I have worked out the risk ratio of different doses of a drug relative to a placebo and want to find out if there is a dose-dependent increase in an outcome/effect I am looking at. I have the (for example) following risk ratios and Confidence intervals:

5 mg: 2.36 (1.56, 4.39) 10 mg: 3.53 (2.04, 5.68) 15 mg 2.22 (1.84, 3.07)

When describing/interpreting the results would it be appropriate to say from 5mg to 10mg a dose-dependent increase in the effect is seen, but the same is not seen from 10 mg to 15 mg?


  • 2
    $\begingroup$ Do you have more than three studies? I think you may be better to work on the log scale as well. $\endgroup$
    – mdewey
    Apr 10, 2016 at 12:46
  • 1
    $\begingroup$ For some doses I have 4 studies, some 3 and there are few doses that are only looked at by 1/2 studies.. $\endgroup$
    – Harose
    Apr 10, 2016 at 12:54
  • 1
    $\begingroup$ Has each study only examined one dosage (compared to placebo) or do you have studies that compared two or more dosages against placebo? $\endgroup$
    – Wolfgang
    Apr 10, 2016 at 13:06
  • 1
    $\begingroup$ No all of the studies I have included have examined 2/more dosages against placebo. Some studies have examined the same dosages so I have pooled the results together, as some dosages were only examined in e.g. 1 trial I only have data from 1 trial for that specific dosage. (I hope that is clear) $\endgroup$
    – Harose
    Apr 10, 2016 at 13:42
  • 1
    $\begingroup$ It is clear but I would question your approach here. You are ignoring the existence of the studies and the correlation between estimates from the same study. I think you need to investigate a multivariate approach or network meta-analysis. $\endgroup$
    – mdewey
    Apr 10, 2016 at 14:20

1 Answer 1


I agree with @Wolfgang and @mdewey that you need to clarify better your goals, and that a multivariate/network approach is most precise.

In case you want to pursue this further, you can refer to the sample R code below, exploiting the netmeta package, and which hypothesizes a dataframe of 10 studies each with 4 arms (Rx 1 [eg placebo], Rx 2, Rx 3, and Rx 4), providing you pairwise as well as network estimates.


id <- c(1:10)
treatment1 <- 1 # eg placebo
treatment2 <- 2
treatment3 <- 3
treatment4 <- 4
events1 <- sample(25:100, 10, replace = T)
events2 <- sample(50:150, 10, replace = T)
events3 <- sample(75:200, 10, replace = T)
events4 <- sample(75:200, 10, replace = T)
patients1 <- sample(300:1000, 10, replace = T)
patients2 <- sample(300:1000, 10, replace = T)
patients3 <- sample(300:1000, 10, replace = T)
patients4 <- sample(300:1000, 10, replace = T)
mydata1 <- data.frame(cbind(id, treatment1, treatment2, treatment3, treatment4, events1, events2, 
                            events3, events4, patients1, patients2, patients3, patients4))

pairwise1 <- pairwise(list(treatment1, treatment2, treatment3, treatment4), list(events1, 
                      events2, events3, events4), list(patients1, patients2, patients3,
                      patients4), studlab=id, sm = "RR", data = mydata1)

netmeta1 <- netmeta(TE, seTE, treat1, treat2, studlab, data=pairwise1, sm = "RR", comb.random = TRUE)

netgraph(netmeta1, points=TRUE, cex.points=3, cex=1.25)

forest(netmeta1, ref="1", xlim=c(0.1, 10))

netrank1 <- netrank(netmeta1, small.values="bad")

  • $\begingroup$ If the answer is fine remember to flag it as correct... $\endgroup$ Apr 11, 2016 at 14:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.