1. Home
  2. BACKTRANSFORM procedure

BACKTRANSFORM procedure

Calculates back-transformed means with approximate standard errors and confidence intervals (V.M. Cave).

Options

PRINT = string tokens Controls printed output (description, means, backmeans); default desc, back
PLOT = string tokens The confidence intervals of the back-transformed means to plot (backtransformed, approximate, both); default * i.e. none
TRANSFORMATION = string tokens Transformation (identity, logarithm, log10, logit, squareroot, reciprocal, power, probit, complementaryloglog,logratio, angular, arcsinesquareoot, calculated); default iden (i.e. no transformation)
CLOG = scalar Constant c for the logarithm and log10 transformations, in form log(mean+c); default 0
EXPONENT = scalar Exponent for power transformation; default -2
KLOGRATIO = scalar Parameter k for logratio transformation, in form log(mean/(mean+k)); default 1
BACKTRANSFORMATION = expression Expression, formed using argument Y, that defines the inverse of the transformation; must be specified when TRANSFORMATION = calculated
DERIVATIVE = expression Expression, formed using argument Y, that defines the first derivative of the transformation; must be specified when TRANSFORMATION = calculated
CIPROBABILITY = scalar Probability for the confidence intervals; default 0.95
DIRECTION = string tokens Order in which the back-transformed means are plotted (ordinal, ascending, descending); default ordi
USEPENS = string tokens Whether to use the current pen definitions for plotting; (yes, no); default no
WINDOW = scalar Window to use for plot; default 3

Parameters

MEANS = tables, variates or scalars Supplies the transformed mean(s)
SEMEANS = tables, variates or scalars Supplies the standard error(s) of the transformed mean(s)
DF = scalars Degrees of freedom to construct the confidence intervals; default *
DECIMALS = scalars Number of decimal places for printing; default *
BACKTRANSFORMEDMEANS = tables, variates or scalars Saves the back-transformed means
SEBACKTRANSFORMEDMEANS = tables, variates or scalars Saves the approximate standard errors for the back-transformed means
CIAPPROXIMATE = pointers Saves the approximate confidence intervals for the back-transformed means
CIBACKTRANSFORMED = pointers Saves the back-transformed confidence intervals for the back-transformed means
TITLE = texts Title for plot; default * i.e. none
YTITLE = texts Title for y-axis; default * i.e. none
XTITLE = texts Title for x-axis; default * i.e. formed automatically

Description

BACKTRANSFORM calculates back-transformed means, with approximate standard errors and confidence intervals. The means and corresponding standard errors, for back-transforming, are supplied using the MEANS and SEMEANS parameters, respectively, as either tables, variates or scalars. If MEANS supplies a table or variate, SEMEANS can be either of the same type or a scalar, whereas if MEANS supplies a scalar, SEMEANS must be a scalar.

The degrees of freedom, used to construct the confidence intervals, can be set using the DF parameter. If these are not supplied, z-scores are used to form the confidence intervals. The probability for the confidence intervals is specified by the CIPROBABILITY option; the default 0.95 gives 95% confidence intervals.

The function that was used to transform the data prior to analysis is specified using the TRANSFORMATION option; the default takes the identity link (i.e. no transformation). The natural logarithm, log10, logit, square root, reciprocal, power, probit, complementary log-log, log-ratio and angular (or arcsine-square root) functions are provided directly by the TRANSFORMATION option. The angular and arcsinesquareroot transformations are synonyms, and the transformed data are assumed to be in radians.

You can also define your own transformation by setting TRANSFORMATION = calculated and providing expressions to calculate the inverse and first derivative of the transformation, using the BACKTRANSFORMATION and DERIVATIVE options, respectively. The calculations are specified in terms of the argument Y. Thus, for example, the logarithm transformation could be specified by setting options

BACKTRANSFORMATION=!E(exp(Y))

and

DERIVATIVE=!E(1/Y).

The CLOG option sets the constant c used by the logarithm and log10 transformations (default 0), the EXPONENT option sets the exponent used by the power transformation (default -2) and the KLOGRATIO option sets the parameter k used by the logratio transformation (default 1).

Back-transformation from the logit, probit, angular or arcsinesquareroot scale returns proportions (rather than percentages).

The PLOT option allows you to request plots of the results, using settings:

    backtransformed for back-transformed means and back-transformed confidence intervals (i.e. confidence intervals that maintain exactly the percent coverage on the transformed scale),
    approximate for back-transformed means with approximate confidence intervals (i.e. derived from the approximate standard errors), and
    both for back-transformed means with both types of confidence interval.

By default no plots are produced.

The TITLE, YTITLE and XTITLE parameters can supply an overall title, a y-axis title and a x-axis title for the plot, respectively. By default, neither an overall title nor a y-axis title is displayed. The default for the x-axis, when MEANS supplies the transformed means as a table, is to use the identifiers of the table’s classifying factors to form a title for the x-axis. If MEANS supplies a variate or scalar, the default is not to display an x-axis title. To omit the x-axis title, you can supply a blank string i.e.

XTITLE=' '

By default, the pen attributes used for plotting are determined automatically within the procedure. However, you can set USEPENS to yes, to request that the current COLOURS, CFILL, SYMBOLS, SMSYMBOL and THICKNESS pen definitions of pens 1 and 2 are used. Pen 1 controls the colour, symbol, symbol size and line thickness used to plot the back-transformed means with back-transformed confidence intervals, whereas Pen 2 controls these attributes when the back-transformed means are plotted with approximate confidence intervals. The WINDOW option specifies the window used for plotting; default 3.

By default, the back-transformed means are plotted in order of their ordinal level, however you can use the DIRECTION option to request that they are plotted in ascending or descending numerical value instead.

The BACKTRANSFORMEDMEANS, SEBACKTRANSFORMEDMEANS parameters can save back-transformed means, approximate standard errors, in data structures of the same type as MEANS. The CIAPPROXIMATE and CIBACKTRANSFORMED parameters can save pointers of approximate confidence intervals and back-transformed confidence intervals, respectively. The pointers contain two data structures elements, of the same type as MEANS, storing the lower and the upper confidence limits, respectively.

Printed output is controlled by the PRINT option, with settings:

    description provides a description of the output,
    means prints the transformed means, with their standard errors and confidence intervals, and
    backmeans prints the back-transformed means, with their approximate standard errors and confidence intervals.

You can set the number of decimals places to appear in the printed output, using the DECIMALS parameter.

Options: PRINT, PLOT, TRANSFORMATION, CLOG, EXPONENT, KLOGRATIO, BACKTRANSFORMATION, DERIVATIVE, CIPROBABILITY, DIRECTION, USEPENS, WINDOW.

Parameters: MEANS, SEMEANS, DF, DECIMALS, BACKTRANSFORMEDMEANS, SEBACKTRANSFORMEDMEANS, CIPAPPROXIMATE, CIBACKTRANSFORMED, TITLE, YTITLE, XTITLE.

Method

BACKTRANSFORM uses a first-order Taylor series expansion, to obtain approximate standard errors for the back-transformed means. The methodology is described in Jørgensen & Pedersen (1998). In brief, let û denote the estimated mean on the transformed scale, seu its standard error, g() the transformation function and g′() the first derivative of the transformation function. On the back-transformed scale the estimated mean (ӯ) and standard error (seӯ) are approximated by g-1(û) and seu / mod(g′(ӯ)), respectively.

A back-transformed confidence interval is given by

(g-1( û – t × seu), g-1( û + t × seu))

where t is the upper (1-CIPROBABILITY)/2 critical value for the t distribution. Note that, when the degrees of freedom have not been set using the DF parameter, the z-score (i.e. the Normal distribution) is used to construct the confidence interval instead of the t-value. For the logratio and reciprocal transformations, when the estimated mean and confidence
limit on the transformed scale lie on different sides of the singularity in the inverse function
of the transformation, the back-transformed confidence limit is set to a missing value.

An approximate confidence interval

(ӯ – t × seӯ, ӯ + t × seӯ)

is also provided. However, this should be used only to evaluate the validity of the approximate standard error. If the back-transformed and approximate confidence intervals differ greatly, the approximate standard error is inadequate.

Action with RESTRICT

BACKTRANSFORM will work when MEANS and/or SEMEANS supplies a restricted variate; however, if both variates are restricted they must be restricted in the same way. Furthermore, their unrestricted length must be the same.

References

Jørgensen, E., & Pedersen, A.R. (1998). How to obtain those nasty standard errors from transformed data – and why they shouldn’t be used. Biometry Research Unit, Danish Institute of Agricultural Sciences.

See also

Directives: AKEEP, PREDICT, VPREDICT.

Procedures: AFMEANS, LSIPLOT.

Example

CAPTION    'BACKTRANSFORM examples',\
           'Pup rat data from Weil (1970, Food and Cosmetics Toxicology)';\
           STYLE=meta,plain
FACTOR     [NVALUES=64; LEVELS=32; VALUES=2(1...32)] Litter
&          [LEVELS=2; LABELS=!t('4','21'); VALUES=(1,2)32] Time
&          [LEVELS=2; LABELS=!t('control','treated'); VALUES=32(1,2)] Diet
VARIATE    [NVALUES=64] Pups
READ       [PRINT=*] Pups
13 13 12 12  9  9  9  9  8  8  8  8  13 12 12 11  
10  9 10  9  9  8 13 11  5  4  7  5  10  7 10  7  
12 12 11 11 10 10  9  9 11 10 10  9  10  9  9  8  
 9  8  5  4  9  7  7  4 10  5  6  3  10  3  7  0 :
CAPTION    'ANOVA Example: Means and standard errors obtained from AKEEP';\
           STYLE=meta
BLOCKS     Litter
TREATMENTS Diet*Time
ANOVA      [PRINT=aov] log(Pups+1)
AKEEP      TERMS=Diet.Time; MEANS=meansA; SEMEANS=semsA; DFCEFFECTS=dfA
BACKTRANSFORM [TRANSFORMATION=log; CLOG=1; PLOT=both]\ 
           meansA; SEMEANS=semsA; DF=dfA
CAPTION    'GLM Example: Means and standard errors obtained from PREDICT';\
           STYLE=meta
MODEL      [DISTRIBUTION=Poisson; LINK=log] Pups
FIT        [PRINT=summary] Diet*Time
RKEEP      DF=dfB
PREDICT    [PRINT=*; PREDICTIONS=meansB; SE=semsB; BACKTRANSFORM=none]\
           Diet,Time
BACKTRANSFORM [TRANSFORMATION=log; PLOT=both] \
           meansB; SEMEANS=semsB; DF=dfB
CAPTION    'REML Example: Means and standard errors obtained from VPREDICT';\
           STYLE=meta
VCOMPONENT [FIXED=Diet*Time] RANDOM=Litter
REML       log(Pups+1)
VKEEP      [WMETHOD=add] TERMS=Diet.Time; DDF=dfC   
VPREDICT   [PRINT=*; PREDICTIONS=meansC; SE=semsC] Diet,Time
BACKTRANSFORM [TRANSFORMATION=log; CLOG=1; PLOT=both]\
           meansC; SEMEANS=semsC; DF=dfC
CAPTION    'GLMM Example: Means and standard errors obtained from VPREDICT';\
           STYLE=meta
GLMM       [PRINT=model,components; DISTRIBUTION=Poisson; LINK=log;\
           FIXED=Diet*Time; RANDOM=Litter; DISPERSION=*] Pups
VKEEP      [WMETHOD=add] TERMS=Diet.Time; DDF=dfD   
VPREDICT   [PRINT=*; PREDICTIONS=meansD; SE=semsD] Diet,Time
BACKTRANSFORM [TRANSFORMATION=log; PLOT=both]\
           meansD; SEMEANS=semsD; DF=dfD
Updated on May 26, 2022

Was this article helpful?