Hi John,
Your interpretation of the model is correct. However, I'm not sure
about the random terms - just to be sure, there are multiple
observations per wardn? With a typical zero-altered model the
random term would be trait:wardn which assumes the between ward
variance is the same for both processes and the correlation between
them is 1. Your model (which is equivalent to idh(trait):units)
assumes a correlation of 0 and different variances. Reality probably
lies somewhere between these two extremes. You might want to see if
the fixed effect coefficients are sensitive to this, and perhaps
even estimate all relevant parameters (us(trait):wardn) if you have
a lot of data. Perhaps try that and report back?
Cheers,
Jarrod
Quoting "Hodsoll, John" <john.hodsoll at kcl.ac.uk> on Thu, 7 Nov 2013
15:26:06 +0000:
Dear all
I am wondering if anyone can help me in interpreting a zero added
model using MCMCglmm. I am analysing a clinical trial for counts of
incidents on a psychiatric ward (per work shift). The data has a
surfeit of zeros and so I am using zero inflated models. The
problem I have is trying to understand what zero added models is
telling me about the zero inflation. I've looked through the
excellent course notes from Jarrod Hadfield but am a bit unsure as
to the take home message as this is the first time I've attempted
to use these models.
Model background: Outcome data is collected at the ward level (i.e.
not individual patient) and so a hurdle model seemed the most
appropriate, i.e. each ward has the potential to generate an
incident on any given shift. I have used the zero altered models to
test for inflation as on p109 of the course notes. In this
(simplified analysis with just a quick test run) I have included
all factors as predictors for both parts of the model; trial
phase: period.x (baseline vs outcome) and experimental condition
expconr (control vs test). Here is my model specification
cf.za.1 <- MCMCglmm(totflct ~ -1 + trait*(expcon.r*period.x),
data = sw.df, family = "zapoisson",
random = ~idh(at.level(trait,2)):wardn +
idh(at.level(trait,1)):wardn,
rcov = ~ trait:units,
#prior = zza.prior,
#nitt = 250000, burnin = 50000, thin = 500,
verbose = TRUE, pr = TRUE, pl = FALSE, saveXL = TRUE)
The outcome I'm interested in is the change between control and
treatment from baseline to outcome, highlighted as the interaction
term in the model below. For shifts with events there is a
reduction in the rate of events for the intervention versus control
shown by the negative coefficient for the expcon.r x period.x.
However, for the zero inflation test this co-efficient is positive.
Just to confirm, does this mean I have zero deflation for the test
condition in the outcome phase relative to the control condition,
i.e. more shifts with incidents.
post.mean l-95% CI u-95% CI eff.samp
trait:units 0.4641 0.4317 0.4947 116.3
Location effects: totflct ~ -1 + trait * (expcon.r * period.x)
post.mean l-95% CI
u-95% CI eff.samp pMCMC
traittotflct 1.395460 1.195803
1.602353 1000.0 <0.001 ***
traitza_totflct 1.012971 0.742179
1.318166 468.3 <0.001 ***
expcon.rtest 0.052641 -0.210311
0.327396 894.5 0.690
period.xoutcome -0.170481 -0.251931
-0.103334 567.0 <0.001 ***
expcon.rtest:period.xoutcome -0.157615 -0.269555
-0.051604 513.6 0.004 **
traitza_totflct:expcon.rtest -0.316590 -0.762917
0.150063 748.1 0.174
traitza_totflct:period.xoutcome -0.189739 -0.345773
-0.059751 162.4 0.008 **
traitza_totflct:expcon.rtest:period.xoutcome 0.237426 0.001023
0.450208 166.0 0.034 *
I find this a bit odd, but then you would expect more zeros for a
condition with a lower mean count in 1 condition relative to the
other so that would reduce zero inflation? If anyone has any
insight it would be much appreciated.
Thanks
John
====================================
John Hodsoll
Institute of Psychiatry
Kings College London
London
SE5 8AF
[[alternative HTML version deleted]]