Outlier Treatment

Introduction

Data treatment is the process of altering indicators to improve their statistical properties, mainly for the purposes of aggregation. Data treatment is a delicate subject, because it essentially involves changing the values of certain observations, or transforming an entire distribution. Like any other step or assumption though, any data treatment should be carefully recorded and its implications understood. Of course, data treatment does not have to be applied, it is simply another tool in your toolbox.

The Treat() function

The COINr function for treating data is called Treat(). This is a generic function with methods for coins, purses, data frames and numeric vectors. It is very flexible but this can add a layer of complexity. If you want to run mostly at default options, see the qTreat() function mentioned below in Simplified function.

The Treat() function operates a two-stage data treatment process, based on two data treatment functions (f1 and f2), and a pass/fail function f_pass which detects outliers. The arrangement of this function is inspired by a fairly standard data treatment process applied to indicators, which consists of checking skew and kurtosis, then if the criteria are not met, applying Winsorisation up to a specified limit. Then if Winsorisation still does not bring skew and kurtosis within limits, applying a nonlinear transformation such as log or Box-Cox.

This function generalises this process by using the following general steps:

  1. Check if variable passes or fails using f_pass
  2. If f_pass returns FALSE, apply f1, else return x unmodified
  3. Check again using f_pass
  4. If f_pass still returns FALSE, apply f2
  5. Return the modified x as well as other information.

For the “typical” case described above f1 is a Winsorisation function, f2 is a nonlinear transformation and f_pass is a skew and kurtosis check. However, any functions can be passed as f1, f2 and f_pass, which makes it a flexible tool that is also compatible with other packages.

Further details on how this works are given in the following sections.

Numeric vectors

The clearest way to demonstrate the Treat() function is on a numeric vector. Let’s make a vector with a couple of outliers:

# numbers between 1 and 10
x <- 1:10

# two outliers
x <- c(x, 30, 100)

We can check the skew and kurtosis of this vector:

library(COINr)

skew(x)
#> [1] 3.063241

kurt(x)
#> [1] 9.741391

The skew and kurtosis are both high. If we follow the default limits in COINr (absolute skew capped at 2, and kurtosis capped at 3.5), this would be classed as a vector with outliers. Indeed we can confirm this using the check_SkewKurt() function, which is the default pass/fail function used in Treat(). This also anyway outputs the skew and kurtosis:

check_SkewKurt(x)
#> $Pass
#> [1] FALSE
#> 
#> $Skew
#> [1] 3.063241
#> 
#> $Kurt
#> [1] 9.741391

Now we know that x has outliers, we can treat it (if we want). We use the Treat() function to specify that our function for checking for outliers f_pass = "check_SkewKurt", and our first function for treating outliers is f1 = "winsorise". We also pass an additional parameter to winsorise(), which is winmax = 2. You can check the winsorise() function documentation to better understand how it works.

l_treat <- Treat(x, f1 = "winsorise", f1_para = list(winmax = 2),
                 f_pass = "check_SkewKurt")

plot(x, l_treat$x)

The result of this data treatment is shown in the scatter plot: one point from x has been Winsorised (reassigned the next highest value). We can check the skew and kurtosis of the treated vector:

check_SkewKurt(l_treat$x)
#> $Pass
#> [1] TRUE
#> 
#> $Skew
#> [1] 1.712038
#> 
#> $Kurt
#> [1] 1.815781

Clearly, Winsorising one point was enough in this case to bring the skew and kurtosis within the specified thresholds.

Data frames

Treatment of a data frame with Treat() is effectively the same as treating a numeric vector, because the data frame method passes each column of the data frame to the numeric method. Here, we use some data from the COINr package to demonstrate.

# select three indicators
df1 <- ASEM_iData[c("Flights", "Goods", "Services")]

# treat the data frame using defaults
l_treat <- Treat(df1)

str(l_treat, max.level = 1)
#> List of 3
#>  $ x_treat       :'data.frame':  51 obs. of  3 variables:
#>  $ Dets_Table    :'data.frame':  3 obs. of  8 variables:
#>  $ Treated_Points:'data.frame':  51 obs. of  3 variables:

We can see the output is a list with x_treat, the treated data frame; Dets_Table, a table describing what happened to each indicator; and Treated_Points, which marks which individual points were adjusted. This is effectively the same output as for treating a numeric vector.

l_treat$Dets_Table
#>      iCode check_SkewKurt0.Pass check_SkewKurt0.Skew check_SkewKurt0.Kurt
#> 1  Flights                FALSE             2.103287             4.508879
#> 2    Goods                FALSE             2.649973             8.266610
#> 3 Services                 TRUE             1.701085             2.375656
#>   winsorise.nwin check_SkewKurt1.Pass check_SkewKurt1.Skew check_SkewKurt1.Kurt
#> 1              1                 TRUE             1.900658            3.3360647
#> 2              2                 TRUE             1.140608            0.1572047
#> 3             NA                   NA                   NA                   NA

We also check the individual points:

l_treat$Treated_Points
#>    Flights Goods Services
#> 1                        
#> 2                        
#> 3                        
#> 4                        
#> 5                        
#> 6                        
#> 7                        
#> 8                        
#> 9                        
#> 10                       
#> 11         winhi         
#> 12                       
#> 13                       
#> 14                       
#> 15                       
#> 16                       
#> 17                       
#> 18                       
#> 19                       
#> 20                       
#> 21                       
#> 22                       
#> 23                       
#> 24                       
#> 25                       
#> 26                       
#> 27                       
#> 28                       
#> 29                       
#> 30   winhi               
#> 31                       
#> 32                       
#> 33                       
#> 34                       
#> 35         winhi         
#> 36                       
#> 37                       
#> 38                       
#> 39                       
#> 40                       
#> 41                       
#> 42                       
#> 43                       
#> 44                       
#> 45                       
#> 46                       
#> 47                       
#> 48                       
#> 49                       
#> 50                       
#> 51

Coins

Treating coins is a simple extension of treating a data frame. The coin method simply extracts the relevant data set as a data frame, and passes it to the data frame method. So more or less, the same arguments are present.

We begin by building the example coin, which will be used for the examples here.

coin <- build_example_coin(up_to = "new_coin")
#> iData checked and OK.
#> iMeta checked and OK.
#> Written data set to .$Data$Raw

Default treatment

The Treat() function can be applied directly to a coin with completely default options:

coin <- Treat(coin, dset = "Raw")
#> Written data set to .$Data$Treated

For each indicator, the Treat() function:

  1. Checks skew and kurtosis using the check_SkewKurt() function
  2. If the indicator fails the test (returns FALSE), applies Winsorisation
  3. Checks again skew and kurtosis
  4. If the indicator still fails, applies a log transformation.

If at any stage the indicator passes the skew and kurtosis test, it is returned without further treatment.

When we run Treat() on a coin, it also stores information returned from f1, f2 and f_pass in the coin:

# summary of treatment for each indicator
head(coin$Analysis$Treated$Dets_Table)
#>     iCode check_SkewKurt0.Pass check_SkewKurt0.Skew check_SkewKurt0.Kurt
#> 1     LPI                 TRUE           -0.3042681           -0.6567514
#> 2 Flights                FALSE            2.1032872            4.5088794
#> 3    Ship                 TRUE           -0.5756680           -0.6814795
#> 4    Bord                FALSE            2.1482360            5.7914905
#> 5    Elec                FALSE            2.2252736            5.7910268
#> 6     Gas                FALSE            2.8294486           10.3346494
#>   winsorise.nwin check_SkewKurt1.Pass check_SkewKurt1.Skew check_SkewKurt1.Kurt
#> 1             NA                   NA                   NA                   NA
#> 2              1                 TRUE             1.900658             3.336065
#> 3             NA                   NA                   NA                   NA
#> 4              1                 TRUE             1.899211             4.346298
#> 5              1                 TRUE             1.717744             2.586062
#> 6              1                 TRUE             1.602518             1.525576

Notice that only one treatment function was used here, since after Winsorisation (f1), all indicators passed the skew and kurtosis test (f_pass).

In general, Treat() tries to collect all information returned from the functions that it calls. Details of the treatment of individual points are also stored in .$Analysis$Treated$Treated_Points.

The Treat() function gives you a high degree of control over which functions are used to treat and test indicators, and it is also possible to specify different functions for different indicators. Let’s begin though by seeing how we can change the specifications for all indicators, before proceeding to individual treatment.

Unless indiv_specs is specified (see later), the same procedure is applied to all indicators. This process is specified by the global_specs argument. To see how to use this, it is easiest to show the default of this argument which is built into the treat() function:

# default treatment for all cols
specs_def <- list(f1 = "winsorise",
                  f1_para = list(na.rm = TRUE,
                                 winmax = 5,
                                 skew_thresh = 2,
                                 kurt_thresh = 3.5,
                                 force_win = FALSE),
                  f2 = "log_CT",
                  f2_para = list(na.rm = TRUE),
                  f_pass = "check_SkewKurt",
                  f_pass_para = list(na.rm = TRUE,
                                     skew_thresh = 2,
                                     kurt_thresh = 3.5))

Notice that there are six entries in the list:

To understand what the individual parameters do, for example in f1_para, we need to look at the function called by f1, which is the winsorise() function:

Here we see the same parameters as named in the list f1_para, and we can change the maximum number of points to be Winsorised, the skew and kurtosis thresholds, and other things.

To make adjustments, unless we want to redefine everything, we don’t need to specify the entire list. So for example, if we want to change the maximum Winsorisation limit winmax, we can just pass this part of the list (notice we still have to wrap the parameter inside a list):

# treat with max winsorisation of 3 points
coin <- Treat(coin, dset = "Raw", global_specs = list(f1_para = list(winmax = 1)))
#> Written data set to .$Data$Treated
#> (overwritten existing data set)

# see what happened
coin$Analysis$Treated$Dets_Table |>
  head(10)
#>       iCode check_SkewKurt0.Pass check_SkewKurt0.Skew check_SkewKurt0.Kurt
#> 1       LPI                 TRUE           -0.3042681           -0.6567514
#> 2   Flights                FALSE            2.1032872            4.5088794
#> 3      Ship                 TRUE           -0.5756680           -0.6814795
#> 4      Bord                FALSE            2.1482360            5.7914905
#> 5      Elec                FALSE            2.2252736            5.7910268
#> 6       Gas                FALSE            2.8294486           10.3346494
#> 7  ConSpeed                 TRUE            0.4622037            0.1873214
#> 8     Cov4G                 TRUE           -1.3725191            0.5419314
#> 9     Goods                FALSE            2.6499733            8.2666095
#> 10 Services                 TRUE            1.7010849            2.3756557
#>    winsorise.nwin check_SkewKurt1.Pass check_SkewKurt1.Skew
#> 1              NA                   NA                   NA
#> 2               1                 TRUE             1.900658
#> 3              NA                   NA                   NA
#> 4               1                 TRUE             1.899211
#> 5               1                 TRUE             1.717744
#> 6               1                 TRUE             1.602518
#> 7              NA                   NA                   NA
#> 8              NA                   NA                   NA
#> 9               1                FALSE             2.469910
#> 10             NA                   NA                   NA
#>    check_SkewKurt1.Kurt check_SkewKurt2.Pass check_SkewKurt2.Skew
#> 1                    NA                   NA                   NA
#> 2              3.336065                   NA                   NA
#> 3                    NA                   NA                   NA
#> 4              4.346298                   NA                   NA
#> 5              2.586062                   NA                   NA
#> 6              1.525576                   NA                   NA
#> 7                    NA                   NA                   NA
#> 8                    NA                   NA                   NA
#> 9              7.087309                 TRUE           0.03104001
#> 10                   NA                   NA                   NA
#>    check_SkewKurt2.Kurt
#> 1                    NA
#> 2                    NA
#> 3                    NA
#> 4                    NA
#> 5                    NA
#> 6                    NA
#> 7                    NA
#> 8                    NA
#> 9            -0.8888965
#> 10                   NA

Having imposed a much stricter Winsorisation limit (only one point), we can see that now one indicator has been passed to the second treatment function f2, which has performed a log transformation. After doing this, the indicator passes the skew and kurtosis test.

By default, if an indicator does not satisfy f_pass after applying f1, it is passed to f2 in its original form - in other words it is not the output of f1 that is passed to f2, and f2 is applied instead of f1, rather than in addition to it. If you want to apply f2 on top of f1 set combine_treat = TRUE. In this case, if f_pass is not satisfied after f1 then the output of f1 is used as the input of f2. For the defaults of f1 and f2 this approach is probably not advisable because Winsorisation and the log transform are quite different approaches. However depending on what you want to do, it might be useful.

Individual treatment

The global_specs specifies the treatment methodology to apply to all indicators. However, the indiv_specs argument (if specified), can be used to override the treatment specified in global_specs for specific indicators. It is specified in exactly the same way as global_specs but requires a parameter list for each indicator that is to have individual specifications applied, wrapped inside one list.

This is probably clearer using an example. To begin with something simple, let’s say that we keep the defaults for all indicators except one, where we change the Winsorisation limit. We will set the Winsorisation limit of the indicator “Flights” to zero, to force it to be log-transformed.

# change individual specs for Flights
indiv_specs <- list(
  Flights = list(
    f1_para = list(winmax = 0)
  )
)

# re-run data treatment
coin <- Treat(coin, dset = "Raw", indiv_specs = indiv_specs)
#> Written data set to .$Data$Treated
#> (overwritten existing data set)

The only thing to remember here is to make sure the list is created correctly. Each indicator to assign individual treatment must have its own list - here containing f1_para. Then f1_para itself is a list of named parameter values for f1. Finally, all lists for each indicator have to be wrapped into a single list to pass to indiv_specs. This looks a bit convoluted for changing a single parameter, but gives a high degree of control over how data treatment is performed.

We can now see what happened to “Flights”:

coin$Analysis$Treated$Dets_Table[
  coin$Analysis$Treated$Dets_Table$iCode == "Flights", 
]
#>     iCode check_SkewKurt0.Pass check_SkewKurt0.Skew check_SkewKurt0.Kurt
#> 2 Flights                FALSE             2.103287             4.508879
#>   winsorise.nwin check_SkewKurt1.Pass check_SkewKurt1.Skew check_SkewKurt1.Kurt
#> 2              0                FALSE             2.103287             4.508879
#>   check_SkewKurt2.Pass check_SkewKurt2.Skew check_SkewKurt2.Kurt
#> 2                 TRUE          -0.09502644           -0.8305217

Now we see that “Flights” didn’t pass the first Winsorisation step (because nothing happened to it), and was passed to the log transform. After that, the indicator passed the skew and kurtosis check.

As another example, we may wish to exclude some indicators from data treatment completely. To do this, we can set the corresponding entries in indiv_specs to "none". This is the only case where we don’t have to pass a list for each indicator.

# change individual specs for two indicators
indiv_specs <- list(
  Flights = "none",
  LPI = "none"
)

# re-run data treatment
coin <- Treat(coin, dset = "Raw", indiv_specs = indiv_specs)
#> Written data set to .$Data$Treated
#> (overwritten existing data set)

Now if we examine the treatment table, we will find that these indicators have been excluded from the table, as they were not subjected to treatment.

External functions

Any functions can be passed to Treat(), for both treating and checking for outliers. As an example, we can pass an outlier detection function ` from the performance package

The following code chunk will only run if you have the ‘performance’ package installed.

library(performance)

# the check_outliers function outputs a logical vector which flags specific points as outliers.
# We need to wrap this to give a single TRUE/FALSE output, where FALSE means it doesn't pass,
# i.e. there are outliers
outlier_pass <- function(x){
  # return FALSE if any outliers
  !any(check_outliers(x))
}

# now call treat(), passing this function
# we set f_pass_para to NULL to avoid passing default parameters to the new function
coin <- Treat(coin, dset = "Raw",
               global_specs = list(f_pass = "outlier_pass",
                                    f_pass_para = NULL)
)
#> Written data set to .$Data$Treated
#> (overwritten existing data set)

# see what happened
coin$Analysis$Treated$Dets_Table |>
  head(10)
#>       iCode outlier_pass0 winsorise.nwin outlier_pass1 outlier_pass2
#> 1       LPI         FALSE              0         FALSE         FALSE
#> 2   Flights         FALSE              1         FALSE          TRUE
#> 3      Ship         FALSE              0         FALSE         FALSE
#> 4      Bord         FALSE              1         FALSE         FALSE
#> 5      Elec         FALSE              1         FALSE          TRUE
#> 6       Gas         FALSE              1         FALSE         FALSE
#> 7  ConSpeed         FALSE              0         FALSE         FALSE
#> 8     Cov4G         FALSE              0         FALSE         FALSE
#> 9     Goods         FALSE              2         FALSE          TRUE
#> 10 Services         FALSE              0         FALSE          TRUE

Here we see that the test for outliers is much stricter and very few of the indicators pass the test, even after applying a log transformation. Clearly, how an outlier is defined can vary and depend on your application.

Purses

The purse method for treat() is fairly straightforward. It takes almost the same arguments as the coin method, and applies the same specifications to each coin. Here we simply demonstrate it on the example purse.

# build example purse
purse <- build_example_purse(up_to = "new_coin", quietly = TRUE)

# apply treatment to all coins in purse (default specs)
purse <- Treat(purse, dset = "Raw")
#> Written data set to .$Data$Treated
#> Written data set to .$Data$Treated
#> Written data set to .$Data$Treated
#> Written data set to .$Data$Treated
#> Written data set to .$Data$Treated

Simplified function

The Treat() function is very flexible but comes at the expense of a possibly fiddly syntax. If you don’t need that level of flexibility, consider using qTreat(), which is a simplified wrapper for Treat().

The main features of qTreat() are that:

The qTreat() function is a generic with methods for data frames, coins and purses. Here, we’ll just demonstrate it on a data frame.

# select three indicators
df1 <- ASEM_iData[c("Flights", "Goods", "Services")]

# treat data frame, changing winmax and skew/kurtosis limits
l_treat <- qTreat(df1, winmax = 1, skew_thresh = 1.5, kurt_thresh = 3)

Now we check what the results are:

l_treat$Dets_Table
#>      iCode check_SkewKurt0.Pass check_SkewKurt0.Skew check_SkewKurt0.Kurt
#> 1  Flights                FALSE             2.103287             4.508879
#> 2    Goods                FALSE             2.649973             8.266610
#> 3 Services                 TRUE             1.701085             2.375656
#>   winsorise.nwin check_SkewKurt1.Pass check_SkewKurt1.Skew check_SkewKurt1.Kurt
#> 1              1                FALSE             1.900658             3.336065
#> 2              1                FALSE             2.469910             7.087309
#> 3             NA                   NA                   NA                   NA
#>   check_SkewKurt2.Pass check_SkewKurt2.Skew check_SkewKurt2.Kurt
#> 1                 TRUE          -0.09502644           -0.8305217
#> 2                 TRUE           0.03104001           -0.8888965
#> 3                   NA                   NA                   NA

We can see that in this case, Winsorsing by one point was not enough to bring “Flights” and “Goods” within the specified skew/kurtosis limits. Consequently, f2 was invoked, which uses a log transform and brought both indicators within the specified limits.