Multiplicative Errors: Log-Normal or Gamma?

Author(s): David Firth

Source: Journal of the Royal Statistical Society. Series B (Methodological) , 1988, Vol. 50,
No. 2 (1988), pp. 266-268
Published by: Wiley for the Royal Statistical Society

Stable URL: https://www.jstor.org/stable/2345764

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Royal Statistical Society and Wiley are collaborating with JSTOR to digitize, preserve and
extend access to Journal of the Royal Statistical Society. Series B (Methodological)

This content downloaded from

172.56.66.42 on Sun, 27 Aug 2023 07:49:54 +00:00
All use subject to https://about.jstor.org/terms
J. R. Statist. Soc. B (1988)
50, No. 2, pp. 266-268

Multiplicative Errors: Log-normal or Gamma?

By DAVID FIRTHt

Imperial College, London, UK

[Received May 1987. Revised September 1987]

SUMMARY
Estimation in regression models with multiplicative error is commonly based on either the
log-normal or the gamma likelihood. When the regression effects are also multiplicative,
the gamma maximum likelihood estimate is found to be the more efficient of the two, but
only slightly so, under reciprocal misspecification.

Keywords: CONSTANT COEFFICIENT OF VARIATION; EFFICIENCY; MULTIPLICATIVE MODEL;

QUASI-LIKELIHOOD

1. INTRODUCTION

There is a well-known correspondence between multiplicative regression models for

positive observations and additive models for their logarithms. For if
Yi = 1-i?'i (i = 1, ... ., n)(1
with log pi = ho + xj,/3 + ... + xip/p3 and {?i} independently i
(IID) with E(ii) = 1, then
log Yi=vi+?li (i= , ...,n) (2)
with vi = {If3 + E(log Pi)} + x ,3 + +.. ? xip
non-intercept parameters ,Bl, ..., ,,p may
equation (2), provided that xir + . .. + xnr = 0 (r = 1, . . ., p); this 'centring' of the
explanatory variables is always possible, and ensures identifiability of fl, . . ., /p. The
intercept parameters in equations (1) and (2) cannot, in general, be similarly identified,
but this will often be unimportant in practice.
If the second moments of both Yi and log Yi are assumed to exist, the quasi-likelihood
method described by Wedderburn (1974) applies to either model (1) or model (2).
Two different estimation procedures result:

(a) gamma (G) solve

E Yi-H ir = = (r = 0, p, P
i=l H-i

and

(b) log-normal (LN) solve

n

Z (log Yi-Vi)Xir = O (r = 0, ..., p),

i = 1

where yi denotes the realised value of Yi and xio = 1 for all i. The estimating equations

t Addressfor correspondence: Department of Mathematics, The University of Texas at Austin, Austin, TX 78712,
USA.

? 1988 Royal Statistical Society 0035-9246/88/50266 $2.00

This content downloaded from

172.56.66.42 on Sun, 27 Aug 2023 07:49:54 +00:00
All use subject to https://about.jstor.org/terms
 1988] Multiplicative Errors 267

in (a) are quasi-likelihood equations appropriate to model (1), which has a constant
coefficient of variation and logarithmic link; alternatively, they are maximum
likelihood equations under the assumption that each Yi has a gamma distrib
with mean pi. Model (2), in contrast, has constant variance and identity lin
to quasi-likelihood equations as in (b); these are just unweighted least squares
equations, or equivalently maximum likelihood equations assuming each log Yi to be
normally distributed.
Which procedure is to be preferred? McCullagh and Nelder (1983) summarise the
main issues. Two important considerations are ease of interpretation and the
applicability of familiar techniques, e.g. diagnostics. Here, though, we concentrate on
efficiency.

2. EFFICIENCY

If it is known that the {1i} are gamma distributed then proce

likelihood and has full asymptotic efficiency; likewise, procedure (b) is fully efficient
if the {Je} have a log-normal distribution. More generally, when the distribution o
the {Je} is not assumed known, the choice is less straightforward. However, an
interesting comparison arises when the results of Cox and Hinkley (1968), on the
efficiency of least squares, are combined with those of Firth (1987), sect. 3, on the
efficiency of quasi-likelihood estimation in models with constant coefficient of
variation.
Cox and Hinkley (1968) show how the asymptotic efficiency of least squares
estimation of (#1, ..., f,p) in a constant variance linear model such as mo
relative to maximum likelihood based on the true distribution for {1ij, may be
as a scalar quantity depending on the shape of the true distribution but not on {Xir}.
One of their examples is the case when the {iij are log-gamma distributed, i.e. eac
exp i1i has a gamma distribution with index y. The efficiency of least squares in this
case is

{y+'(y)} -' = effG(LN), (3)

say, where f'( ) is the trigamma function; by the correspondence described between
additive and multiplicative models, equation (3) is precisely the asymptotic efficiency
of method (b) for estimating (,6, ..., fl,p) in model (1) when the {le} are gamma
distributed with index y.
Firth (1987), sect. 3, demonstrates an analogous property for quasi-likelihood
estimation, i.e. maximisation of the gamma likelihood, in the multiplicative model
(1): the asymptotic efficiency for the non-intercept parameters (fl1, ..., ,Bp) is a sc
depending on the shape of the true distribution of the {1i}. In the particular
when each ei has a log-normal distribution with mean unity and variance 0, the
efficiency is

log(1 + f/k= effLN(G), (4)

say.
Identifying yT with q as the variance of ?i, the efficiencies in equations (3) and (4)
may be directly compared (Table 1).
It is apparent from Table 1 that, at least at these realistic values of var(86), the lo
of efficiency involved in using procedure (a) when the errors are log-normally
distributed is less than that incurred by use of procedure (b) when the data have a

This content downloaded from

172.56.66.42 on Sun, 27 Aug 2023 07:49:54 +00:00
All use subject to https://about.jstor.org/terms
 268 FIRTH [No. 2,

TABLE 1
Asymptotic efficiencies of estimators G and
LN in the multiplicative model (1)

4 = var(e) effLN (G) effG(LN)

0.1 0.953 0.951

0.2 0.912 0.904
0.5 0.811 0.775
1.0 0.693 0.608
2.0 0.549 0.405

gamma distribution. In this sense it is safer to use quasi-likelihood estimation on the

original scale than to apply least squares to the log-transformed data. However, the
difference is slight. For small 4, effLN(G) and effG(LN) may be approximated as follows

effLN(G)= 1 2 + %02 + (2),

and, using a well-known formula for the trigamma function (Abramowitz and Stegun,
1965),

effG(LN)= 1- -q- + 1 2 + o(qY).

The difference, then, between effLN(G) and effG(LN) is of order 42, an order of
magnitude smaller than might have been anticipated.

3. DISCUSSION

The efficiency comparison of the previous section indicates that, although gamma-
based estimation performs better than the log-normal method under reciprocal
misspecification, the difference is unlikely to be a major factor affecting choice between
the two methods. Other criteria will be more important in practice. For example, a
least squares analysis of the logarithms of the data allows the use of well-developed
diagnostic tools, while analysis on the original scale may have advantages for
interpretation. A point worth noting here concerns behaviour under failure of the
assumption that the multiplicative errors are identically distributed. For while
quasi-likelihood estimation on the original scale remains consistent for ( *--, /3,
the same is not generally true of least squares applied to {log Y,}; if E(log si) depe
on i, the quantities that are consistently estimated by method (b) may have no
interpretation on the original scale.

ACKNOWLEDGEMENTS

I thank Professor Sir David Cox for helpful discussions. This work was supported
by the Science and Engineering Research Council.

REFERENCES

Abramowitz, M. and Stegun, I. A. (eds) (1965) Handbook of Mathematical Functions, p. 260. Washington DC: National
Bureau of Standards, US Government Printing Office.
Cox, D. R. and Hinkley, D. V. (1968) A note on the efficiency of least-squares estimates. J. R. Statist. Soc. B, 30,284-289.
Firth, D. (1987) On the efficiency of quasi-likelihood estimation. Biometrika, 74, 233-245.
McCullagh, P. and Nelder, J. A. (1983) Generalized Linear Models, sect. 7.1. London: Chapman and Hall.
Wedderburn, R. W. M. (1974) Quasi-likelihood functions, generalized linear models, and the Gauss-Newton method.
Biometrika, 61, 439-447.

This content downloaded from

172.56.66.42 on Sun, 27 Aug 2023 07:49:54 +00:00
All use subject to https://about.jstor.org/terms