Introduction Open access (OA) to the scientific literature means the removal of barriers (including price barriers) from accessing scholarly work. There are two parallel “roads” towards OA: OA journals and self-archiving [ 1, 2]. OA journals make published articles immediately freely available on their Web site, a model mostly funded by charges paid by the author (usually through a research grant). The alternative for a researcher is “self-archiving” (i.e., to publish in a traditional journal, where only subscribers have immediate access, but to make the article available on their personal and/or institutional Web sites [including so-called repositories or archives]), which is a practice allowed by many scholarly journals. OA raises practical and policy questions for scholars, publishers, funders, and policymakers alike, including what the return on investment is when paying an article processing fee to publish in an OA journal, or whether investments into institutional repositories should be made and whether self-archiving should be made mandatory, as contemplated by some funders [ 3]. Among the arguments of OA proponents (and an expectation of scientists who publish OA articles) is that “open” work is more quickly recognized, as measured by citations. Critics of OA dispute this fact and argue that there is “no evidence that this will happen.” [ 4] Representatives of traditional publishers argue that the “established system of scientific/technical/medical publishing provides excellent levels of open access to scientists and the public alike,” implying that scientists have access to the literature anyway and that there would be little advantage to publish OA. [ 5] In fact, the evidence on the “OA advantage” is controversial. Previous research has based claims of an OA citation advantage mainly on studies looking at the impact of self-archived articles or articles that are found online (“openly accessible,” which some have argued to be different from open access in the narrower sense [ 6]). Most studies show an association between being online and being cited more often [ 1, 7– 9], although another study in the field of pediatrics seemed to suggest the opposite [ 10]. All these previous studies are cross-sectional and are subject to numerous limitations. The first problem is self-selection. As most of these previous studies broadly define OA as “being found freely available online,” [ 7, 9] alternative explanations for citation differences include that important (high-citation) articles are more likely to be posted online by authors or users as a result of the articles' importance; for example, because they are used for journal clubs [ 6] or coursework, or because authors post them on their homepages because they get so many requests from peers (Wren found that online accessible papers are clearly biased towards publications with “higher popular demand” [ 6]). In other words, one could argue that the articles are online because they are highly cited, rather than being highly cited because they are online. A mere association in a cross-sectional study tells us nothing about the direction of the relationship. Kurtz even argues that “the claims that the citation rate ratio of papers openly available on the internet versus those not available is caused by the increased readership of the open articles…(“OA advantage”) are somewhat overstated.” [ 11] Similarly, while the usual line of argument is that self-archiving leads to higher citations [ 8], alternative explanations include that top authors are more likely to be at top institutions that may be more likely to have an institutional repository, which smaller institutions do not have, or that authors selectively self-archive their best work as a “trophy.” [ 6] A recent analysis of articles published in four mathematics journals indicates that articles deposited in the arXiv ( http://arXiv.org) received more citations than nondeposited articles, but the authors do not attribute OA as the cause of more citations, but self-selection (quality differential) [ 12]. Secondly, especially in fields like physics, where pre- and post-publication on http://arXiv.org is quasi-standard, a relationship between self-archiving and higher citation may be due to other factors, such as earlier dissemination of results through preprints [ 11], a quality improvement through discussion of preprints [ 13, 14], or an “outsider” position of authors who do not self-archive. Thirdly, previous studies reported crude, unadjusted rate ratios, where differences in author and article characteristics between OA and non-OA publications were not taken into account and corrected for. One could argue that the observed citation advantages of self-archived papers are a result of confounders; for example, publications with more authors are more likely to be self-archived (as it takes only one author to self-archive) and are also (independently from any OA effect) cited more often (e.g., through increased self-citations or because they might be of higher quality). Limited or no evidence is available on the citation impact of articles originally published as OA that are not confounded by the various biases and additional advantages of self-archiving or “being online” that contribute to the previously observed OA effects. A “journal-level” analysis of journal impact factors concluded that OA journals are more often in the lower half of their subject category, although within the collection of OA titles, these journals ranked higher by immediacy index than by impact factor [ 15]. However, comparing the impact of OA journals against non-OA journals ignores differences in the journals' novelty, editorial policies, quality of peer review, and acceptance policies, which are strong confounders that are difficult to adjust for. To answer the question of whether OA publications lead to a citation advantage I chose an article-level approach, comparing the bibliometric impact of a cohort of articles from the same journal (Proceedings of the National Academy of Sciences [PNAS]) that offers both an OA and a non-OA publishing option, adjusted for different article and author characteristics. Results Article and Author Characteristics A total of 1,492 original articles were included: 212 (14.2% of all articles) were published as immediate OA articles on the journal site, and 1280 (85.8%) as non-OA articles. On December 31, 2004, the articles in the cohort had been published (in most cases, electronically before print publication) within the last 194 d (mean, 101 d; SD = 57.5), with OA articles being on average younger (83.6 d; SD = 50.2) than non-OA articles (104.0 d; SD = 58.1) ( p 0.05). Citations Crude analysis. In the crude analysis, the mean number of citations as well as the proportion of articles cited at least once was significantly higher in the OA group in both the April 2005 and the October 2005 analyses, with the relative “risk” for non-OA articles of not being cited increasing over time ( Table 2). In an analysis stratified by subject, there was a trend for an OA citation advantage in almost all subjects, although, due to the limited number of articles per subject area, few of these differences reached statistical significance (unpublished data). Adjusted analyses. A number of potential confounders must be considered and adjusted for to correct for differences in the number of authors, past productivity (or author seniority), time since publication, and submission track, which differ between the article groups and could be independently related to the probability of getting cited. As shown above, in the crude analysis, the proportion of uncited articles differed significantly between the groups at all three analysis points ( Table 2). In order to determine whether these differences remained significant when adjusted for potential confounders, a logistic regression predicting “cited” status as dependent variable and with stepwise backwards elimination of potential predictors and confounders that were not statistically significant was conducted, controlling for first and last author's lifetime publication count, first and last author's lifetime average citations per paper , number of days since publication (categorized), number of authors (categorized), country of the corresponding author (12 most common countries and “other”), funding type, subject area (14 most common subjects and “other”), and submission track. OA status remained an independent predictor for being cited for all three analysis points, with an increasing odds ratio over time in favor of OA articles ( Table 3). Similarly, using a stepwise backward linear regression model including the same covariates, OA status remained a significant independent predictor for the number of citations (transformed on a logarithmic scale) both in the April 2005 analysis (beta for OA status = 0.187; p < 0.001; overall model fit r 2 = 0.21) as well as in the October 2005 analysis (beta for OA status = 0.263; p < 0.001; overall model fit r 2 = 0.15). Secondary Analysis PNAS allows authors to “self-archive” their research on the Internet even if they choose the non-OA option. This means that some of the articles in the non-OA group may in fact have been “openly accessible” online through the author's homepage or an institutional repository. In a secondary analysis I also analyzed citations of “self-archived OA” articles (i.e., self-archived or otherwise openly accessible on other Web sites than http://www.pnas.org or http://pubmedcentral.org), with the explicit caveat that articles which are found on the Internet are subject to self-selection and other biases as discussed in the introduction (i.e., it is impossible to discriminate whether they are on the Internet because they are important, or whether they are highly cited because they are on the Internet). Citation rates (as of October 2005) of four separate subgroups were analyzed, as an article could be either published under the PNAS immediate OA option or “self-archived,” or both, or none. There was a clear relationship between the level of openness and the citation levels ( Table 4). While 36 of 212 (17.0%) of immediate journal OA articles were also self-archived, only 121 of 1,280 (10.6%) of non-OA papers on the journal site were self-archived (i.e., papers published originally as OA were more likely to be self-archived [Fisher's exact test p = 0.002]). The 1,159 papers which were neither self-archived nor immediate journal OA articles had on average 4.4 citations, whereas the 334 papers which were either self-archived or published originally as OA (or both) had 5.9 (Z = 4.215, p < 0.001) citations. The risk of not being cited for papers which were either published originally as OA or self-archived was only 6.9%, while it was 13.81% for articles in the non-OA group (neither self-archived nor published originally as OA; relative risk = 2.0 [1.31–3.04]; Fisher's exact test p < 0.001). While in the crude analysis self-archived papers had on average significantly more citations than non–self-archived papers (mean, 5.46 versus 4.66; Wilcoxon Z = 2.417; p = 0.02), these differences disappeared when stratified for journal OA status ( p = 0.10 in the group of articles published originally as non-OA articles, and p = 0.25 in the group of articles published originally as OA). In a logistic regression model with backward elimination, which included original OA status and self-archiving OA status as separate independent variables as well as all potential confounders, self-archiving OA status did not remain a significant predictor for being cited. In a linear regression model, the influence of the covariate “article published originally as OA, without being self-archived” (beta = 0.250, p < 0.001) on citations remained stronger than self-archiving status (beta = 0.152, p = 0.02). Discussion Main Findings This comparison of the impact of OA and non-OA articles from the same journal in the first 4–16 mo after publication shows that OA articles are cited earlier and are, on average, cited more often than non-OA articles. To my knowledge, this is the first longitudinal study of a cohort of OA and non-OA articles providing direct and strong evidence for preferential or earlier citation of articles published originally as OA. It is also the first study showing an advantage of publishing an article as OA on the journal site over self-archiving (i.e., making the article otherwise online accessible). The strength of the OA effect is particularly surprising because PNAS is a widely available journal that is accessible for most researchers through their library. In addition, articles are made freely available to nonsubscribers 6 mo after publication. The effect of OA publishing may be even higher in fields where journals are not widely available and where articles from the control group remain “toll-access.” Limitations This study offers only a short-term glimpse at what happens at the left-hand (early) side of the temporal citation curve. A citation curve plotting over time the number of new citations per year to articles published in a certain year would show a sharp increase of new citations with a citation peak after 1 or 2 y (which is typically the time needed for citing authors to prepare and publish their papers), with a slow and steady decline of new citations year after year. Despite the narrow observation window of this study, it appears that publishing OA does not merely lead to a steeper increase and “left-shift” of the citation curve by 6 mo, because such a left-shift is incompatible with the observation in this study that there are still increased citation rates for OA articles at our third observation point 10–16 mo after publication. Rather, there seems to be a sustained effect on the absolute number of citations. In other words, there seems to be not only an advantage in terms of immediacy (defined as the average number of times that an article, published in a specific year within a specific journal, is cited over the course of the same year), but also in terms of total impact (as measured by the absolute number of citations received over a longer period of time). Future follow-up analyses following this cohort over a number of years will provide a more complete picture on how long-lasting the citation advantage of OA articles is. As this was an observational study and not a randomized trial, we were able to statistically control only for known confounders. There is a possibility of selection bias in authors judging the importance of their work and making a deliberate decision to publish their most important work as OA, with quality differences between articles contributing to citation differences. However, results from our author survey—in which we asked authors to self-rate the importance and quality of their work—did not show significant differences between the groups and do not make this likely, in particular because PNAS is a high-impact journal and most of the authors considered their work high quality in the first place. PNAS has a broad interdisciplinary scope and different submission tracks, which are features that should enhance the generalizability of the results. However, it has to be acknowledged that our data come from a single, rather atypical journal, and should be replicated with data from other newer hybrid (“author-choice OA”) journals. Other publishers have started to offer an author-choice option, but only recently and with limited sample sizes both in terms of articles and in terms of citations. While a study on these journals is under way, these results will be available only in a few years. This study used citations as a proxy for impact, and some may argue that “it is hard to see how science will benefit by increased citation rates [of OA articles].” [ 4] Our data do suggest that OA articles are more quickly recognized and their results are picked up and discussed by peers to a larger extent. It is hard to see how faster and increased utilization and uptake of research results will not benefit science, at least in terms of accelerating the speed by which new results are verified, falsified, or built upon by others. By focusing on citations this study only addresses the impact on other research users, not on the knowledge user (i.e., policymakers, consumers, or health professionals), but it can be hypothesized (and should be tested in future studies) that there is also a “knowledge translation advantage” in terms of increased and accelerated knowledge uptake by consumers and policymakers. Conclusions OA journals and hybrid journals like PNAS, as well as traditional publishers like Blackwell Publishing (“Online Open”), Oxford University Press (“Oxford Open”), and Springer (“Springer Open Choice”) are now offering authors an immediate OA option if the author pays a fee. Researchers, publishers, and policymakers confronted with the question of whether or not to invest in OA publishing have reason to believe that OA accelerates scientific advancement and knowledge translation of research into practice. While more work remains to be done to evaluate citation patterns over longer periods of time and in different fields and journals, this study provides evidence and new arguments for scientists and granting agencies to invest money into article processing fees to cover the costs of OA publishing. It also provides an incentive for publishers seeking to increase their impact factor to offer an OA option. The findings indirectly also support policies of granting agencies which made (or consider to make) OA publishing (be it only through self-archiving) mandatory for grantees [ 3], as it illustrates the advantage of openess in the dissemination of knowledge. However, this study suggests that publishing papers as OA articles on the journal site facilitates knowledge dissemination to a greater degree than self-archiving, presumably because few scientists search the Internet or Google for articles if they have encountered an access problem on the journal Web site. Materials and Methods Article cohort. PNAS announced on June 8, 2004, that authors could pay US$1,000 if they wanted their article to be immediately OA [ 16] (as opposed to the usual non-OA “moving wall” model, where articles become freely accessible after 6 mo). The resulting mix of OA and non-OA articles published between June 8, 2004 (page 8745 of Volume 101), and December 20, 2004, in PNAS constituted the article cohort. Included were the article types “Journal Article,” “Clinical Trials,” and “Case Reports,” excluding editorials, commentaries, biographies, retractions, and retracted articles, “classics” papers, and supplement articles. To determine citation rates, Web of Science (Thomson ISI, Stamford, Connecticut) was searched on December 31, 2004, April 4, 2005, and again on October 4, 2005, for all articles citing papers from the cohort. Citation errors (e.g., wrong volume or misspelling of the author name) were manually corrected. The following article characteristics were extracted and used in the multivariate models in order to control for potential confounders: days since publication, number of authors, article type, country of the corresponding author, funding (as indexed in Medline), subject area (as classified in the PNAS Table of Contents), and submission track. PNAS has three different submission tracks: the majority (80%) of submissions are made through a regular submission track (called “track II”), where authors submit manuscripts to the editorial office, which assigns an Academy member as editor to guide the paper through peer review [ 17]. PNAS has two further unique peer-review and submission models, called track I (authors submit their work directly to an Academy member, who solicits two reviews and then sends it to the editorial office for publication) and track III (Academy members can send their own articles to the editorial office, accompanied by peer-review reports which they themselves solicited) [ 18]. Including the submission track into the multivariate models allows control of different levels of rigor in peer review and quality of the contributions. To control for possible differences in “author quality,” information on the first and last authors' total number of published articles and lifetime citations up to 2004 was gathered from Web of Science and was also included as adjustment variables in the multivariate models. In addition, in order to test equivalence of article quality between the OA and non-OA groups, all authors with published and valid e-mail addresses ( n = 313) received a Web-based survey in which they were asked to rate the relative urgency, importance, and quality of their particular PNAS article relative to other articles on a five-point Likert scale. Secondary analysis. To identify “self-archived” articles I used a computer program (in-house development, see Acknowledgments) that conducted for each article a Google ( http://www.google.com) phrase search, with the first sentence from the results section of the article or, if this did not lead to any hits, the digital object identifier as query. An article was considered “self-archived” if the article was found on Web sites other than on http://www.pnas.org or http://pubmedcentral.org. While this is not a perfect method (as Web coverage by search engines is incomplete [ 19]), it is currently the best we have, and if an article is not accessible through Google, it may not be found by the research community anyway and may not be considered “openly accessible.” Statistics. For crude comparisons of continuous variables (number of days since publication, number of authors, and citations), the nonparametric two-sided Wilcoxon Mann-Whitney test was used, as these data were not normally distributed. Proportions were compared using the Fisher's exact test, while categorical variables with multiple categories were compared between the two groups using the Freeman-Halton test (an extension of Fisher's exact test for r by c tables). In multivariate models I tried to predict the number of citations from several article characteristics, including OA status, adjusted for potential confounders as independent variables. In a linear regression model, the number of citations as the dependent variable was transformed into a log scale, as the distribution was skewed, predicting log(citations+1). In a logistic regression model, the number of citations as the dependent variable was dichotomized into 0 (uncited) and ≥ 1 (cited). All data were analyzed using SAS, version 8.02 (Cary, NC).