{smcl}
{* March 2007}{...}
{hline}
help for {hi:smfit}{right:Stephen P. Jenkins (March 2007)}
{hline}
{title:Fitting a Singh-Maddala distribution by ML to unit record data}
{p 8 17 2}{cmd:smfit} {it:var} [{it:weight}] [{cmd:if} {it:exp}]
[{cmd:in} {it:range}] [{cmd:,}
{cmdab:a:var(}{it:varlist1}{cmd:)} {cmdab:b:var(}{it:varlist2}{cmd:)}
{cmdab:q:var(}{it:varlist3}{cmd:)} {cmdab:abq(}{it:varlist}{cmd:)}
{cmdab:st:ats} {cmdab:f:rom(}{it:string}{cmd:)} {cmdab:poor:frac(}{it:#}{cmd:)}
{cmdab:cdf(}{it:cdfname}{cmd:)} {cmdab:pdf(}{it:pdfname}{cmd:)}
{cmdab:r:obust} {cmdab:cl:uster(}{it:varname}{cmd:)} {cmdab:svy:}
{cmdab:l:evel(}{it:#}{cmd:)} {it:maximize_options} {it:svy_options} ]
{p 4 4 2}{cmd:by} {it:...} {cmd::} may be used with {cmd:smfit}; see help
{help by}.
{p 4 4 2}{cmd:pweight}s, {cmd:aweight}s, {cmd:fweight}s, and {cmd:iweight}s
are allowed; see help {help weights}. To use {cmd:pweight}s, you must first
{cmd:svyset} your data and then use the {cmd:svy} option.
{title:Description}
{p 4 4 2}
{cmd:smfit} fits by ML the 3 parameter Singh-Maddala (1976) distribution to
sample observations on a random variable {it:var}. Unit record data
are assumed (rather than grouped data). Otherwise known as
the Burr Type 12 distribution, the Singh-Maddala distribution has been
shown to provide a good fit to empirical income data relative to other
parametric functional forms: see e.g. McDonald (1984). It is closely
related to the Dagum (Burr Type 3) distribution (Dagum, 1977,1980).
Both are special cases of the Generalized Beta of the Second Kind
distribution (see {help gb2fit}). For a comprehensive review of these
and other related distributions, see Kleiber and Kotz (2003).
For derivation of Lorenz orderings of pairs of income distributions
in terms of their Singh-Maddala parameters, see Wifling and
Kraemer (1993) and Kleiber (1996). Of course the Singh-Maddala
distribution might be suitable for describing any skewed variable,
not only income.
{p 4 4 2}
The likelihood function for a sample of observations on {it:var} is specified
as the product of the densities for each observation (weighted where relevant), and
is maximized using {cmd:ml model lf}.
{title:Options}
{p 4 8 2}{cmd:avar(}{it:varlist1}{cmd:)}, {cmd:bvar(}{it:varlist2}{cmd:)}, and
{cmd:qvar(}{it:varlist3}{cmd:)} allow the user to specify each
parameter as a function of the covariates specified in the respective
variable list. A constant term is always included in each equation.
{p 4 8 2}{cmd:abq(}{it:varlist}{cmd:)} can be used instead of the previous option
if the same covariates are to appear in each parameter equation.
{p 4 8 2}{cmd:from(}{it:string}{cmd:)} specifies initial values for the
parameters, and is likely to be used only rarely. You can specify the initial
values in one of three ways: the name of a vector containing the initial values
(e.g., from(b0) where b0 is a properly labeled vector); by specifying coefficient
names with the values (e.g., from(a:_cons=1 b:_cons=5 q:_cons = .16);
or by specifying an ordered list of values (e.g., from(1 5 0 .16, copy)).
Poor values in from() may lead to convergence problems. For more details,
including the use of copy and skip, see {help:maximize}.
{p 8 8 2}If covariates are specified, the next four options are not available.
Use {help smpred} to generate statistics at particular values of the
covariates, or {cmd:nlcom}. {cmd:predict} can be used to generate the
observation-specific parameters corresponding to the covariate values of
each sample observation: see Examples below.
{p 4 8 2}{cmd:stats} displays selected distributional statistics implied by the
Singh-Maddala parameter estimates: quantiles, cumulative
shares of total {it:var} at quantiles (i.e. the Lorenz curve
ordinates), the mode, mean, standard deviation, variance, half the
coefficient of variation squared, Gini coefficient, and
quantile ratios p90/p10, p75/p25.
{p 4 8 2}{cmd:poorfrac(}{it:#}{cmd:)} displays the estimated proportion with values of {it:var}
less than the cut-off specified by {it:#}. This option may be specified when replaying
results.
{p 4 8 2}{cmd:cdf(}{it:cdfname}{cmd:)} creates a new variable {it:cdfname} containing the
estimated Singh-Maddala c.d.f. value F(x) for each x.
{p 4 8 2}{cmd:pdf(}{it:pdfname}{cmd:)} creates a new variable {it:pdfname} containing the
estimated Singh-Maddala p.d.f. value f(x) for each x.
{p 4 8 2}{cmd:robust} specifies that the Huber/White/sandwich estimator of
variance is to be used in place of the traditional calculation; see
{hi:[U] 23.14 Obtaining robust variance estimates}. {cmd:robust} combined
with {cmd:cluster()} allows observations which are not independent within
cluster (although they must be independent between clusters). If you
specify {help pweight}s, {cmd:robust} is implied.
{p 4 8 2}{cmd:cluster(}{it:varname}{cmd:)} specifies that the observations are
independent across groups (clusters) but not necessarily within groups.
{it:varname} specifies to which group each observation belongs; e.g.,
{cmd:cluster(personid)} in data with repeated observations on individuals.
See {hi:[U] 23.14 Obtaining robust variance estimates}. {cmd:cluster()} can be
used with {help pweight}s to produce estimates for unstratified
cluster-sampled data. Specifying {cmd:cluster()} implies {cmd:robust}.
{p 4 8 2}{cmd:svy} indicates that {cmd:ml} is to pick up the {cmd:svy} settings
set by {cmd:svyset} and use the robust variance estimator. Thus, this option
requires the data to be {cmd:svyset}; see help {help svyset}. {cmd:svy} may not be
combined with weights or the {cmd:strata()}, {cmd:psu()}, {cmd:fpc()}, or
{cmd:cluster()} options.
{p 4 8 2}{cmd:level(}{it:#}{cmd:)} specifies the confidence level, in percent,
for the confidence intervals of the coefficients; see help {help level}.
{p 4 8 2}{cmd:nolog} suppresses the iteration log.
{p 4 8 2}{it:maximize_options} control the maximization process.. The options
available are those shown by {help maximize}, with the exception of {cmd:from()}.
If you are seeing many "(not concave)" messages in the iteration
log, using the {cmd:difficult} or {cmd:technique} options may help convergence.
{p 4 8 2}{it:svy_options} specify the options used together with the {cmd:svy} option.
{title:Saved results}
{p 4 4 2}In addition to the usual results saved after {cmd:ml}, {cmd:smfit} also
saves the following, if there are no covariates have been specified and
the relevant options used:
{p 4 4 2}{cmd:e(ba)}, {cmd:e(bb)}, and {cmd:e(bq)} are the estimated Singh-Maddala
parameters.
{p 4 4 2}{cmd:e(cdfvar)} and {cmd:e(pdfvar)} are the variable names specified for the
c.d.f. and the p.d.f.
{p 4 4 2}
{cmd:e(mode)}, {cmd:e(mean)}, {cmd:e(var)}, {cmd:e(sd)}, {cmd:e(i2)}, and {cmd:e(gini)}
are the estimated mode, mean, variance, standard deviation, half coefficient of
variation squared, Gini coefficient. {cmd:e(pX)}, and {cmd:e(LpX)} are the
quantiles, and Lorenz ordinates, where X = {1, 5, 10, 20, 25, 30, 40, 50,
60, 70, 75, 80, 90, 95, 99}.
{p 4 4 2}The following results are saved regardless of whether covariates have been
specified or not.
{p 4 4 2}{cmd:e(b_a)}, {cmd:e(b_b)}, and {cmd:e(b_q)} are row vectors containing the
parameter estimates from each equation.
{p 4 4 2}{cmd:e(length_b_a)}, {cmd:e(length_b_b)}, and {cmd:e(length_b_q)} contain
the lengths of these vectors. If no covariates have been specified in an equation,
the corresponding vector has length equal to 1 (the constant term);
otherwise, the length is one plus the number of covariates.
{title:Formulae}
{p 4 4 2}
The Singh-Maddala distribution has distribution function (c.d.f.)
{p 8 8 2}
F(x) = 1 - { 1/[ 1 + (x/b)^a ]^q }
{p 4 4 2}
where a, b, q, are parameters, each positive, for random variable x > 0.
Parameters a and q are the key distributional 'shape' parameters; b is a scale parameter.
{p 4 4 2}
Letting z = 1 + (x/b)^a, then F(x) = 1 - [1/(z^q)], and the probability
density function (p.d.f.) is
{p 8 8 2}
f(x) = (aq/b)*{z^-(q+1)}*[(x/b)^(a-1)].
{p 4 4 2}
The likelihood function for a sample of observations on {it:var} is specified
as the product of the densities for each observation (weighted where relevant), and
is maximized using {cmd:ml model lf}.
{p 4 4 2}
The formulae used to derive the distributional summary statistics
presented (optionally) are as follows. The r-th moment about the origin
is given by
{p 8 8 2}
b^r*B(1+r/a,q-r/a)/B(1,q)
{p 4 4 2}
where B(u,v) is the Beta function = G(u).G(v)/G(u+v) and G(.) is the
gamma function [exp({cmd:lngamma}(.)], which by substitution and using G(1) = 1,
implies the moments can be written
{p 8 8 2}
b^r*G(1+r/a)*G(q-r/a)/G(q)
{p 4 4 2}
and hence
{p 8 8 2}
mean = b*G(1+1/a)*G(q-1/a)/G(q)
{p 8 8 2}
variance = b*b*G(1+2/a)*G(q-2/a)/G(q) - (mean^2)
{p 4 4 2}
from which the standard deviation and half the squared coefficient of
variation can be derived. The mode is
{p 4 4 2}
mode = b*((a-1)/(aq+1))^(1/a) if a > 1, and 0 otherwise.
{p 4 4 2}
The quantiles are derived by inverting the distribution function:
{p 8 8 2}
x_s = b*((1-s)^(-1/q) - 1)^(1/a) for each s = F(x_s).
{p 4 4 2}
The Gini coefficient of inequality is given by
{p 8 8 2}
1-Gini = G(q)*G(2q - 1/a) / { G(q-1/a)*G(2q) }.
{p 4 4 2}
The Lorenz curve ordinates at each s = F(x_s) use the incomplete Beta function:
{p 8 8 2}
L(s) = {cmd:ibeta}(1+1/a, q- 1/a, 1-(1-s)^(1/q) ).
{title:Examples}
{p 4 8 2}{inp:. smfit x [w=wgt] }
{p 4 8 2}{inp:. smfit }
{p 4 8 2}{inp:. smfit, stats poorfrac(100) }
{p 4 8 2}{inp:. smfit x, a(age sex) b(age sex) q(age sex) }
{p 4 8 2}{inp:. smfit x, abq(age sex) }
{p 4 8 2}{inp:. predict double a_i, eq(a) xb }
{p 4 8 2}{inp:. predict double b_i, eq(b) xb }
{p 4 8 2}{inp:. predict double q_i, eq(q) xb }
{p 4 4 2}See also the examples in the presentation by
{browse "http://www.stata.com/meeting/2german/Jenkins.pdf":Jenkins (2004)}.
{title:Author}
{p 4 4 2}Stephen P. Jenkins , Institute for Social
and Economic Research, University of Essex, Colchester CO4 3SQ, U.K.
{title:Acknowledgements}
{p 4 4 2}N.J. Cox made numerous helpful comments and suggestions, and also wrote
programs for distributional diagnostic plots ({help qsm}, {help psm}).
{title:References}
{p 4 8 2}Dagum, C. (1977). A new model of personal income distribution:
specification and estimation. {it:Economie Appliqu{c e'}e} 30: 413-437.
{p 4 8 2}Dagum, C. (1980). The generation and distribution of income, the
Lorenz curve and the Gini ratio. {it:Economie Appliqu{c e'}e} 33: 327-367.
{p 4 8 2}Jenkins, S.P. (2004). Fitting functional forms to distributions, using {cmd:ml}. Presentation
at Second German Stata Users Group Meeting, Berlin. {browse "http://www.stata.com/meeting/2german/Jenkins.pdf"}
{p 4 8 2}Kleiber, C. (1996). Dagum vs. Singh-Maddala income distributions.
{it: Economics Letters} 53: 265-268.
{p 4 8 2}Kleiber, C. and Kotz, S. (2003).
{it:Statistical Size Distributions in Economics and Actuarial Sciences}. Hoboken, NJ: John Wiley.
{p 4 8 2}McDonald, J.B. (1984). Some generalized functions for the size
distribution of income. {it:Econometrica} 52: 647-663.
{p 4 8 2}Singh, S.K. and G.S. Maddala (1976). A function for the size
distribution of income. {it:Econometrica} 44: 963-970.
{p 4 8 2}Wifling, B. and W. Kraemer (1993). The Lorenz-ordering of Singh-
Maddala income distributions. {it:Economics Letters} 43: 53-57.
{title:Also see}
{p 4 13 2}
Online: help for {help smpred}, {help psm}, {help qsm}, {help dagumfit}, {help gb2fit},
{help lognfit}, if installed.