CoDaImpact is a package for the analysis of covariate impacts in compositional regression models. It is designed to work in conjunction with the compositions package. The compositions package is used to fit a regression model involving compositional variables and the CoDaImpact package provides new functions to interpret the estimation results.
The interface between the two packages is implemented by the
CoDaImpact::ToSimplex() function which converts the usual
lm object into an object of class lmCoDa. This new
class is associated with new interpretation methods that make use of
elasticities and semi-elasticities (see Morais and Thomas-Agnan, 2021)
and share ratio elasticities (see Dargel and Thomas-Agnan, 2024). A
finite increments interpretation is also provided in the function
CoDaImpact::VariationTable(). Finally, some visualization
tools are provided.
The Dargel and Thomas-Agnan (2024) paper is illustrated by vignette and all materials for reproducing the results in the paper and the vignette can be found on a dedicated GitHub repository.
The following three sections illustrate the use of the CoDaImpact package for three classes of regression models:
<! – Section 4 provides implementation details. –> <! – The
package contains three data sets used as examples:
rice_yields, car_market,
elections. –>
The following code loads all the packages required for this vignette. The CoDaImpact package has to be installed from GitHub for now and should be released on CRAN later on.
This case is illustrated using an example from Trinh et al. (2023) studying the impact of climate change on rice yield in Vietnam.
The rice_yields dataset contains the rice production
“YIELD” in each province of Vietnam, between 1987 and 2016. The data on
rice yields originate from “International Rice Research Institute”. The
maximum daily temperature data originate from the database “Climate
Prediction Center (CPC)”, created by “National Oceanic and Atmospheric
Administration (NOAA)”. The temperatures have been interpolated over the
63 provinces using areal interpolation methods. For this illustration,
we group the temperatures in three classes LOW ([-6,25.1]),
MEDIUM ([25.1,35.4]) and HIGH ([35.4,45]). The
precipitation data come from the project “Asian
Precipitation-Highly-Resolved Observational Data Integration Towards
Evaluation (APHRODITE)”.
##           PROVINCE                                        REGION   
##  ANGIANG      :  30   Central Highlands                      :180  
##  BACGIANG     :  30   Mekong Delta                           :420  
##  BACKAN       :  30   Midlands and Northern Mountainous Areas:390  
##  BACLIEU      :  30   Northern and Coastal Central Region    :420  
##  BACNINH      :  30   Red River Delta                        :330  
##  BARIA-VUNGTAU:  30   Southeastern Area                      :150  
##  (Other)      :1710                                                
##       YEAR          YIELD       PRECIPITATION   
##  Min.   :1987   Min.   :1.120   Min.   :0.1693  
##  1st Qu.:1994   1st Qu.:3.094   1st Qu.:1.1656  
##  Median :2002   Median :4.096   Median :1.4040  
##  Mean   :2002   Mean   :4.077   Mean   :1.3824  
##  3rd Qu.:2009   3rd Qu.:5.027   3rd Qu.:1.6411  
##  Max.   :2016   Max.   :6.637   Max.   :3.7032  
##                                                 
##    TEMPERATURES.LOW    TEMPERATURES.MEDIUM    TEMPERATURES.HIGH  
##  Min.   :0.0000003     Min.   :0.5150684     Min.   :0.00000020  
##  1st Qu.:0.0027399     1st Qu.:0.6438356     1st Qu.:0.00273983  
##  Median :0.1945206     Median :0.7657533     Median :0.02185802  
##  Mean   :0.1704263     Mean   :0.7957867     Mean   :0.03378702  
##  3rd Qu.:0.3095891     3rd Qu.:0.9726024     3rd Qu.:0.05205488  
##  Max.   :0.4699453     Max.   :0.9999995     Max.   :0.25479457  
## The following plot is a ternary diagram of maximum temperature colored according to rice yield.
whitered_pal <- colorRampPalette(rev(c("beige","orange","red","darkred")))
top100 <- head(rice_yields, 250)
whitered_col <- whitered_pal(length(top_100$YIELD))## Error in eval(expr, envir, enclos): object 'top_100' not found## Error in eval(expr, envir, enclos): object 'top_100' not found## Error in eval(expr, envir, enclos): object 'whitered_col' not foundlayout(matrix(1:2,ncol=2), width = c(2,1),height = c(1,1))
plot(acomp(top_100$TEMPERATURES),
     col = ylcol_top100,
     pch = 16)## Error in eval(expr, envir, enclos): object 'top_100' not foundlegend_image <- as.raster(matrix(whitered_pal(20), ncol=1))
plot(c(0,2),c(0,1),type = 'n', axes = F,xlab = '', ylab = '', main = 'Rice yield')
yield_range <- range(top_100$YIELD)## Error in eval(expr, envir, enclos): object 'top_100' not found## Error in eval(expr, envir, enclos): object 'yield_range' not found## Error in eval(expr, envir, enclos): object 'yield_range' not foundThis model explains a scalar variable YIELD with a
combination of scalar and compositional variables. Before estimating, we
have to store our data in a data.frame in such a way that scalar
variables are column vectors and compositional variables are matrices.
In our dataset, the TEMPERATURES variable is already in the
matrix format.
## $PROVINCE
## [1] "factor"
## 
## $REGION
## [1] "factor"
## 
## $YEAR
## [1] "numeric"
## 
## $YIELD
## [1] "numeric"
## 
## $PRECIPITATION
## [1] "numeric"
## 
## $TEMPERATURES
## [1] "matrix" "array"##             LOW          MEDIUM            HIGH 
## 0.0000002999999 0.9616434547947 0.0383562452054We continue to fit a regression model with the tools provided by base
R and the compositions package, in particular, the
lm() function and the alr() transformation for
the TEMPERATURES variable.
## [1] "lm"In the next step the function ToSimplex() transforms the
above output into a “lmCoDa” class which serves as the interface between
the lm-output and the interpretation functions.
## [1] "lmCoDa" "lm"The same can be achieved in a single step using the
lmCoDa() function.
fit_X_compo <- lmCoDa(YIELD ~ PRECIPITATION + alr(TEMPERATURES),data = rice_yields)
class(fit_X_compo)## [1] "lmCoDa" "lm"The estimated parameters can be retrieved by the coef()
function. For a compositional explanatory variable the option
space allows choosing between the simplex version, the clr
version or the coordinate version of the parameters.
##                   YIELD
## (Intercept)   3.5835270
## PRECIPITATION 0.2528452
## LOW           0.3338069
## MEDIUM        0.3413303
## HIGH          0.3248628##                      YIELD
## (Intercept)    3.583526963
## PRECIPITATION  0.252845153
## LOW            0.001623843
## MEDIUM         0.023911877
## HIGH          -0.025535720##                               YIELD
## (Intercept)             3.583526963
## PRECIPITATION           0.252845153
## alr(TEMPERATURES)LOW    0.001623843
## alr(TEMPERATURES)MEDIUM 0.023911877The fitted() function retrieves the fitted values.
##        1        2        3        4        5        6 
## 3.924848 3.736444 3.948922 4.013668 3.871424 3.952286Finite increments interpretation as in Muller et al. (2018) can be obtained using outputs in coordinate space. For the case of scalar on composition regression, more interpretations are found in Coenders and Pawlovsky-Glahn (2020). The different formulations they propose correspond to different parametrizations of the model and we illustrate below their interpretation based on clr coordinates.
Using the approach of Muller et al. (2018) we interpret an additive
increment for a specific ilr coordinate that corresponds to a so-called
principal balance. To use their interpretation we first need to
transform the parameters into the appropriate ilr space, which is
defined by the contrast matrix. The ilrBase() function
without further arguments leads to the principal balance of the last
component against all others.
##         [,1]       [,2]
## 1 -0.7071068 -0.4082483
## 2  0.7071068 -0.4082483
## 3  0.0000000  0.8164966We can use this contrast matrix and the parameters in clr space to obtain the ilr parameters.
clr_TEMP <- coef(fit_X_compo, space = "clr", split = TRUE)[["alr(TEMPERATURES)"]]
ilr_TEMP <- t(V_TEMP) %*% clr_TEMP
ilr_TEMP##            YIELD
## [1,]  0.01576002
## [2,] -0.03127474The interpretation is that an additive increment of one unit in the
second ilr coordinate leads to an expected decrease in the rice yield by
0.03 tonnes per hectare. This increment means that the ratio between
HIGH and the geometric mean of LOW and
MEDIUM is approximately multiplied by \(\exp(1/0.82)\sim 3.44,\) where \(O.82\) is the element \((3,2)\) of the contrast matrix.
In terms of temperature distribution, the above increment implies
that the third component HIGH of the temperature
composition grows at the expense of the others (the geometric mean of
the other two) in a way that the ratio of LOW to
MEDIUM remains constant. This corresponds to a change in
the TEMPERATURES simplex in the direction of the vertex
HIGH. Changing for another vertex is simple by a
permutation of the rows of the contrast matrix.
This method only allows to evaluate the impacts of changing a compositional covariate in the direction of a vertex. In Dargel and Thomas-Agnan (2024) it is shown how the impact of more general changes in explanatory compositions can be interpreted.
The clr parameter estimates for the temperatures are given by:
##               YIELD
## LOW     0.001623843
## MEDIUM  0.023911877
## HIGH   -0.025535720We follow Coenders and Pawlovsky-Glahn (2020) but without changing
the base of the logarithm and the normalization constant. In these
conditions, the first clr parameter for the level LOW of
temperature is interpreted as the expected change of the response when
the ratio between LOW and the geometric mean of
LOW, MEDIUM and HIGH is
multiplied by \(\exp(1)\sim 2.7\). The
resulting change in expected yield is therefore an increase of \(0.0016\) tons per hectare.
To evaluate the impact of a given compositional covariate, we create scenarios of changes in its simplex space described by a linear equation \(x(h)=x(0)\oplus h\odot u,\) where \(x(0)\) is the initial point, \(h\) is the signed-intensity of change and \(u\) is a vector of the simplex defining the direction of change. We then compute the predicted value of \(y(h)\) corresponding to \(x(h)\).
We consider a change in the direction of the vertex “LOW” of the variable “TEMPERATURE”, and we predict new values of the response variable according to the change in percentage points of this vertex.
VariationScenario(
  fit_X_compo,
  Xvar = "TEMPERATURES",
  Xdir = c(0.2, 0.5, 0.3),
  inc_size = 2,
  n_steps = 5,
  add_opposite = TRUE,
  obs=43)##       YIELD           X.LOW        X.MEDIUM          X.HIGH
## -5 3.819922 0.9989636850642 0.0000018748701 0.0010344400657
## -4 3.855516 0.9963715150059 0.0000314423589 0.0035970426351
## -3 3.891110 0.9870530721012 0.0005237290466 0.0124231988521
## -2 3.926704 0.9498471187296 0.0084740730707 0.0416788081997
## -1 3.962298 0.7674684242080 0.1151254523609 0.1174061234311
## 0  3.997893 0.2465753931507 0.6219176838356 0.1315069230137
## 1  4.033487 0.0220905961112 0.9368346502406 0.0410747536483
## 2  4.069081 0.0013878377613 0.9896156353605 0.0089965268782
## 3  4.104675 0.0000832426310 0.9980354895354 0.0018812678336
## 4  4.140269 0.0000049585605 0.9996043543177 0.0003906871218
## 5  4.175863 0.0000002949985 0.9999186719961 0.0000810330054For a good visual impression, it is better to use a smaller step size and a larger number of steps. Below is the ternary diagram showing the path in the simplex: the initial point is in black and the other ones are colored according to the value of rice yield.
VS2 <- VariationScenario(
  fit_X_compo,
  Xvar = "TEMPERATURES",
  Xdir = c(0.2, 0.5, 0.3),
  inc_size = .1,
  n_steps = 40,
  add_opposite = TRUE,
  obs=43)
yield_color <- whitered_pal(length(VS2$YIELD))
yield_color <- yield_color[rank(VS2$YIELD,ties.method = "first")]
plot(acomp(VS2$X), col = yield_color, pch = 16)
plot(acomp(VS2["0", "X"]), add = TRUE, pch = 16)The use of semi-elasticities for the interpretation of compositional
covariates is described in Morais and Thomas-Agnan (2021). For a scalar
on composition model the Impacts() method computes the
semi-elasticities for all compositional covariates.
RY_Impacts <- Impacts(fit_X_compo, Xvar = "TEMPERATURES")
RY_Impacts # here, the semi-elasticity is computed for the variable TEMPERATURES ##               YIELD
## LOW     0.001623843
## MEDIUM  0.023911877
## HIGH   -0.025535720Dargel and Thomas-Agnan (2024) use semi-elasticities to interpret
infinitesimal changes due to variation scenarios. The function
VariationTable computes the impact of incremental changes
in covariates on the response variable. The covariate increments,
corresponding to linear changes in the simplex, are indexed by a
direction vector and a signed intensity parameter inc_size
and are specific to a given observation. Two cases are illustrated below
with a direction pointing to a vertex of the simplex and a general
direction. For the vertex direction, the inc_rate is
related to the inc_size parameter and measures the relative
change of the share corresponding to this vertex.
The direction pointing to a vertex corresponds to the increments
considered by Muller et al. (2018). For example, we can measure the
change caused by increasing the HIGH share of the
TEMPERATURES variable by \(5\%\), assuming that the ratio between the
other two components remains constant, using the function
VariationTable().
VariationTable(
  fit_X_compo,
  obs      =  1,             # indicator of the observation (1 is default)
  Xvar     = "TEMPERATURES", # covariate for which the impact is computed
  Xdir     = 'HIGH',         # vertex in the covariate simplex 
  inc_rate = 0.05)           # signed intensity ##                           YIELD
## Initial value       3.924848006
## New value           3.919636938
## Semi elasticity    -0.031274742
## Variation in %     -0.132771206
## Variation in units -0.005211068The interpretation of this result is as follows: a \(5\%\) change of the HIGH share
assuming the ratio between the LOW and the
MEDIUM share remains constant will result in a \(-0.0052\) unit change in
YIELD.
Let us now consider changing the covariate in the direction \((0.2,0.55,0.25)\), with an increment size of \(0.05\). Note that this direction vector will be normalized before being used within the function, and will later be given as output in attributes.
fit_X_compo.varTab <- VariationTable(
  fit_X_compo,
  obs      = 10,               # indicator of the observation
  Xvar     = "TEMPERATURES",   # covariate for which the impact is computed
  Xdir     = c(0.2,0.55,0.25), # general direction in the covariate simplex
  inc_size = 0.05)             # signed intensity
fit_X_compo.varTab##                          YIELD
## Initial value      4.010224763
## New value          4.015157689
## Semi elasticity    0.024601746
## Variation in %     0.123008728
## Variation in units 0.004932926## $names
## [1] "YIELD"
## 
## $class
## [1] "data.frame"
## 
## $row.names
## [1] "Initial value"      "New value"          "Semi elasticity"   
## [4] "Variation in %"     "Variation in units"
## 
## $`X(0)`
##               LOW            MEDIUM              HIGH 
## "0.0000002999999" "0.9506845561646" "0.0493151438356" 
## attr(,"class")
## [1] "acomp"
## 
## $`X(h)`
## [1] 0.0000002805874 0.9531436237124 0.0468560957002
## 
## $Xdir
## [1] 0.1616218 0.6208904 0.2174878
## 
## $inc_size
## [1] 0.05
## 
## $inc_rates
## [1] -0.064708144  0.002586628 -0.049863955The move between the temperature share vector of \((0.0000002999999, 0.9506845561646,
0.0493151438356)\) to \((0.0000002805874, 0.9531436237124,
0.0468560957002)\) results in an increase of the
YIELD by \(0.004932926\)
units. Note that the Aitchison distance between these two points is
equal to the increment size \(0.05\).
This case is illustrated using an example from Morais et al. (2018) studying the impact of some socio-economic variables on the market shares in each of the five market segments A, B, C, D and E.
This dataset shows monthly data on the French automobile market
between 2003 and 2015. The market is divided into 5 main segments
(SEG_A to SEG_E), according to the size of the
vehicle chassis. Apart from data on market shares, we have four
classical variables, such as HOUSEHOLD_EXPENDITURE, gross
domestic product (GDP), the average national price of
diesel fuel (GAS_PRICE), and
SCRAPPING_SUBSIDY. We also have the DATE,
generally the beginning of each month in which the sale has taken
place.
##       DATE                SEG_A             SEG_B            SEG_C       
##  Min.   :2003-01-01   Min.   :0.05178   Min.   :0.3051   Min.   :0.2959  
##  1st Qu.:2006-02-22   1st Qu.:0.06979   1st Qu.:0.3721   1st Qu.:0.3398  
##  Median :2009-04-16   Median :0.07841   Median :0.3909   Median :0.3593  
##  Mean   :2009-04-16   Mean   :0.08737   Mean   :0.3910   Mean   :0.3654  
##  3rd Qu.:2012-06-08   3rd Qu.:0.09893   3rd Qu.:0.4125   3rd Qu.:0.3883  
##  Max.   :2015-08-01   Max.   :0.16036   Max.   :0.4576   Max.   :0.4484  
##      SEG_D             SEG_E              GDP         HOUSEHOLD_EXPENDITURE
##  Min.   :0.06481   Min.   :0.01845   Min.   :134773   Min.   :71647        
##  1st Qu.:0.09159   1st Qu.:0.02596   1st Qu.:151629   1st Qu.:81181        
##  Median :0.10665   Median :0.03619   Median :164700   Median :88666        
##  Mean   :0.11257   Mean   :0.04362   Mean   :162472   Mean   :86882        
##  3rd Qu.:0.13434   3rd Qu.:0.06109   3rd Qu.:173575   3rd Qu.:93219        
##  Max.   :0.18964   Max.   :0.10123   Max.   :182620   Max.   :96548        
##    GAS_PRICE     SCRAPPING_SUBSIDY
##  Min.   :0.755   Min.   :0.0000   
##  1st Qu.:1.019   1st Qu.:0.0000   
##  Median :1.125   Median :0.0000   
##  Mean   :1.140   Mean   :0.1645   
##  3rd Qu.:1.317   3rd Qu.:0.0000   
##  Max.   :1.445   Max.   :1.0000opar <- par(mar=c(5.1, 4.1, 4.1, 8.1), xpd=TRUE)
plot(x = car_market$DATE, y = car_market$SEG_A,type = "l", col = "red",
     main = "French vehicles market shares from 2003 to 2015", 
     xlab = "DATE", ylab = "VALUE", ylim = c(0,0.5))
lines(x = car_market$DATE, y = car_market$SEG_B,type = "l", col = "blue" )
lines(x = car_market$DATE, y = car_market$SEG_C,type = "l", col = "green")
lines(x = car_market$DATE, y = car_market$SEG_D,type = "l", col = "orange")
lines(x = car_market$DATE, y = car_market$SEG_E,type = "l", col = "black")
legend("topright",
       legend = paste0("Segment ", LETTERS[1:5]),
       col = c("red", "blue", "green", "orange", "black"),
       lty = 1,
       inset=c(-0.35,0))We directly use the lmCoDa() function to fit a
regression model explaining the market shares as a function of household
expenditure, GDP, gas price and an indicator of the scrapping incentive
period. As previously, the compositional variable must be included as a
matrix.
Below we transform the response by an ilr, using the
ilr() function of the compositions
package.
fit_Y_compo <- lmCoDa(
  ilr(SEG) ~ HOUSEHOLD_EXPENDITURE + GDP + GAS_PRICE + SCRAPPING_SUBSIDY,
  data=car_market)The estimated coefficients are retrieved using the
coef() function.
##                            SEG_A      SEG_B     SEG_C     SEG_D     SEG_E
## (Intercept)           0.00510613 0.02979117 0.0299574 0.1738140 0.7613313
## HOUSEHOLD_EXPENDITURE 0.20004025 0.20000871 0.2000042 0.1999921 0.1999548
## GDP                   0.19998123 0.19999792 0.2000000 0.2000025 0.2000184
## GAS_PRICE             0.13249844 0.12769598 0.1583048 0.1951823 0.3863185
## SCRAPPING_SUBSIDY     0.25071823 0.19046324 0.1729588 0.1825495 0.2033102##                            ilr(SEG)1      ilr(SEG)2      ilr(SEG)3
## (Intercept)            1.24717388192  0.72459957006  2.03502320226
## HOUSEHOLD_EXPENDITURE -0.00011150825 -0.00008293645 -0.00011093018
## GDP                    0.00005900412  0.00004246356  0.00004089033
## GAS_PRICE             -0.02610541082  0.16036864757  0.29475342664
## SCRAPPING_SUBSIDY     -0.19436277707 -0.19093033990 -0.08827046806
##                           ilr(SEG)4
## (Intercept)            2.8974651795
## HOUSEHOLD_EXPENDITURE -0.0002527637
## GDP                    0.0001027078
## GAS_PRICE              0.8389659267
## SCRAPPING_SUBSIDY      0.0279658835For models with compositional response the fitted() also
has a space argument allowing to retrieve the fitted values in
coordinate space, clr space or in the simplex.
##   SEG_A        SEG_B       SEG_C       SEG_D       SEG_E       
## 1 "0.06550853" "0.3518854" "0.3312294" "0.1636691" "0.08770752"
## 2 "0.06506734" "0.3490900" "0.3309363" "0.1646582" "0.09024814"
## 3 "0.06450031" "0.3455116" "0.3305134" "0.1659005" "0.09357414"
## 4 "0.06266230" "0.3512741" "0.3315112" "0.1640017" "0.09055064"
## 5 "0.06327936" "0.3553624" "0.3319284" "0.1625658" "0.08686404"
## 6 "0.06335595" "0.3558713" "0.3319754" "0.1623847" "0.08641256"
## attr(,"class")
## [1] "acomp"##        SEG_A     SEG_B     SEG_C      SEG_D      SEG_E
## 1 -0.9018499 0.7792754 0.7187808 0.01381641 -0.6100227
## 2 -0.9123999 0.7675070 0.7141032 0.01604937 -0.5852597
## 3 -0.9258272 0.7525292 0.7081499 0.01889132 -0.5537431
## 4 -0.9439949 0.7798119 0.7219066 0.01812205 -0.5758457
## 5 -0.9286494 0.7969295 0.7287105 0.01487411 -0.6118646
## 6 -0.9267313 0.7990692 0.7295609 0.01446811 -0.6163670##       [,1]      [,2]       [,3]       [,4]
## 1 1.188735 0.6369230 -0.1601446 -0.6820261
## 2 1.187874 0.6422151 -0.1504177 -0.6543402
## 3 1.186777 0.6489506 -0.1380381 -0.6191037
## 4 1.218916 0.6564617 -0.1453068 -0.6438151
## 5 1.220169 0.6487640 -0.1594550 -0.6840855
## 6 1.220325 0.6478018 -0.1612235 -0.6891192
## attr(,"class")
## [1] "rmult"The same functionality is available for the residuals.
##   SEG_A       SEG_B       SEG_C       SEG_D       SEG_E      
## 1 "0.2100350" "0.2099886" "0.1900386" "0.1819359" "0.2080019"
## 2 "0.2241523" "0.1884051" "0.2165972" "0.1887411" "0.1821043"
## 3 "0.1990092" "0.1761677" "0.2160209" "0.2230781" "0.1857242"
## 4 "0.2015812" "0.2391805" "0.2238778" "0.2106168" "0.1247437"
## 5 "0.2229353" "0.2138702" "0.2243390" "0.1821381" "0.1567175"
## 6 "0.1971703" "0.2164442" "0.2369534" "0.1682997" "0.1811324"
## attr(,"class")
## [1] "acomp"##          SEG_A       SEG_B       SEG_C       SEG_D       SEG_E
## 1  0.050722716  0.05050182 -0.04932391 -0.09289665  0.04099603
## 2  0.117542428 -0.05618878  0.08325649 -0.05440689 -0.09020325
## 3 -0.001038231 -0.12295300  0.08098568  0.11313241 -0.07012686
## 4  0.032029651  0.20305589  0.13693794  0.07587790 -0.44790138
## 5  0.117957913  0.07644556  0.12423438 -0.08415879 -0.23447906
## 6 -0.006779935  0.08648514  0.17701586 -0.16510139 -0.09161968##            [,1]        [,2]        [,3]        [,4]
## 1 -0.0001561993 -0.08159755 -0.09543328  0.04583496
## 2 -0.1228465125  0.04293111 -0.08886310 -0.10085030
## 3 -0.0862067605  0.11674374  0.11039017 -0.07840421
## 4  0.1209338141  0.01583609 -0.04168174 -0.50076897
## 5 -0.0293536636  0.02207206 -0.16486648 -0.26215556
## 6  0.0659483699  0.11199333 -0.21709099 -0.10243392
## attr(,"class")
## [1] "rmult"Finite increments interpretation as in Muller et al. (2018) can be obtained using coordinate space outputs.
##                            ilr(SEG)1      ilr(SEG)2      ilr(SEG)3
## (Intercept)            1.24717388192  0.72459957006  2.03502320226
## HOUSEHOLD_EXPENDITURE -0.00011150825 -0.00008293645 -0.00011093018
## GDP                    0.00005900412  0.00004246356  0.00004089033
## GAS_PRICE             -0.02610541082  0.16036864757  0.29475342664
## SCRAPPING_SUBSIDY     -0.19436277707 -0.19093033990 -0.08827046806
##                           ilr(SEG)4
## (Intercept)            2.8974651795
## HOUSEHOLD_EXPENDITURE -0.0002527637
## GDP                    0.0001027078
## GAS_PRICE              0.8389659267
## SCRAPPING_SUBSIDY      0.0279658835Following Muller et al. (2018), using the default contrast matrix below, we can interpret the impact of an additive change in the household expenditure variable on the ratio between segment E and the geometric mean of the other segments.
##         [,1]       [,2]       [,3]       [,4]
## 1 -0.7071068 -0.4082483 -0.2886751 -0.2236068
## 2  0.7071068 -0.4082483 -0.2886751 -0.2236068
## 3  0.0000000  0.8164966 -0.2886751 -0.2236068
## 4  0.0000000  0.0000000  0.8660254 -0.2236068
## 5  0.0000000  0.0000000  0.0000000  0.8944272If the household expenditure, expressed in chain-linked prices of the previous year, increases by \(250\) million euros, approximately one percent of its range, the ratio between segment E and the geometric mean of the other segments is multiplied by \(\exp((-0.0025) * 250/0.89)\sim 0.49,\) hence approximately divided by two. However, it is difficult to describe the simultaneous changes in the ratio between the other segments with this approach.
In this part, we are interested in presenting on a graph the
variations of the fitted shares corresponding to a change scenario of
the variable “HOUSEHOLD_EXPENDITURE”. In this case, since the variable
to be changed is scalar, the scenario of change is just a regular grid
of values defined by a step size, a number of steps and an initial
observation. The first step VariationScenario() creates the
grid for the covariate (\(X\)) and
computes the corresponding fitted values (\(Y\)).
vs_exp2 <- VariationScenario(
  fit_Y_compo,
  Xvar = "HOUSEHOLD_EXPENDITURE",
  obs = 1,
  inc_size = 100,
  n_steps = 150,
  add_opposite = TRUE)plot(x = vs_exp2$HOUSEHOLD_EXPENDITURE, y = vs_exp2$Y[,1],type = "l", col = "red",
     main = "Variation scenario of houshold expenditure for observation 1",
     xlab = "Household expenditure", ylab = "Market share of segment")
lines(x = vs_exp2$HOUSEHOLD_EXPENDITURE, y = vs_exp2$Y[,2],type = "l", col = "blue" )
lines(x = vs_exp2$HOUSEHOLD_EXPENDITURE, y = vs_exp2$Y[,3],type = "l", col = "green")
lines(x = vs_exp2$HOUSEHOLD_EXPENDITURE, y = vs_exp2$Y[,4],type = "l", col = "orange")
lines(x = vs_exp2$HOUSEHOLD_EXPENDITURE, y = vs_exp2$Y[,5],type = "l", col = "black")
legend("topleft",
       legend = paste0("SEG_", LETTERS[1:5]),
       col = c("red", "blue", "green", "orange", "black"),
       lty = 1)The use of semi-elasticities for the interpretation of compositional
covariates is described in Morais and Thomas-Agnan (2021). The only
difference with the previous section is that it involves derivatives of
the logarithm of the \(Y\) components
with respect to \(X\), in contrast with
derivatives of \(Y\) with respect to
the logarithm of the components of \(X\). The Impacts() function
computes these elasticities.
##                              SEG_A         SEG_B         SEG_C          SEG_D
## HOUSEHOLD_EXPENDITURE 0.0001933935 0.00003569699 0.00001296924 -0.00004740457
##                               SEG_E
## HOUSEHOLD_EXPENDITURE -0.0002339346
## attr(,"obs")
## [1] 4The semi-elasticities are then used for evaluating the impact on \(Y\) of an infinitesimal additive change of any of the scalar explanatory variables.
For example, we can measure the change caused by an increase of \(2500\) million euros in
HOUSEHOLD_EXPENDITURE using the function
VariationTable().
fit_Y_compo.VarTab <- VariationTable(
  fit_Y_compo,                      # model output
  Xvar = "HOUSEHOLD_EXPENDITURE",   # variable to be changed
  inc_size = 2500,                  # additive increment of X
  obs = 1)                          # observation index
fit_Y_compo.VarTab##                              SEG_A         SEG_B         SEG_C           SEG_D
## Initial parts          0.065508533 0.35188544895 0.33122939357   0.16366910523
## New parts              0.096976210 0.38218939438 0.34093421621   0.14376119345
## Semi elasticity        0.000192144 0.00003444751 0.00001171976  -0.00004865405
## Variation in %        48.035998361 8.61187796119 2.92994004579 -12.16351231860
## Variation in % points  3.146767770 3.03039454268 0.97048226455  -1.99079117759
##                                SEG_E
## Initial parts           0.0877075195
## New parts               0.0361389855
## Semi elasticity        -0.0002351841
## Variation in %        -58.7960237233
## Variation in % points  -5.1568533997For the first month, an increase of \(2500\) million euros of total
HOUSEHOLD_EXPENDITURE will result in a decrease of 2
percentage points in the “SEG_D” share, and an increase of 3 percentage
points in the “SEG_B” share. We can visualize these changes more
globally with a barplot.
This case is illustrated using an example from Nguyen et. al (2020) studying the impact of some socio-economic variables on the outcome of French departmental elections in 2015.
This dataset contains the vote shares for 3 groups of parties
(left, right, extreme_right) in
different departments of France in the 2015 French departmental election
for 95 departments in France. The election results originate from the
French ‘Ministry of the Interior’. The corresponding socio-economic data
(for 2014) have been downloaded from the INSEE website. The regression
involves the following explanatory variables:
unemp_rate), the proportion of people who own assets
(asset_owner_rate), the proportion of people who pay income
tax (income_taxpayer_rate), the proportion of foreigners
(forgeigner_rate).Age_1839 for people
from 18 to 39 years old, Age_4064 for people from 40 to 64
years old, and Age_65plus for people over 65.Educ_BeforeHighschool for
people with at most some secondary education,
Educ_Highschool for people with at least secondary
education and at most a high school diploma, and
Educ_Higher for people with a university diploma.##       left            right        extreme_right    
##  Min.   :0.1740   Min.   :0.1195   Min.   :0.05563  
##  1st Qu.:0.3108   1st Qu.:0.3296   1st Qu.:0.19956  
##  Median :0.3598   Median :0.3760   Median :0.26102  
##  Mean   :0.3682   Mean   :0.3841   Mean   :0.24777  
##  3rd Qu.:0.4322   3rd Qu.:0.4421   3rd Qu.:0.30400  
##  Max.   :0.7228   Max.   :0.6367   Max.   :0.41485##     Age_1839         Age_4064        Age_65plus    
##  Min.   :0.2270   Min.   :0.3836   Min.   :0.1536  
##  1st Qu.:0.2756   1st Qu.:0.4208   1st Qu.:0.2243  
##  Median :0.3077   Median :0.4315   Median :0.2613  
##  Mean   :0.3140   Mean   :0.4295   Mean   :0.2566  
##  3rd Qu.:0.3486   3rd Qu.:0.4384   3rd Qu.:0.2872  
##  Max.   :0.4320   Max.   :0.4511   Max.   :0.3364##  Educ_BeforeHighschool Educ_Highschool   Educ_Higher    
##  Min.   :0.3484        Min.   :0.1444   Min.   :0.1617  
##  1st Qu.:0.5613        1st Qu.:0.1596   1st Qu.:0.2034  
##  Median :0.5918        Median :0.1677   Median :0.2324  
##  Mean   :0.5889        Mean   :0.1689   Mean   :0.2422  
##  3rd Qu.:0.6349        3rd Qu.:0.1780   3rd Qu.:0.2670  
##  Max.   :0.6867        Max.   :0.2234   Max.   :0.4998##    unemp_rate      asset_owner_rate income_taxpayer_rate forgeigner_rate  
##  Min.   :0.08353   Min.   :0.3998   Min.   :0.4499       Min.   :0.01342  
##  1st Qu.:0.10513   1st Qu.:0.5894   1st Qu.:0.5143       1st Qu.:0.02994  
##  Median :0.11648   Median :0.6189   Median :0.5432       Median :0.04436  
##  Mean   :0.11731   Mean   :0.6158   Mean   :0.5525       Mean   :0.05049  
##  3rd Qu.:0.12423   3rd Qu.:0.6645   3rd Qu.:0.5777       3rd Qu.:0.05886  
##  Max.   :0.16246   Max.   :0.7229   Max.   :0.7452       Max.   :0.22824First, we must ensure that all compositional variables are stored as a matrix.
election$VOTE <- as.matrix(election[,c("left","right","extreme_right")])
election$AGE  <- as.matrix(election[,c("Age_1839","Age_4064","Age_65plus")])
election$EDUC <- as.matrix(election[,c("Educ_BeforeHighschool","Educ_Highschool","Educ_Higher")])The next step consists of applying a log-ratio transformation (alr or
ilr) to the compositional variables and then using the “lm” function to
perform regression in coordinates space. In this example, the response
variable, VOTE, is transformed by an ilr, the variable
Age is transformed by an alr, and EDUC is
transformed by an ilr.
fit_YX_compo <- lmCoDa(
  ilr(VOTE) ~
  alr(AGE) + unemp_rate + asset_owner_rate +
  ilr(EDUC) + income_taxpayer_rate + forgeigner_rate,
  data = election)With the coef() function we can access the estimated
coefficients in coordinate space, as well as in the simplex:
##                       ilr(VOTE)1 ilr(VOTE)2
## (Intercept)          -0.07020983 -6.9109264
## alr(AGE)Age_1839     -0.10463431  0.5592385
## alr(AGE)Age_4064      0.23336593 -0.2278400
## unemp_rate           -8.16701488 17.2562747
## asset_owner_rate     -2.96545832  2.8971115
## ilr(EDUC)1           -1.21486657 -2.2816886
## ilr(EDUC)2           -0.76776003  0.9077666
## income_taxpayer_rate  2.91875389  1.7559045
## forgeigner_rate      -1.45607088 -0.5496067##                                   left                right extreme_right
## (Intercept)            0.5247472819482  0.47514740845404052  0.0001053096
## Age_1839              -0.1543205229982 -0.30229578360147880  0.4566163066
## Age_4064              -0.0719993474717  0.25802991714544993 -0.1860305697
## Age_65plus             0.2263198704699  0.04426586645602888 -0.2705857369
## unemp_rate             0.0000002135099  0.00000000000205759  0.9999997865
## asset_owner_rate       0.1892554610820  0.00285566966007837  0.8078888693
## Educ_BeforeHighschool -1.3364388553996  0.32169418104455549  1.0147446744
## Educ_Highschool        1.1957612695548  0.42416115809884758 -1.6199224277
## Educ_Higher            0.1406775858448 -0.74585533914340330  0.6051777533
## income_taxpayer_rate   0.0076516015515  0.47468732197433922  0.5176610765
## forgeigner_rate        0.7635086355050  0.09739055674009525  0.1391008078Standard errors and confidence intervals are provided for the clr of
parameters by the function confint(). The following example
is for the case of the compositional explanatory variable
AGE.
##               Y          X         EST        SD      2.5 %     97.5 %
## 1          left   Age_1839 -0.15432052 0.4909885 -1.1303733 0.82173229
## 2         right   Age_1839 -0.30229578 0.3747233 -1.0472210 0.44262946
## 3 extreme_right   Age_1839  0.45661631 0.3953292 -0.3292722 1.24250484
## 4          left   Age_4064 -0.07199935 0.6721951 -1.4082790 1.26428030
## 5         right   Age_4064  0.25802992 0.5130205 -0.7618211 1.27788092
## 6 extreme_right   Age_4064 -0.18603057 0.5412314 -1.2619630 0.88990183
## 7          left Age_65plus  0.22631987 0.1919484 -0.1552609 0.60790069
## 8         right Age_65plus  0.04426587 0.1464954 -0.2469573 0.33548902
## 9 extreme_right Age_65plus -0.27058574 0.1545511 -0.5778232 0.03665172When providing the argument y_ref the function returns
the same information for the differences of clr parameter values, which
coincide with differences in elasticities (Dargel and Thomas-Agnan
2024).
##   Y_ref             Y          X       DIFF        SD      2.5 %    97.5 %
## 1  left          left   Age_1839  0.0000000 0.0000000  0.0000000 0.0000000
## 2  left         right   Age_1839 -0.1479753 0.7789026 -1.6963824 1.4004319
## 3  left extreme_right   Age_1839  0.6109368 0.8088834 -0.9970701 2.2189437
## 4  left          left   Age_4064  0.0000000 0.0000000  0.0000000 0.0000000
## 5  left         right   Age_4064  0.3300293 1.0663682 -1.7898406 2.4498991
## 6  left extreme_right   Age_4064 -0.1140312 1.1074138 -2.3154970 2.0874346
## 7  left          left Age_65plus  0.0000000 0.0000000  0.0000000 0.0000000
## 8  left         right Age_65plus -0.1820540 0.3045064 -0.7873926 0.4232846
## 9  left extreme_right Age_65plus -0.4969056 0.3162271 -1.1255443 0.1317331The following example is for the case of the scalar explanatory
variable unemp_rate.
##               Y          X        EST       SD      2.5 %    97.5 %
## 1          left unemp_rate  -1.269893 2.465914  -6.171968  3.632182
## 2         right unemp_rate -12.819796 1.881990 -16.561068 -9.078524
## 3 extreme_right unemp_rate  14.089689 1.985480  10.142685 18.036693In this case the difference in clr parameters coincides with a semi-elasticities difference.
##   Y_ref             Y          X      DIFF       SD      2.5 %    97.5 %
## 1  left          left unemp_rate   0.00000 0.000000   0.000000  0.000000
## 2  left         right unemp_rate -11.54990 3.911918 -19.326539 -3.773267
## 3  left extreme_right unemp_rate  15.35958 4.062492   7.283616 23.435549Fitted values and residuals, both in the simplex space, can be accessed directly using fitted, and residuals functions as shown previously.
Finite increments interpretation as in Muller et al. (2018) can be
obtained using coordinate space outputs, but this time we need to
consider two contrast matrices, one for the response and one for the
explanatory composition. In the present model, the VOTE and
the EDUC compositions are both ilr transformed with the
default contrast matrix.
##         [,1]       [,2]
## 1 -0.7071068 -0.4082483
## 2  0.7071068 -0.4082483
## 3  0.0000000  0.8164966This means we can interpret the element in the last row and last
column of the corresponding parameter matrix that is expressed using the
ilr spaces of both the response VOTE and the “EDUC”
variables (whose contrast matrices coincide in our example).
##            ilr(VOTE)1 ilr(VOTE)2
## ilr(EDUC)1  -1.214867 -2.2816886
## ilr(EDUC)2  -0.767760  0.9077666The value of about \(0.90777\)
indicates that an additive unit increase in the second ilr coordinate of
the explanatory EDUC composition leads to an additive
increase of the second ilr coordinate of the response VOTE
composition. Given the form of the contrast matrix, these ilr
coordinates can be interpreted as principal balances of the last
component against the first two in both compositions. The above change
in education corresponds to a move in the direction of the high
education vertex in the simplex. Therefore, a multiplication of the
ratio between the higher education level to the geometric mean of the
other two by \(\exp(1/081) \sim 3.43\)
induces a multiplication of the ratio of the vote share for the extreme
right party to the geometric mean of the other two parties by \(\exp(0.91/081)\sim 3.07.\) If one wishes to
evaluate the impact of this change on other ratios of \(Y\) it would mean changing the contrast
matrix of the response. Similarly, if one wishes to evaluate the impact
of changing the education composition in the direction of the other two
vertices, it would mean changing the \(X\) contrast matrix.
As in the case of scalar on composition regression, to evaluate the impact of a given compositional covariate, we create scenarios of changes in its simplex space described by a linear equation \(x(h)=x(0)\oplus h\odot u,\) where \(x(0)\) is the initial point, \(h\) is the signed-intensity of change and \(u\) is a vector of the simplex defining the direction of change. We then compute the predicted value \(y(h)\) corresponding to \(x(h)\), keeping the other covariates at their value for the initial point.
To present the variations of \(Y\)
induced by a change scenario of \(X\),
we propose two plots. The first one is a graph of the shares of \(Y\) as a function of one of the shares of
\(X\) and the second one presents the
evolution of the other shares of \(X\).
In the following illustration, we choose the direction of change joining
the initial point (in this case, observation 1, Landes department) to
the vertex Age_1839.
VS_election<-VariationScenario(
  fit_YX_compo,
  Xvar = "AGE",
  Xdir = "Age_1839",
  n_steps = 100,
  obs = 1)
opar <- par(mfrow = c(2,1), mar = c(5,4,1,2))
plot(x = VS_election$X[,1],  y = VS_election$Y[,1], col = "Orange",
     xlim = c(0,1), ylim = c(0,1), xlab = "% Age_1839", ylab = "% VOTE")
lines(x = VS_election$X[,1],  y = VS_election$Y[,1], lwd = 1.5, col = "orange")
lines(x = VS_election$X[,1],  y = VS_election$Y[,2], lwd = 1.5, col = "darkblue")
lines(x = VS_election$X[,1],  y = VS_election$Y[,3], lwd = 1.5, col = "red")
legend("top",
       legend = c("left", "right", "extreme_right"),
       col = c("orange", "darkblue", "red"),
       lty = 1)
plot(x = VS_election$X[,1],  y = VS_election$X[,1], col = "Orange",
     xlim = c(0,1), ylim = c(0,1), xlab = "% Age_1839", ylab = "% AGE")
lines(x = VS_election$X[,1],  y = VS_election$X[,1], lwd = 1.5, col = "orange")
lines(x = VS_election$X[,1],  y = VS_election$X[,2], lwd = 1.5, col = "darkblue")
lines(x = VS_election$X[,1],  y = VS_election$X[,3], lwd = 1.5, col = "red")
legend("top",
       legend = c("18-39","40-64", "65_plus"),
       col = c("orange", "darkblue", "red"),
       lty = 1)The vertical line corresponds to the initial department of Landes,
and we can read the values of its fitted vote shares on the first graph
and the values of its age characteristics on the second.
For the specified direction, the bottom graph shows that an increase in
the share of voters aged between \(18\)
and \(39\) years corresponds to a
linear decrease in the other shares. On the top graph, we see that an
increase in the share of voters aged between \(18\) and \(39\) years induces an increase in the share
of the extreme right, a decrease in the share of the right parties, and
a simultaneous decrease in the left party’s share.
The function Impacts() computes elasticities (or
semi-elasticities, depending on the nature of the considered covariate)
of covariates for a given observation. This function takes a “lmCoDa”
object as input, along with the covariate’s name and the index of the
observation.
##                  left      right extreme_right
## Age_1839   -0.1279812 -0.2759565     0.4829556
## Age_4064   -0.1051949  0.2248344    -0.2192261
## Age_65plus  0.2331761  0.0511221    -0.2637295
## attr(,"obs")
## [1] 3The semi-elasticities are then used for evaluating the impact on \(Y\) of an infinitesimal additive change of any of the scalar explanatory variables.
For example, we can measure the change caused by increasing the
Higher share of the EDUC variable by 5%,
assuming that the ratio between the other two components remains
constant, using the function VariationTable(). Assuming
that the voting population for the first department is \(100\;000\) individuals (we do not have this
information in the dataset). The total of the response must be specified
by the user in case he wants to know the variation in units.
##                               left        right extreme_right
## Initial parts           0.34161097    0.4756646     0.1827245
## New parts               0.34061774    0.4693200     0.1900622
## Elasticity             -0.05051647   -0.2317484     0.6977253
## Variation in %         -0.29074696   -1.3338252     4.0157495
## Variation in % points  -0.09932235   -0.6344534     0.7337757
## Variation in units    -99.32234990 -634.4533879   733.7757378A \(5\%\) increase in the direction
of Age_1839 will result in a \(0.099\) percentage points decrease in the
left share, a \(0.63\)
percentage points decrease in the right proportion, and a
\(0.73\) percentage points increase in
the extreme_right share. In units these changes correspond
respectively to a decrease of \(99\)
votes for the left parties, a decrease of \(634\) votes for the right
parties and an increase of \(733\)
votes for the extreme_right share. It is easy to check that
the sum of these changes is zero.
For a general direction \((0.45,0.2,0.35)\), with an increment size of \(0.1\), we get the following results.
VT <- VariationTable(
  fit_YX_compo,
  Xvar = "AGE",
  Xdir = c(0.45,0.2,0.35) ,
  inc_size=0.1,
  Ytotal = 100000)
VT##                              left         right extreme_right
## Initial parts           0.3416110     0.4756646     0.1827245
## New parts               0.3454412     0.4630204     0.1915383
## Elasticity              0.1121240    -0.2658202     0.4823571
## Variation in %          1.1212404    -2.6582020     4.8235706
## Variation in % points   0.3830280    -1.2644125     0.8813844
## Variation in units    383.0280319 -1264.4124527   881.3844208Note that this direction vector will be normalized before being used within the function, and will later be given as output in the table’s attributes.
## $names
## [1] "left"          "right"         "extreme_right"
## 
## $class
## [1] "data.frame"
## 
## $row.names
## [1] "Initial parts"         "New parts"             "Elasticity"           
## [4] "Variation in %"        "Variation in % points" "Variation in units"   
## 
## $`X(0)`
##    Age_1839    Age_4064  Age_65plus 
## "0.2906800" "0.4380762" "0.2712438" 
## attr(,"class")
## [1] "acomp"
## 
## $`X(h)`
## [1] 0.3116451 0.4091594 0.2791955
## 
## $Xdir
## [1] 0.5254862 0.1320255 0.3424884
## 
## $inc_size
## [1] 0.1
## 
## $inc_rates
## [1]  0.07212431 -0.06600860  0.02931571Coenders, G. and Pawlowsky-Glahn, V. (2020). “On interpretations of tests and effect sizes in regression models with a compositional predictor.” SORT-Statistics and Operations Research Transactions, pages 201–220.
Dargel, L., and C. Thomas-Agnan (2024). “Pairwise share-ratio interpretations of compositional regression models.” Computational Statistics & Data Analysis, Volume 195, pages 107945
Morais, J., and C. Thomas-Agnan. (2021) “Impact of covariates in compositional models and simplicial derivatives.” Austrian Journal of Statistics 50.2 (2021): 1-15.
Muller I, Hron K, Fiserova E, Smahaj J, Cakirpaloglu P, Vancakova J (2018). “Interpretation of Compositional Regression with Application to Time Budget Analysis.” Austrian Journal of Statistics, 47(2).