# How to estimate standard error of journal impact factor based on impact factor and number of citable papers?

The impact factor (and other indicators of citations per document per year) is often used (controversially) to evaluate the research impact of a journal.

Critics of the impact factor often point out that it based on a skewed distribution with outliers. For example, many papers may have zero citations, another set may have a small number of citations, and finally a small set of papers may have a large number of citations (e.g., 50 or 100, etc.). In general, I've read that this profile is typical of many mid-tier journals. Thus, this is used to suggest that the impact factor may be influenced by outliers.

I agree that this distribution of citations does highlight that the impact factor overestimates the median citations per paper (as is the case with any positively skewed count distribution).

But another critique is that the impact factor is unreliable because of this distribution.

Thus, my question is:

What formula can be used to calculate standard error of the impact factor for a journal?

Presumably, this would help both critics and advocates of the impact factor to be able to quantify the uncertainty in estimating the impact factor. It would also assist in assessing whether it is appropriate to say that the differences in impact factor over time or between journals are a reliable difference.

Initial thoughts:

• The standard error of the mean is commonly estimated as: s / sqrt(n), where s is the standard deviation of citations per article, and n is the number of citable articles. So presumably, the main factor that would reduce the standard error would be the number of citable articles published in the journal over the period. But unless we have the raw data, we wont have the standard deviation and the skewed distribution might lead this formula to underestimate the standard error.
• Use existing literature to determine the distribution of citations over articles. Hopefully, its a distribution that can be approximate by knowing only the mean, or where the mean provides a good estimate of the standard deviation
• Then apply a formula to get the standard error of the impact factor.
• Why the down vote? Honestly, curious. Impact factors are relevant to academics. Presumably, being able to quantify the standard error of measurement would be beneficial both for people who like them and those that do not. For example, for those who are critical of impact factors, they could say "look, the difference between 1.7 and 2.2 is not reliable"; but that requires them to have some estimate of the standard error. Such information would presumably be useful in promotion, grant, and other contexts, where you are making an argument about research impact. – Jeromy Anglim Jun 22 '17 at 1:35
• I wonder if journals provide their own raw data that will have a huge impact on their publicity policies. – Mithun Jun 22 '17 at 2:31
• @Mithun you could probably get raw data from scopus or web of science fairly easily for one journal. You'd just need to work out how to do the correct dates for publication and citation windows. – Jeromy Anglim Jun 22 '17 at 2:57
• Oh, I forgot. Another issue will arise as: when we try to obtain the citation count for an individual article within a time frame, how the citations outside this time will reflect the std? I know there is overall citations count and last five years citations count. – Mithun Jun 22 '17 at 3:29
• @AlexanderWoo that is not true. Easiest counter example: the sampling distribution of regression coefficients in a linear regression model follows a t-distribution if the null hypothesis is true. More generally, the standard error is used to characterize the spread of the sampling distribution. You can use that for any sampling distribution. If you want to derive more from that (e.g. confidence intervals) you need to be careful to correctly apply the appropriate formula, but it is always good advise to be careful and do things correctly... – Maarten Buis Jun 22 '17 at 8:53