epitools::oddsratio error numbers close to zero in cells?

You can actually also get results from fisher.test when one of the entries is 0.
One of your confidence limits will just be either 0 or Inf.
But that should hardly be a surprise.

You could also try the twoby2 command from the Epi package which will summarize your
table analysis nicely.

m

? ? [,1] [,2]
[1,] ?360 ? ?0
[2,] ? ?7 ?120

fisher.test(m)

? ? ? ?Fisher's Exact Test for Count Data

data: ?m
p-value < 2.2e-16
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
?1204.706 ? ? ?Inf
sample estimates:
odds ratio
? ? ? Inf

Another approach is to run a logistic regression model. Using the same
data as above:

(dp <- data.frame (a=c(320,7,0,0), b=c(0,0,4,120), t=c(1,0,1,0)))

? ?a ? b t
1 320 ? 0 1
2 ? 7 ? 0 0
3 ? 0 ? 4 1
4 ? 0 120 0

fit <- glm(cbind(a,b) ~ t, binomial, dp)

The estimated odds ratio is here (OR=1371):

exp(coef(fit))

(Intercept) ? ? ? ? ? t
?5.833e-02 ? 1.371e+03


And its confidence interval:

exp(confint(fit))

Waiting for profiling to be done...
                2.5 %    97.5 %
(Intercept)   0.02466    0.1159
t           442.87000 5551.2516


With less unbalanced data, glm() gives results very close to those
from fisher.test()

dp2

a   b t
1 320   0 1
2 200   0 0
3   0 100 1
4   0 150 0

fit2 <- glm(cbind(a,b) ~ t, binomial, dp2)

exp(coef(fit2))

(Intercept)           t
      1.333       2.400

exp(confint(fit2))

Waiting for profiling to be done...
            2.5 % 97.5 %
(Intercept) 1.080  1.650
t           1.766  3.274

fisher.test(matrix(c(320,200,100,150),2))

Fisher's Exact Test for Count Data

data:  matrix(c(320, 200, 100, 150), 2)
p-value = 2.294e-08
alternative hypothesis: true odds ratio is not equal to 1
95 percent confidence interval:
 1.742 3.309
sample estimates:
odds ratio
     2.397



--Philippe
Senior Epidemiologist,
World Health Organization
Geneva, Switzerland