Handling interactions in Stata

Stata,
especially with continuous
predictors
di t

Patrick Royston & Willi Sauerbrei

German Stata Users’ meeting,
g, Berlin,, 1 June 2012
Interactions – general concepts

• General idea of a (two-way)

(two way) interaction in
multiple regression is effect modification:
• η(x1,x x2) = f1(x1) + f2(x2) + f3(x1,x x2)
• Often, η(x1,x2) = E(Y | x1,x2), with obvious
extension to GLM, GLM Cox regression,
regression etc.etc
• Simplest case: η(x1,x2) is linear in the x’s and
f3((x1,,x2) is
s tthe
epproduct
oduct o
of tthe
e x’s:
s
• η(x1,x2) = β1x1 + β2x2 + β3x1x2
• Can extend to more general, non non-linear
linear
functions

1
The simplest type of interaction

• Binary x binary Treatment White cell White cell

group count low count high
• E.g. in the MRC RE01 (>10)
(<=10)
trial in kidney cancer
• 12 month % survival MPA 34% (se 4) 24% (se 4)
since randomisation
• Substantial treatment
effect in patients with
low WCC Interferon 49% (se 4) 21% (se 7)

• Little or no treatment
effect in those with
high
g WCC
• But really, WCC is a
continuous variable …
Overview

• Interactions and factor variables (Stata 11/12)

• Note: I am not an expert on factor
variables! I sometimes use them.
them
• General interactions between continuous
covariates in observational studies
• Focus on continuous covariates …
• … because people don
don’tt appear to know how
to handle them!
• Special case: interactions between treatment
and continuous covariates in randomized
controlled trials
3
Interactions and factor variables

4
Scope

• We introduce the topic with a brief

introduction to factor variables
• In this part
part, we consider only linear
interactions:
• Binary x binary (2 x 2 table)
• Binary x continuous
• Continuous x continuous

5
 Factor variables: brief notes

• Implemented via prefixes (unary operators) and

binary interaction operators
• see help fvvarlist
• There are four factor-variable operators:
Operator Description
i. unary operator to specify indicators (dummies)
c. unary operator to treat as continuous
# binary operator to specify interactions
## binary operator to specify factorial interactions
• Dummy variables are ‘virtual’ – not created per se
• Names of regression parameters easily found by
inspecting the post-estimation
post estimation result matrix e(b)
(b)
Factor variables: i.
i prefix

• Example from Stata manual [U]11.4.3:

[U]11 4 3:

. li
list
t group i.group
i i
in 1/5
group 1b.group 2.group 3.group
1
1. 1 0 0 0
2. 1 0 0 0
3
3. 2 0 1 0
4. 2 0 1 0
5. 3 0 0 1
Example dataset

• MRC RE01 trial in advanced kidney cancer

• Of 347 patients, only 7 censored, the rest died
• For
F simplicity,
i li it as a continuous
ti response
variable, Y, we use months to death, _t
• Ignore the small amount of censoring
• There are several prognostic factors that may
influence time to death
• Some are binary, some categorical, some
continuous

8
Example: factor variable parameters

. regress _t
t i.who
Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 2, 344) = 15.72
Model | 7780.81413 2 3890.40707 Prob > F = 0.0000
R id l | 85126
Residual 85126.5686
5686 344 247
247.460955
460955 R-squared
R d = 0 0837
0.0837
-------------+------------------------------ Adj R-squared = 0.0784
Total | 92907.3828 346 268.518447 Root MSE = 15.731

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
who |
1 | -4.2783 2.026215 -2.11 0.035 -8.263629 -.2929707
2 | -12.94365
12.94365 2.354542 -5.50
5.50 0.000 -17.57476
17.57476 -8.312534
8.312534
|
_cons | 19.17358 1.622518 11.82 0.000 15.98227 22.36488
------------------------------------------------------------------------------

. matrix
i li
list e(b)
(b)

e(b)[1,4]
0b.
0b 1.
1 2.
2
who who who _cons
y1 0 -4.2782996 -12.943646 19.173577
Basic analysis
y to understand binary
y x binary:
y
the 2 x 2 table of means

• Example:
. table rem sex, contents(mean _t) format(%6.2f)

--------------------------
x6: |
kid
kidney |
removed | x2: sex
(Y/N) | male female
----------+---------------
0 | 13.50 9.34
1 | 13.77 18.19
--------------------------

10
 Fitting an interaction model
• Consider 3 methods:
• Method 1: binary operator to specify interactions
• regress
g _t rem#sex
• Method 2: binary operator to specify factorial
interactions
• regress _t rem##sex
• Method 3: create multiplicative
p term(s)
( )yyourself
• gen byte remsex = rem * sex
• regress
g _t rem sex remsex
• Models are identical - all give the same fitted values
• Parameterisation of method 1 is different
11
Method 1: Binary operator #

. regress _t
t rem#sex
Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 3, 343) = 2.73
Model | 2168.10384 3 722.701279 Prob > F = 0.0437
Residual | 90724.0188 343 264.501513 R-squared = 0.0233
-------------+------------------------------ Adj R-squared = 0.0148
Total | 92892.1227 346 268.474343 Root MSE = 16.264

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
rem#sex |
0 1 | -4.160995 2.888457 -1.44 0.151 -9.842314 1.520323
1 0 | .2724983
2724983 2
2.137064
137064 0.13
0 13 0
0.899
899 -3.930903
3 930903 4
4.475899
475899
1 1 | 4.68417 2.581164 1.81 0.070 -.3927323 9.761072
|
_cons | 13.49939 1.610327 8.38 0.000 10.33204 16.66675
------------------------------------------------------------------------------

Notes on parameters:
rem#sex 0 1 (-4.16) is (sex=1) - (sex=0) at rem=0
rem#sex 1 0 (+0.27) is (rem=1) - (rem=0) at sex=0
rem#sex 1 1 (+4.68)
(+4 68) is [(rem=1)
[(rem 1) | sex=1]
sex 1] - [(rem=0)
[(rem 0) | sex=0]
sex 0]
_cons (13.50) is intercept (mean of Y at rem=0 & sex=0)
I don’t recommend this parameterisation! 12
Method 2: Binary operator ##

. regress _t
t rem##sex
Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 3, 343) = 2.73
Model | 2168.10384 3 722.701279 Prob > F = 0.0437
Residual | 90724.0188 343 264.501513 R-squared = 0.0233
-------------+------------------------------ Adj R-squared = 0.0148
Total | 92892.1227 346 268.474343 Root MSE = 16.264

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.rem | .2724983 2.137064 0.13 0.899 -3.930903 4.475899
1.sex | -4.160995 2.888457 -1.44 0.151 -9.842314 1.520323
|
rem#sex |
1 1 | 8.572667 3.792932 2.26 0.024 1.112333 16.033
|
_cons | 13.49939 1.610327 8.38 0.000 10.33204 16.66675
------------------------------------------------------------------------------

Notes on parameters:
1.rem (0.27) is (rem=1) - (rem=0) at sex=0
1
1.sex (-4.16)
( 4 16) is
i (sex=1)
( 1) - (sex=0)
( 0) att rem=0
0
rem#sex (+8.57) is [(rem=1) - (rem=0) at sex=1] - [(rem=1) - (rem=0) at sex=0]
_cons (13.50) is intercept (mean of Y at rem=0 & sex=0) 13
This is a ‘standard’ parameterisation with P-value as given above
 Method 3: DIY multiplicative term

. generate byte remsex = rem * sex

. regress _t rem sex remsex

Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 3, 343) = 2.73
Model | 2168.10384 3 722.701279 Prob > F = 0.0437
Residual | 90724.0188 343 264.501513 R-squared = 0.0233
-------------+------------------------------
+ Adj R-squared
R squared = 0.0148
0 0148
Total | 92892.1227 346 268.474343 Root MSE = 16.264

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
rem | .2724983 2.137064 0.13 0.899 -3.930903 4.475899
sex | -4.160995 2.888457 -1.44 0.151 -9.842314 1.520323
remsex | 8.572667 3.792932 2.26 0.024 1.112333 16.033
_cons | 13.49939 1.610327 8.38 0.000 10.33204 16.66675
------------------------------------------------------------------------------
Notes on parameters:
rem is 1.rem in Method 2
sex is 1.sex in Method 2
remsex is rem#sex in Method 2
This is the same parameterisation as Method 2 14
Interactions in non-normal
non normal errors models

Key points:

1 In
1. I a 2 x 2 table,
t bl an interaction
i t ti iis a ‘diff
‘difference
of differences’
2 Tabulate the 2 x 2 table of mean values of the
2.
linear predictor
3 May back
3. back-transform
transform values via the inverse link
function
• e.g. exponentiation in hazards models

15
 Binary x continuous interactions

• Use c.
c prefix to indicate continuous variable
• Use the ## operator
. regress _t
t trt##c.wcc
trt##c wcc
Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 3, 343) = 7.44
Model
d l | 5678.62935 3 11892.87645 Prob
b > F = 0.0001
1
Residual | 87228.7534 343 254.311234 R-squared = 0.0611
-------------+------------------------------ Adj R-squared = 0.0529
Total | 92907.3828 346 268.518447 Root MSE = 15.947

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1
1.trt | 12
12.81405
81405 4
4.124167
124167 3.11
3 11 0.002
0 002 4.702208
4 702208 20
20.92589
92589
wcc | -.2867831 .2741174 -1.05 0.296 -.8259457 .2523796
|
trt#c.wcc |
1 | -1.034239
1 034239 .4327233
4327233 -2.39
2 39 0
0.017
017 -1.885365
1 885365 -.1831142
1831142
|
_cons | 14.45292 2.712383 5.33 0.000 9.117919 16
19.78791
------------------------------------------------------------------------------
Binary x continuous interactions (cont.)
(cont )

• The main effect of wcc

cc is the slope in group 0
• The interaction parameter is the difference
between the slopes in groups 1 & 0
• Test of trt#c.wcc provides the interaction
parameter and test
• Results are nicely presented graphically
• Predict linear predictor xb
• Plot xb by levels of the factor variable
• Also,
Also ‘treatment
treatment effect plot
plot’ (coming later)
Plotting a binary x continuous interaction

. regress _t
t trt##c.wcc
trt##c wcc
. predict fit
. twoway (line fit wcc if trt==0, sort) (line fit wcc
if trt==1,
1 sort lp(-)),
l ( )) l
legend(lab(1
d(l b(1 "trt 0 (MPA)")
( ) )
lab(2 "trt 1 (IFN)") ring(0) pos(1))

trt 0 (MPA) trt 1 (IFN)

20
10
1
Fitted values
0
F
-10
-20

0 20 40 60
x9: white cell count (per l x 10^-9)
 Continuous x continuous interaction

• Just use c.
c prefix on each variable

. regress _
_t c.age##c.t_mt
_
Source | SS df MS Number of obs = 347
-------------+------------------------------ F( 3, 343) = 10.35
Model | 7714.26052 3 2571.42017 Prob > F = 0.0000
Residual | 85193.1223 343 248.37645 R-squared = 0.0830
-------------+------------------------------ Adj R-squared = 0.0750
Total | 92907.3828 346 268.518447 Root MSE = 15.76

------------------------------------------------------------------------------
_t | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | .0719063 .0876542 0.82 0.413 -.1005011 .2443137
t_mt | .0659781 .0128802 5.12 0.000 .040644 .0913122
|
c.age#c.t_mt | -.0008783 .0001861 -4.72 0.000 -.0012443 -.0005124
|
_cons | 8.055213 5.256114 1.53 0.126 -2.28306 18.39349
------------------------------------------------------------------------------
Continuous x continuous interaction

• Results are best explored graphically

• Consider in more detail next
Continuous x continuous
interactions

21
Motivation: continuous x continuous intn.
intn

• Many people only consider linear by linear

interactions
• Not sensible if main effect of either variable is
non-linear
• Mi
Mismodelling
d lli th
the main
i effect
ff t may iintroduce
t d
spurious interactions
•E
E.g. ffalse
l assumption
ti off li
linearity
it can create
t
a spurious linear x linear interaction
• Or
O they
h categorise
i the
h continuous
i variables
i bl
• Many problems, including loss of power
The MFPIgen approach (1)

• MFP = multivariable fractional polynomials

• I = interaction
• gen = generall
• Fractional polynomials (FPs) can be used to
model relationships that may be non-linear
non linear
• In Stata, FPs are implemented through the
standard fracpoly and mfp commands
• MFPIgen is implemented through a user-
written command,
o a d, mfpigen
p g
The MFPIgen approach (2)

• MFPIgen aims to identify non-linear

non linear main
effects and their two-way interactions
• Assume
u x1, x2 continuous
o uou and
a d z confounders
o ou d
• Apply MFP to x1 and x2 and z
• Force x1 and x2 into the model
• FP functions FP1(x1) and FP2(x2) are
selected for x1 and x2
• Linear
Li ffunctions
i could
ld b
be selected
l d
• Add term FP1(x1) × FP2(x2) to the model
chosen
• Apply likelihood ratio test of interaction
The MFPIgen approach in practice

• Start with a list of covariates

• Check all pairs of variables for an interaction
• Simultaneously,
Si lt l apply l MFP tto adjust
dj t ffor
confounders
• Use a low significance level to detect
interactions, e.g. 1%
• Present interactions graphically
• Check interactions for artefacts graphically
• Use forward stepwise if more than one
interaction remains
Example: Whitehall 1

• Prospective cohort study of 17

17,260
260 Civil
Servants in London
• Studied various standard risk factors for
common causes of death
• Also studied social factors,
factors particularly job
grade
• Wee consider
co s de 10-year
0 yea all-cause
a cause mortality
o ta ty as the
t e
outcome
• Logistic
g regression
g analysis
y
 Example: Whitehall 1 (2)

• Consider weight and age

. mfpigen:
p g logit
g all10 age
g wt
MFPIGEN - interaction analysis for dependent variable all10
------------------------------------------------------------------------------
variable 1 function 1 variable 2 function 2 dev. diff. d.f. P Sel
------------------------------------------------------------------------------
age Linear wt FP2(-1 3) 5.2686 2 0.0718 0
------------------------------------------------------------------------------
Sel = number of variables selected in MFP adjustment
j model

• Age function is linear, weight is FP2(-1, 3)

• No strong interaction (P = 0.07)
 Plotting the interaction model

. mfpigen,
mfpigen fplot(40 50 60)
60): logit all10 age wt
t

1
0
-1
-2
-3
-4
4
age by wt

40 60 80 100 120 140

x7: weight (kg)

fit o
on wt a
at age 40
0 fit o
on wt a
at age 50
fit on wt at age 60
 Mis specifying the main effects function(s)
Mis-specifying

• Assume age and weight are linear

• The dfdefault(1) option imposes linearity

. mfpigen, dfdefault(1): logit all10 age wt

MFPIGEN - interaction analysis for dependent variable all10
------------------------------------------------------------------------------
variable 1 function 1 variable 2 function 2 dev. diff. d.f. P Sel
------------------------------------------------------------------------------
age Linear wt Linear 8.7375
8 7375 1 0.0031
0 0031 0
------------------------------------------------------------------------------
Sel = number of variables selected in MFP adjustment model

• There appears to be a highly significant

interaction (P = 0.003)
Checking the interaction model

• Linear age x weight interaction seems

important
• Check if it’s real,
a , or
o the result
u ofo mismodelling
od g
• Categorize age into (equal sized) groups
p
• for example, 4ggroups
p
• Compute running line smooth of the binary
outcome on weight in each age group,
transform to logits
• Plot results for each group
• Compare with the functions predicted by the
interaction model
Whitehall 1: Check of age
g x weight
g linear
interaction
0
1st quartile 2nd quartile
3rd quartile 4th quartile
-1
ath))
Logit(pr(dea
-2
-3
-4
-

40 60 80 100 120 140

weight
Interpreting the plot

• Running line smooths are roughly parallel

across age groups ⇒ no (strong) interactions
• Erroneously assuming that the effect of weight
is linear ⇒ estimated slopes of weight in age-
groups
g p indicate strong
g interaction between
age and weight
• We should have been more careful when
modelling the main effect of weight
 Whitehall 1: 7 variables,
variables any interactions?

. mfpigen, select(0.05): logit all10 cigs

sysbp age ht wt chol i.jobgrade
i jobgrade
MFPIGEN - interaction analysis for dependent variable all10
------------------------------------------------------------------------------
variable 1 function 1 variable 2 function 2 dev. diff. d.f. P Sel
------------------------------------------------------------------------------
cigs FP1(.5) sysbp FP2(-2 -2) 0.7961 2 0.6716 5
FP1(.5) age Linear 0.0028 1 0.9576 5
FP1(.5) ht Linear 2.1029 1 0.1470 5
FP1(.5) wt FP2(-2 3) 0.1560 2 0.9249 5
FP1(.5) chol Linear 1.7712 1 0.1832 5
FP1(.5) i.jobgrade Factor 4.3061 3 0.2303 5

sysbp FP2(-2 -2) age Linear 3.1169 2 0.2105 5

(
(remaining
i i output
t t omitted)
itt d)
33
What mfpigen is doing

• FP functions for each pair of continuous

variables are selected
• Functions are simplified if possible
• Closed test procedure in mfp
• Controlled by the alpha() option
• The select(0.05) option tests confounders
for inclusion in each interaction model at the
5% significance level
• The Sel column in the output shows how
many variables are actually included in each
confounder
f d modeld l
34
Results: P values for interactions
P-values

. mfpigen, select(0.05):
mfpigen select(0 05): logit all10 cigs sysbp ///
age ht wt chol (gradd1 gradd2 gradd3)

*FP transformations were selected; otherwise, linear

 Graphical presentation of age x chol interaction

. fracgen cigs .5,

5 center(mean)
. fracgen sysbp -2 -2, center(mean)
. fracgen wt -2 3, center(mean)

. mfpigen, linadj(cigs_1 sysbp_1 sysbp_2

> wt_1 wt_2 ht i.jobgrade) df(1)
> fplot(%10 35 65 90): logit all10 age chol

Alternatively:

. logit all10 c.age##c.chol cigs_1 sysbp_1 sysbp_2

> wt_1 wt_2 ht i.jobgrade
. sliceplot age chol, sliceat(10 35 65 90) percent

• sliceplot is a new user-written command

 Logit(pr(de
eath)) Logit(pr(dea
L ath))
-5 -4 -3 -2 -1 -5 -4 -3 -2 -1

40
45
5

50
age
chol

55
10

60
65
15

Pr(death) Pr(death
h)
0 .05 .1 .15 .2 .25 0 .05 .1 .15 .2 .25
0

40
45
5

50
age
chol

55
10

60
65
15
intn
Graphical presentation of age x chol intn.

37
Check of chol x age interaction

2 Q1: slope 0.17 (SE 0.06) Q2: slope 0.22 (SE 0.05)

2
0

0
-2

-2
-4

-4
-6

-6
Logit(pr(death))
-8

-8
0 2.5 5 7.5 10 12.5 0 2.5 5 7.5 10 12.5

Q3: slope 0.14 (SE 0.04) Q4: slope -0.01 (SE 0.04)
2

2
0

0
--2

--2
-4

-4
-6

-6
-8

-8

0 2.5 5 7.5 10 12.5 0 2.5 5 7.5 10 12.5

38
chol
 Interactions with continuous
covariates in randomized trials

39
MFPI method (Royston & Sauerbrei 2004)

• Continuous covariate x of interest

interest, binary
treatment variable t and other covariates z
• Independent of x and t, t use MFP to select an
‘adjustment’ (confounder) model z* from z
• Find best FP2 function of x (in all patients)
adjusting for z* and t
• Test
est FP2(x)
( ) × t interaction
te act o (2( dd.f.))
• Estimate β’s in each treatment group
• Standard test for equality of β β’ss
• May also consider simpler FP1 or linear
functions – choose e.g.g byy min AIC
40
MFPI in Stata

• MFPI is implemented as a user command,

command
mfpi
• mfpi is available on SSC
• Details are given by Royston & Sauerbrei,
Stata Journal 9(2): 230-251
230 251 (2009)
• Program was updated in 2012 to support
acto variables
factor a ab es

41
Treatment effect function

• Have estimated two FP2 functions – one per

treatment group
• Plot the difference between functions against x
to show the interaction
• i.e.
i e the treatment effect at different x
• Pointwise 95% CI shows how strongly the
te act o is
interaction s supported
suppo ted at different
d e e t values
a ues of
o x
• i.e. variation in the treatment effect with x

42
Example: MRC RE01 trial in kidney cancer

• Main analysis: Interferon improves survival

• HR: 0.76 (0.62 - 0.95), P = 0.015
• Is
I the
h treatment effect
ff similar
i il iin all
ll patients?
i ?
• Nine possible covariates available for the
investigation of treatment-covariate
interactions – only one is significant (WCC)

43
 Kaplan Meier showing treatment effect
Kaplan-Meier
00
1.0
(1) MPA
(2) Interferon
75
0.7
alive
50
portion a
0.5
25
Prop
0.2
00
0.0

At risk 1: 175 55 22 11 3 2 1

At risk 2: 172 73 36 20 8 5 1

0 12 24 36 48 60 72
44
Follow-up (months)
 The mfpi command

. mfpi, select(0.05) fp2(wcc) with(trt) gendiff(d): stcox

(whod1 whod2) t
t_dt
dt t
t_mt
mt rem mets haem
Interactions with trt (347 observations). Flex-1 model (least flexible)

-------------------------------------------------------------------------------
Var Main Interact idf Chi2 P Deviance tdf AIC
-------------------------------------------------------------------------------
wcc FP2(-1 -.5) FP2(-1 -.5) 2 6.91 0.0316 3180.194 7 3194.194
-------------------------------------------------------------------------------
idf = interaction degrees of freedom; tdf = total model degrees of freedom

. mfpi_plot wcc
[ i
[using variables
i bl created
d b
by gendiff(d)]
diff(d)]

45
 Treatment effect plot for wcc

2
1
Log relativve hazard
-1 -2
2
-3 0

5 10 15 20 25
White cell count

About 25% of patients, those with WCC > 10 seem not to

benefit from interferon 46
Concluding remarks

• MFPIgen and MFPI should help researchers

detect, model and visualize interactions with
continuous covariates
• Usually, we are searching for interactions, so
small P-values are required
q
• Other methods not considered
•S
STEPP – mainly
a y graphical
g ap ca
•…

47
Thank you
you.

48
Cox (1984) paper: Interaction

• Cox identifies 3 types of variable that might

appear in interactions:
• Treatment variables
• Can be modified or imposed
• Treatments,
Treatments e.g.
e g chemotherapy,
chemotherapy surgery
• Behaviours, e.g. smoking, drinking
• Intrinsic variables
• Cannot be modified
• Often demog
demographic,
aphic e.g.
e g sex,
se age
• Unspecific variables
• e.g. structural
t t l bl
blocks,
k `random’
d ’ factors
f t
49