'adjusting for bias' in a Poisson model

  PS I would be somewhat concerned with your number of
random-effects levels (4 for group 1, 2 for group 2).  It's
not at all surprising that you got an estimated value
of zero variance for group 2 ... there are various schools
of thought on this, but I would suggest that you might
alternatively try (1) fitting the grouping variables as
fixed effects and (??) (2) Bayesian analysis with appropriate
non-informative priors on the variances [this is a big
can of worms, though].

  [Forwarding back to the list]

  Ben Bolker


Henrik Parn wrote:

Thanks a lot Ben for your very rapid answer! I will try it out!

Cheers,

Henrik

Ben Bolker wrote:

  "offset" is you want to look for.
  offsets apply on the scale of the linear predictor
(log scale in this case), so I think adding something
like offset=((x==1)*log(0.8)) to the model will do
the trick.

Henrik Parn wrote:

Dear all,

Basically, I have fitted a mixed model with poisson error to analyse how
number of offspring (y) depend on a fixed factor (x). The data is
grouped by two random factors (gr1, gr2).

# Some test data with at least a similar structure:

set.seed(100)
test.data <- data.frame(
y = c(rpois(40, 5.5), rpois(40, 3.5)),
x = factor(rep(0:1, each = 40)),
gr1 = factor(rep(1:4, each = 10)),
gr2 = factor(rep(1:2, each = 5)))

# The model
model <- lmer(y ~ x + (1|gr1) + (1|gr2),
data = test.data, family = poisson)

summary(model)
Generalized linear mixed model fit by the Laplace approximation
Formula: y ~ x + (1 | gr1) + (1 | gr2)
   Data: test.data
   AIC   BIC logLik deviance
 61.01 70.54 -26.51    53.01
Random effects:
 Groups Name        Variance   Std.Dev.
 gr1    (Intercept) 0.00045437 0.021316
 gr2    (Intercept) 0.00000000 0.000000
Number of obs: 80, groups: gr1, 4; gr2, 2

Fixed effects:
            Estimate Std. Error z value Pr(>|z|)
(Intercept)  1.75331    0.06666  26.302  < 2e-16 ***
x1          -0.48660    0.10665  -4.563 5.05e-06 ***


Thus, y is significantly higher for x = 0 than for x = 1. However, it
has been suggested that the estimate of x may be biased (for biological
reasons) and actually be less negative than what is found above.
Specifically, even in absence of an effect of x, the bias could cause y
for x = 1 to be 20% lower than y for x = 0.
(y for x=0 - y for x=1)/ y for x=0 = 20%; y for x=1 could be 0.8*y for
x=0 due to bias.

Thus, the estimate of x would be (1.75331 - 0.2 * 1.75331) - 1.75331 =
-1.75331 * 0.2 = -0.350662 just due to the bias.

I wish to test if there is an effect of x on y over and above the
potential 20% bias.

My naive starting point: Instead of testing if the estimate of x
(-0.48660) differs from zero, I would need to test if -0.48660 -
(-0.350662) = -0.135938 differs fro zero.

Or could I somehow make the adjustment already in the data set, e.g.
adjusting the y's for x = 1? But I assume that I cannot 'just add ?? %
to y in group x', because the response has to be integers?

Does anyone have a suggestion of a convenient way of performe test for
significance of x on y while taking the bias into account?


Thanks a lot in advance!