688  1391526  1456173  1515403  1568266   939820   986642
       PS        88144   111282   122520   133567   173422    44362    88891
       TW      3356944  3459185  4060151  5027427  5267750  1669048  2377671
       UA     15941096 14731587 17801128 21060697 20920843  9174567 13688577
       WN     12438738  8978339 12215989 21556781 26787623  4972506 15973176
       ML (1)    20735    50927    55881    75030    62855    44549    18173
       KH        20744   -26425   -24265    30078    98529    33595    43026
       MQ      7065052  5152746  5893882  7136735  8087443  2947023  5540099
       B6      1886261  1340744  1736450  2373151  2930423  1012698  1969546
       DH       795614   527649   708129   801968   986930   227907   504644
       EV      5733212  3684210  4005374  5262924  5874647  1753361  4418290
       FL      2666677  1694294  1810548  2928247  3068538   819827  2188420
       OO      4717107  3106319  3438056  4725854  5481441  2797745  4764041
       XE      4870453  3904752  4532069  5349375  5315818  1636826  3531446
       TZ       228508   147963   197371   224693   275340    39722   148940
       HA       -72468   -92714   -66578     4840   153830    -2082   -22196
       OH      2276399  1567510  1830571  2336032  2702519   922531  1659708
       F9       551932   484426   566122   858027   729273   337695   526887
       YV      1959906  1419073  1463954  1930992  2152270  1270104  1830749
       9E       787776   579608   590038   709161   869358   304151   586378
       VX        10208    37079    12956    42661    73457     2943    39987

Using Sparse Cubes

An additional tool that may be useful when using rxCube and rxCrossTabs with large data is the useSparseCube parameter.  Compiling cross-tabulations of categorical data can sometimes result in a large number of cells with zero counts, yielding at its core a “sparse matrix”.  In the usual case, memory is allocated for every cell in the cube, but large cubes may overwhelm memory resources. If we instead allocate space only for cells with positive counts, such operations may often proceed successfully.

As an example, let’s look at the airline data again and construct a case where the cross-tabulation yields many zero entries.  As the overwhelming number of flights in the data set were not canceled, by appending the Cancelled predictor in the formula, we would expect a large number of categorical predictor combinations to have zero observations.  Because the Cancelled predictor is a logical rather than a factor variable, we need to use the F(.) function to convert it.

bigDataDir <- "C:/MRS/Data"
bigAirData <- file.path(bigDataDir, "AirOnTime87to12/AirOnTime87to12.xdf")

arrDelaySparse <- rxCube(ArrDelay ~ UniqueCarrier:DayOfWeek:F(Cancelled),
		 data = bigAirData, blocksPerRead = 30, useSparseCube = TRUE)
print(arrDelaySparse)

This gives the following output. We get 210 rows with F_Cancelled = 0. By default, if useSparseCube=TRUE, rows with zero counts are removed from the result.

Call:
rxCube(formula = ArrDelay ~ UniqueCarrier:DayOfWeek:F(Cancelled), 
   data = bigAirData, useSparseCube = TRUE, blocksPerRead = 30)

Cube Results for: ArrDelay ~ UniqueCarrier:DayOfWeek:F(Cancelled)
File name: C:/data/AirOnTime87to12/AirOnTime87to12.xdf
Dependent variable(s): ArrDelay
Number of valid observations: 145576737
Number of missing observations: 3042918 
Statistic: ArrDelay means 
   UniqueCarrier DayOfWeek F_Cancelled ArrDelay    Counts 
1   AA            Mon       0            6.54609466 2437614
2   US            Mon       0            5.20151371 2268064
3   AS            Mon       0            6.37231471  493475
4   CO            Mon       0            6.49381077 1303473
5   DL            Mon       0            6.63422763 2735223
6   EA            Mon       0            6.13500730  127482
7   HP            Mon       0            6.85447467  521916
8   NW            Mon       0            5.09910096 1520066
9   PA (1)        Mon       0            4.24550765   45021
10  PI            Mon       0            9.32556128  124892
11  PS            Mon       0            7.11355016   12391
12  TW            Mon       0            6.21971889  539726
13  UA            Mon       0            7.51524443 2121168
14  WN            Mon       0            4.11051602 3026077
15  ML (1)        Mon       0            2.05724774   10079
… [rows omitted] …
201 FL            Sun       0            6.91461396  316492
202 OO            Sun       0            6.30954516  755053
203 XE            Sun       0            7.72464706  457166
204 TZ            Sun       0            5.19715263   28658
205 HA            Sun       0           -0.28030915   79184
206 OH            Sun       0            7.12036827  233093
207 F9            Sun       0            5.61844996   93778
208 YV            Sun       0            8.35988986  218992
209 9E            Sun       0            4.18506623  140112
210 VX            Sun       0            5.06036446    7902

While this particular example will likely run successfully to completion even on a minimally equipped modern computer without setting the useSparseCube flag to TRUE, it illustrates how one can quickly start to see the number of zero entries accumulate in an rxCube computation.  With larger data sets and a larger number of categorical variable combinations, however, this setting may allow computations of cubes that would not otherwise fit in memory.

For the rxCrossTabs function, the useSparseCube option works exactly the same internally. However, because rxCrossTabs always returns a table, it may require more memory to format its result than rxCube. If you have an extremely large contingency table, we recommend rxCube with useSparseCube=TRUE for the greatest chance of completing the computation.  The useSparseCube flag may also be used with rxSummary.

Tests of Independence on Cross-Tabulated Data

One common use of contingency tables is to test whether the tabulated variables are independent. RevoScaleR includes several tests of independence, all of which expect data in the standard R xtabs format. You can get data in this format from the rxCrossTabs function by using the argument returnXtabs=TRUE:

#  Tests of Independence on Cross-Tabulated Data

bigDataDir <- "C:/MRS/Data"
bigAirData <- file.path(bigDataDir, "AirOnTime87to12/AirOnTime87to12.xdf")
arrDelayXTab <- rxCrossTabs(ArrDel15~ UniqueCarrier:DayOfWeek, 
    data = bigAirData, blocksPerRead = 30, returnXtabs=TRUE)

The blocksPerRead argument is ignored if run locally using R Client. Learn more...

You can then use this as input to any of the following functions:

  • rxChiSquaredTest: performs Pearson’s chi-squared test of independence.
  • rxFisherTest: performs Fisher’s exact test of independence.
  • rxKendallCor: performs a Kendall tau test of independence. There are three flavors of test, a, b, and c; by default, the b flavor, which accounts for ties, is used.

(In fact, regular rxCrossTabs or rxCube output can be used as input to these functions, but they are converted to xtabs format first, so it is somewhat more efficient to have rxCrossTabs return the xtabs format directly.)

Here we use the arrDelayXTab data created preceding and perform a Pearson’s chi-squared test of independence on it:

rxChiSquaredTest(arrDelayXTab)

Gives the following output:

Chi-squared test of independence between UniqueCarrier and DayOfWeek 
 X-squared  df p-value
  105645.8 174       0

For large contingency tables such as this one, the chi-squared test is the tool of choice. For smaller tables, particularly those with cells with expected counts fewer than five, Fisher’s exact test is useful. On a large table, however, Fisher’s exact test may not be an option. For example, if we try it on our airline table, it returns an error:

rxFisherTest(arrDelayXTab)

 Error in FUN(tbl[, , i], ...) : FEXACT error 40.
 Out of workspace. 

To show the Fisher test, we return to the admissions data from the beginning of the article. This time we use rxCrossTabs to return an xtabs object:

UCBADF <- as.data.frame(UCBAdmissions)
admissCTabs <- rxCrossTabs(Freq ~ Gender:Admit, data = UCBADF, 
    returnXtabs=TRUE)

We then call rxFisherTest on the resulting table:

rxFisherTest(admissCTabs)
Fisher's Exact Test for Count Data 
 estimate 1 95% CI Lower 95% CI Upper      p-value
   1.840856     1.621356     2.091246 4.835903e-22
   HA: two.sided 
   H0:  odds ratio = 1

The chi-squared test works equally well on this example:

rxChiSquaredTest(admissCTabs)	
Chi-squared test of independence between Gender and Admit 
 X-squared df      p-value
   91.6096  1 1.055797e-21

In both cases, we are given indisputable evidence of the independence of our two predictor factors. For this example, we could have as easily used the standard R functions chisq.test and fisher.test. The RevoScaleR enhancements, however, permit rxChisSquaredTest and rxFisherTest to work on xtabs objects with multiple tables. For example, if we expand our examination of the admissions data to include the department info, we obtain a multi-way contingency table:

admissCTabs2 <- rxCrossTabs(Freq ~ Gender:Admit:Dept, data = UCBADF,
    returnXtabs=TRUE)

The chi-squared and Fisher’s exact test results are shown as follows; notice that they provide a test of independence between Gender and Admit for each level of Dept:

rxChiSquaredTest(admissCTabs2)

 Chi-squared test of independence between Gender and Admit 
	  X-squared df      p-value
 Dept==A 16.37177373  1 5.205468e-05
 Dept==B  0.08509801  1 7.705041e-01
 Dept==C  0.63322380  1 4.261753e-01
 Dept==D  0.22159370  1 6.378283e-01
 Dept==E  0.80804765  1 3.686981e-01
 Dept==F  0.21824336  1 6.403817e-01

rxFisherTest(admissCTabs2)

 Fisher's Exact Test for Count Data 
	odds ratio 95% CI Lower 95% CI Upper      p-value
 Dept==A  0.3495628    0.1970420    0.5920417 1.669189e-05
 Dept==B  0.8028124    0.2944986    2.0040231 6.770899e-01
 Dept==C  1.1329004    0.8452173    1.5162918 3.866166e-01
 Dept==D  0.9213798    0.6789572    1.2504742 5.994965e-01
 Dept==E  1.2211852    0.8064776    1.8385155 3.603964e-01
 Dept==F  0.8280944    0.4332888    1.5756278 5.458408e-01
   HA: two.sided 
   H0:  odds ratio = 1

Like Fisher’s exact test, the Kendall tau correlation test works best on smaller contingency tables. Here is an example of what it returns when applied to our admissions data (the results differ from run to run as the underlying algorithm relies on sampling):

rxKendallCor(admissCTabs2)

	       taub p-value
Dept==A -0.13596550   0.000
Dept==B -0.02082575   0.666
Dept==C  0.02865045   0.380
Dept==D -0.01939676   0.585
Dept==E  0.04140240   0.383
Dept==F -0.02319366   0.530
   HA: two.sided

Odds Ratios and Risk Ratios

Another common task associated with 2 x 2 contingency tables is the calculation of odds ratios and risk ratios (also known as relative risk). The two functions rxOddsRatio and rxRiskRatio in RevoScaleR can be used to compute these quantities. The odds ratio and the risk ratio are closely related: the odds ratio computes the relative odds of an event among two or more groups, while the risk ratio computes the relative probabilities of an event. Consider again the contingency table admissCTabs:

#  Odds Ratios and Risk Ratios

admissCTabs

		Admit
Gender   Admitted Rejected
  Male       1198     1493
  Female      557     1278

In this example, the odds of being admitted as a male are 1198/1493, or about 4 to 5 against. The odds of being admitted as a female are 557/1278, or about 4 to 9 against. The odds ratio is (1198/1493)/(557/1278), or 1.8 greater odds that a male will be admitted as opposed to a woman.

rxOddsRatio(admissCTabs)

  data:  
  Z = 0.6104, p-value < 2.2e-16
  alternative hypothesis: two.sided 
  95 percent confidence interval:
   1.624377 2.086693 
  sample estimates:
  oddsRatio 
    1.84108

The risk ratio, by contrast, compares the probabilities of being rejected, that is, 1493/(1198+1493) for a man versus 1278/(557+1278) for a woman. So here the risk ratio is 0.697 (the probability of a woman being rejected) divided by 0.555 (the probability of a man being rejected), or 1.255:

rxRiskRatio(admissCTabs)

 data:  
 Z.Female = 0.2274, p-value < 2.2e-16
 alternative hypothesis: two.sided 
 95 percent confidence interval:
  1.199631 1.313560 
 sample estimates:
 riskRatio.Female 
	 1.255303