Search on Cross Validated… 1

SPONSORED BY

Home
Mboxcox, interpreting difficult regressions Ask Question

Questions Asked 7 years, 8 months ago Active 2 months ago Viewed 3k times
Tags

Users Mboxcox in Stata suggests transforming my variables using a power of 0.1 for the independent
Featured on Meta
variable, and 0.4 for the dependent variable.
Unanswered
3 Opt-in alpha test for a new Stacks editor
I have run the model, and it fixes problems associated with the assumptions of OLS. But certainly,
it complicates matters in terms of interpretations. Visual design changes to the review
queues
Please outline possible interpretations and solutions.
2
Hot Meta Posts
Dependent variable is in millions of dollars, and independent variables are number of years. and a
number of dummies which do not require transformation. 25 Would Cross Validated want Machine
Learning Theory questions that are no…

regression data-transformation stata 14 Some tags need better names

9 What are some “best practices” for

structuring/formatting a post?
edited May 30 '13 at 20:48 asked May 30 '13 at 9:43
Share Cite Edit Follow Cesare Camestre
549 3 12 27 Linked

0 Why transform distribution could help

Why would you need to transform your independent variable? – Glen_b May 30 '13 at 13:51 model accuracy?

0 Transformations in Simple Linear

2 If you carefully read the article on which this code is based, you will find this key bit of advice: "Following
the advice of Sheather (2009), we round the suggested powers to the closest interpretable fractions" Regression
(emphasis added). For many people, those would be the "fractions" 0 (in place of 0.1) and 1/2 (or
122 What if residuals are normally distributed,
possibly 1/3 , in place of 0.4). Some might even round 0.4 to 0 or 1. You need to select the rounding
but y is not?
based on the possible interpretations and what any underlying theory suggests. – whuber ♦ May 30 '13 at
15:49 57 Box-Cox like transformation for
independent variables?
1 @glen_b One excellent reason to transform the independent variable is explained and illustrated (with an
example) here. – whuber ♦ May 30 '13 at 15:51 46 Normality of dependent variable = normality
of residuals?
2 @whuber Thanks; indeed I am aware of that as a reason to do so (and do so transform sometimes, for
that reason). I wondered why the OP in particular was doing it (hoping it was something like that, but 4 transformation to normality of the
considering the possibility that they might instead have thought IVs need to have a normal distribution). –
dependent variable in multiple regression
Glen_b May 30 '13 at 16:05
1 Log-linear transformation
1 I'd definitely support @whuber's mention of the advice of Sheather for rounding the values (indeed,
Sheather's advice is often worth considering); Nick Cox's answer below says essentially the same thing.
2 Is log transformation a proper way to
I'd also suggest looking at Tukey's discussion of his 'ladder of powers' for similar advice about taking
interpretable round numbers. – Glen_b May 30 '13 at 16:10 reduce the weight of high vs. low values in
logistic regression, and how do I diagnose
Rounding to 0 doesn't make much sense! I would end up with everything 1! – Cesare Camestre May 30 when the DV is binary?
'13 at 19:17
0 Box-Cox Transformation for predictor
If so, the implication is beautifully simple and clear: no transformations are required. – Nick Cox May 30 variables in R
'13 at 19:40
1 What is the largest n root transformation I
4 A Box-Cox parameter of zero corresponds to the logarithm, not the zero power. It sounds like you would should consider for making a time series
benefit from reading some more background about this technique before proceeding further. – whuber ♦ stationary?
May 30 '13 at 19:45
See more linked questions
I am quite sure a transformation is needed, because I have issues with the assumptions of the regression
model. Using logs, seem to solve most problems apart from normal distribution, but some questioned
adding a constant (such as 1) in view that one of the variables is age.. – Cesare Camestre May 30 '13 at Related
19:49
21 How to interpret regression coefficients
Ok - agreed but then I have age.. and i was criticised for proposing taking ln(age+1). – Cesare Camestre
when response was transformed by the 4th
May 30 '13 at 21:01
root?
@Glen_b "they might instead have thought IVs need to have a normal distribution" Is it a statistical issue
0 Estimating the constant in multiple
for IVs to have a non-normal distribution? And if not, why? Being a parametric procedure, I was lead to
regression
believe that this is among the basic requirements. – landroni Mar 6 '14 at 14:28
2 Using Poisson instead of transforming the
2 @landroni Since we condition on the IVs in ordinary regression, there's no distributional assumption
data
whatever on them. Indeed, there's not even any unconditional distributional assumption on the DV itself
(i.e. looking at a histogram or QQ plot etc of the DV is no use, since we make no particular assumption
8 Non-normality in residuals
about that). If one is doing the usual normal-theory inference (hypothesis tests, CIs, PIs), there's an
assumption about the distribution of the error term, which is assessed by residual diagnostics. This is
7 What is the difference between a
discussed in comments or answers on dozens of questions here by various people. – Glen_b Mar 6 '14 at
21:24 hierarchical linear regression and an
ordinary least squares (OLS) regression?
1 @landroni Even then, the normality assumption may not be particularly crucial; in large samples it's
usually only much of an issue for prediction intervals. (In small samples it matters much more.) – Glen_b 1 Discrete choice model
Mar 6 '14 at 21:29
0 Adding trend does not impact coefficients
@Glen_b Thanks for the explanations. I find all this curious as I've read conflicting arguments in different
sources (although most likely it's just me who is confusing things up!). I'm currently reading through the 1 Control variables in Seemingly Unrelated
reams of comments on these issues throughout the site, and will come back with a separate question if Regression (SUR)
I'm still unconvinced. – landroni Mar 6 '14 at 21:38

1 @landroni A list of the assumptions is here (the list can vary somewhat because people can add or Hot Network Questions
remove things at the periphery which may not be explicitly used in the derivations; the core assumptions
don't change). If you'd like to discuss it some more, we can take it to chat – Glen_b Mar 6 '14 at 22:02 Plot many curves on the same graph

@Glen_b Thanks for the link, and for the chat invitation! I'll try to first properly do my homework before Does voltage depend on the surface area of a
taking up that. :) – landroni Mar 6 '14 at 22:37 conductor?

Is this system causal or not?

@Glen_b This question has excellent answers on the issue of normality: What if residuals are normally
distributed, but y is not?. – landroni Mar 7 '14 at 14:54 Why do some PCB designers put pull-up resistors
on pins where there is already an internal pull-up?
Add a comment
I bring villages to my compound but they keep
going back to their village. Why is that?
2 Answers Active Oldest Votes Chamishit - mi yodeya?

Programmatically define macro within the body of

\foreach
In my experience, likelihood methods for finding Box-Cox transformations of data are both
poor (in performance) and unstable. They are contrary to the spirit and intended use of How to avoid an hyperref bug (?) in theorems

11 transformations, too, which include: Where does Martian meaning inhabitant of Mars
come from?
Finding interpretable re-expressions of data, Can a country be only de jure sovereign ? (Such
as Andorra)
Attempting to linearize relationships,
What Justification can I give as for why my
Attempting to achieve homoscedastic relationships, vampires sleep specifically in coffins?

What is the diference betwen 電気製品 and 電化製

Allowing interactive exploration of data analysis options, 品?
Using calculations that are resistant to outlying values, and confirming this will blink LED at given frequency

Being robust to alternative (but plausible) assumptions about data behavior. Is attempted murder the same charge regardless
of damage done?

Instead, by its very nature, a likelihood-based method (such as mboxcox , which aims to achieve Do wormholes really exist?
[approximate] multinormality), violates all these aims, as you can check, one-by-one. Three queens and two rooks covering the chess
board... again!
Nevertheless, almost since the time Box-Cox transformations were first described, people have
Why is it difficult to drag with the Magic Trackpad
been coming up with automated ways to estimate them. Few work well, but many sometimes give 2?
an approximate starting point, or range of starting points, to streamline the exploration. How can a deep discount bond with a longer time
to maturity have a LOWER duration than an
Before we go on, let's establish the correct use of mboxcox . A careful reading of the article on otherwise identical bond with a shorter time to
maturity?
which this code is based find this important suggestion: "Following the advice of Sheather (2009),
Ginormous number
we round the suggested powers to the closest interpretable fractions" (emphasis added).
Traditionally, an "interpretable fraction" [sic] is a value that might appear in a physical theory: 1/2 , Is it a vampire number?

1/3 (and their negatives) along with whole values 0, 1, 2 (and their negatives). Thus it was never How can I draw the switch in circuitikz?
intended that the user accept the output of mboxcox as-is: it has to be rounded according to
Bosch Drawer Microwave causes ARC Faults
knowledge of the data and the objectives of the analysis.
Are there any 3rd level spells a Lore Bard could
pick at 6th character level to provide food and
water to the party?
As a quick test of mboxcox , I applied it to the classical Mercury vapor pressure dataset
popularized by John Tukey. It is the simplest multivariate dataset possible, containing 19 What is black's idea in this puzzle?

(temperature, pressure) pairs with extremely small errors, leading to little uncertainty in what the
Question feed
best Box-Cox parameters ought to be. An exploratory data analysis (EDA) of this dataset is
described in my answer at Box-Cox like transformation for independent variables?. It finds,
correctly, that (after converting temperature to absolute temperature) the Box-Cox powers should
be 0 for the pressure and −1 for the temperature. (In this case, unlike in most data analyses, there
is a correct answer given by a well-known physical law. That is why it makes a fine proving ground
for any automated procedure.)

By contrast, when we apply mboxcox , at first it complains that it cannot deal with the zero
temperature. If we simply exclude it--it will later found to be an outlier, anyway--it reports that the
Box-Cox parameter for pressure should be 0.3420578 (comfortably close to the convenient 1/3 ,
with a 95% CI from 0.22 to 0.46 ) and for temperature should be 2.2386 (CI from 1.5 to 3.0 ),
which could be taken as close to 2. Good, right? Both are highly significant.

But, as we know--and can see in the nonlinear trend in the scatterplot--these are awful results,
because we really need to be using the absolute temperature. Let's start over after adding 273
degrees to the temperatures. Because there will no longer be a problem with zero, we will include
all the data. This time mboxcox reports that the Box-Cox parameters should be 0.0712114 for
pressure and 0.2411739 for temperature (CI from −0.8 to 1.3 ). Even after rounding both values
and accounting for the long confidence interval for temperature, the results are far from correct--
even though the 𝑅2 value in the resulting regression of (transformed) pressure on (transformed)
temperature actually exceeds what is achieved when the correct parameters are used!

Although this is a beautifully linear relationship on this scale, examination of the residuals shows it
leaves much to be desired.

NB: The y-axes on these plots are not directly comparable, because they represent different re-
expressions of the pressures. What is of concern are the apparent patterns of non-linear behavior
and serial correlation in each plot. The right plot does a better job at identifying the outlier (at a
temperature of 0) and is much more horizontal than the left plot ( mboxcox ), which shows a clear
curvilinear trend and is nowhere horizontal.

If we remove the case with the lowest temperature (which is the main source of the difficulty, even
though it's not much of an outlier), mboxcox finally gets it right: it estimates a parameter of 0.013
for pressure, which clearly rounds to 0, and −0.8559 for temperature (CI from −1.4 to −0.3),
which anyone would round to −1 , with a possible −1/2 contained in the confidence interval. But it
took three tries and required an insight (the use of absolute temperatures) that dropped out of the
original EDA but had to be supplied by the analyst using mboxcox .

With the results of this quick look we may deduce that mboxcox indeed has the potential to deliver
useful starting points for an EDA of multivariate data, provided it is carefully protected by first
identifying outliers, that the estimates are appropriately rounded, and that the data are further
explored to make sure that other Box-Cox parameters (even those far from the "optimal" ones)
might not serve better. I would give little weight or credence to the mboxcox results without
extensive follow-on analysis, because although it aims at establishing an approximate multivariate
normal distribution of the data, that does little to assure either linearity or homoscedasticity and
likely places too much emphasis on transforming the independent variables instead of the
dependent variable itself.

edited Nov 26 '20 at 19:10 answered May 30 '13 at 16:57

Share Cite Edit Follow kjetil b halvorsen ♦ whuber ♦
52.5k 22 119 383 257k 49 566 1002

6 Excellent answer. I'd underline that identifying a good functional form for the relationship between
variables is widely underestimated in importance, while being right about the distribution of errors is widely
overestimated. – Nick Cox May 30 '13 at 17:15

1 Thanks @Nick. I'm actually expecting comments and debate about what makes the result of an EDA
"correct." The example here is interesting in that if we had no prior theory to guide us, it would be much
tougher to decide whether to reject the second mboxcox result: it hinges on a somewhat delicate
assessment of the residuals. In practice, especially with sociodemographic and biological data, different
analysts using different procedures can legitimately arrive at equally valid but strikingly different ways of
expressing the variables, leading to different functional forms. – whuber ♦ May 30 '13 at 17:27

I do not have issues with linearity and homoscedasticity, only with the normality if I adopt a log-linear
model (and ignore the mboxcox approach), but then an issue arose because of age.. and age of firm
could be 0. – Cesare Camestre May 30 '13 at 19:31

2 You question indicates the contrary: since mboxcox returned different values of the parameter for the
DV and IV, it sounds like you do have an issue with linearity; and because the value it returned for the DV
differs substantially from 1, you do have an issue with homoscedasticity! Moreover, it would be rare to
have any reason to want regression data to be normal; this is neither expected nor assumed of standard
regression techniques. The presence of a zero in your data also indicates you will have problems with
mboxcox , as my own experience attests. – whuber ♦ May 30 '13 at 19:37

In the initial data, you are correct there are issues with linearity, and homoskedasticity. These could be
fixed using logs: ln(investment) = ln(age+1) + d1 + d2 + d2.. which seems to be a suitable transformation
(solves most problems) but this model does not solve (a) normal distribution problems (it is negatively
skewed) (b) there is an issue with adding a constant to the ln(age), BUT this has to be done because age
can be 0. So I went back to original data, and tried to use mboxcox, to see if there is a beter alternative
and came up with those strange powers I mentioned previously. – Cesare Camestre May 30 '13 at 20:19

1 That's a good start. (Experience and some theory suggest automatically taking the log of investment and
then considering using the square root of the age, should that be necessary.) But what is negatively
skewed? Log(investment) or their residuals? Only the latter matters, not the former. As far as adding 1 to
age goes, there are several good threads here discussing this issue, but this won't be an issue if you use
the root of age or don't transform it at all. – whuber ♦ May 30 '13 at 21:22

The residuals are slightly negatively skewed (-0.7) after applying the log to the investment –
Cesare Camestre May 30 '13 at 21:41

The values you get after applying the log are not the residuals: the residuals are the differences between
the logs and the fitted values in your linear model. – whuber ♦ May 30 '13 at 21:41

Whuber you are right, but I'm applying a Shapiro-Wilk test on THE RESIDUALS. – Cesare Camestre
Jun 5 '13 at 13:58

I tried using log of investment and the root of age- this way I still do not get normal distribution. –
Cesare Camestre Jun 5 '13 at 14:12

@whuber "Moreover, it would be rare to have any reason to want regression data to be normal; this is
neither expected nor assumed of standard regression techniques." I'm no expert, but I am surprised.
Being a parametric procedure, I was lead to believe that normality of the underlying data was among the
basic requirements of OLS. Is it not? – landroni Mar 6 '14 at 15:24

2 @landroni The distinction being made is between the data (the response values) and their residuals.
Some regression techniques make parametric assumptions about the distributions of the residuals. It is
rare that additional assumptions are made about the response itself: its distribution is determined by the
explanatory variables and the distributions of the residuals. – whuber ♦ Mar 6 '14 at 23:11

@whuber Thanks for the explanation. Very curious to learn that what matters is the distribution of the
response conditional on the IVs. What about the IVs themselves? No distributional assumptions on them,
either? – landroni Mar 7 '14 at 7:31

2 @landroni Not for fixed-effects models: they do not assume the IVs are random variables at all. –
whuber ♦ Mar 7 '14 at 13:54

@whuber This question has excellent answers on the issue of normality: What if residuals are normally
distributed, but y is not?. – landroni Mar 7 '14 at 14:51

@landroni Thank you. This is a FAQ and is discussed (I suspect) in several hundred threads and at least
as many comments. – whuber ♦ Mar 7 '14 at 14:53

Add a comment

Although you have given some details, this is too close to "I have some data, want to fit a
regression, and can't interpret my model easily" to allow much to be said easily that is likely to be
6 really helpful. Too much depends on what your field is, what models make sense or are interesting
substantively in that field, etc., not to mention finer details of your data. Not least, what is
"interpretation"? It can mean anything from "I don't understand the statistics here, so need
technical explanation on my level" to "What does this imply in subject-matter terms?".

But (personal opinions mixed in here)

If your response or dependent variable is a count, I would expect Poisson regression to make
much more sense than regression. Even if it is a measured number of years that is zero
upwards, I would still expect that. http://blog.stata.com/tag/poisson-regression/ is one account
rich in Stata context.

The idea of Box-Cox is letting your data indicate which transformations make most sense.
However, Box-Cox like much else is a knife that you can cut yourself with. The original
examples are instructive: Box and Cox didn't use the precise powers indicated, but logarithm
and reciprocal, which made sense on other grounds. Unless you are fitting a power law, it is
usually more practical to regard Box-Cox as pointing to one of a small number of standard
transformations, most commonly log, root or reciprocal. It is rare that (say) powers such as 0.4
can be related to substantive literature unless there is good theory underpinning the use of
fractional powers in the first place. The fact that most common transformations can be
regarded as members of a family doesn't mean that all members of that family are equally
helpful.

edited May 30 '13 at 19:29 answered May 30 '13 at 10:11

Share Cite Edit Follow Nick Cox
45.2k 8 103 149

I do not have a count in the independent variable. As I said, its millions of pounds, and therefore I would
not use poisson regression. Counts are in the dependent variables, and previously I was told that when
that happens there is no case to use poisson regression. – Cesare Camestre May 30 '13 at 19:22

2 Sorry, my ambiguity, which I have fixed in an edit. I was referring to your response (dependent variable). (I
did say "number of years".) Contrary to your assertion, counts as response are the canonical case for
Poisson regression. – Nick Cox May 30 '13 at 19:29

1 There's confusion here, Nick, but it's not your fault: the question still states the DV is "number of years,"
which is manifestly not a count, it's a duration. – whuber ♦ May 30 '13 at 19:44

@IdiotAbroad You are telling us currently that your dependent variable is a count. But on checking your
post again, I see that you refer to several dependent variables. Are you confusing dependent
(conventionally 𝑦 ) and independent variables (conventionally 𝑥 )? – Nick Cox May 30 '13 at 19:49

Sorry - I do apologise. I cannot concentrate right now. My regression is investment = age + d1 + d2 + d3

etc – Cesare Camestre May 30 '13 at 19:58

The dependent variable is not a count! – Cesare Camestre May 30 '13 at 20:06

Whuber, i fixed the question - apologies, I have mixed the terms when I actually wrote the question; and
only realised just now! – Cesare Camestre May 30 '13 at 20:11

Add a comment

Your Answer

Links Images Styling/Headers Lists Blockquotes Code HTML Tables Advanced help

Post Your Answer

Not the answer you're looking for? Browse other questions tagged regression data-transformation stata

or ask your own question.

CROSS VALIDATED COMPANY STACK EXCHANGE Blog Facebook Twitter LinkedIn Instagram
NETWORK
Tour Stack Overflow
Technology
Help For Teams
Life / Arts
Chat Advertise With Us
Culture / Recreation
Contact Hire a Developer
Science
Feedback Developer Jobs
Other
Mobile About
Disable Responsiveness Press
Legal
Privacy Policy
Terms of Service
site design / logo © 2021 Stack Exchange Inc; user contributions licensed
Cookie Settings under cc by-sa. rev 2021.2.9.38523

Accept all cookies Customize settings By clicking “Accept all cookies”, you agree Stack Exchange can store cookies on your device and disclose information in accordance with our Cookie Policy.