#### data source

We searched the published literature for studies of primate species in forest fragments (defined here as 100 km^{2} or less) that provided data on number of primate species in relation to either size of fragment or distance of fragment to adjacent intact forest (defined as greater than 100 km^{2}). Our final sample consisted of 136 fragments at 33 study sites (Table 1; see the Appendix). We omitted sites where only one species occurred in the region and therefore only one was available to be in any fragment, for example at the north edge of species’ ranges in central America and the south edge in Africa (Lawes 1992; Estrada & Coates-Estrada 1996; Estrada *et al*. 1999; Lawes, Mealin & Piper 1999). We also omitted nine sites of one fragment each, because we could not be sure from the original papers that more than one species was present in the fragments. These nine sites are listed and marked in the Appendix.

Table 1. Characteristics of forest fragments that contained primates. Area, *significantly different from each other; no. fragments, number of fragments of < 1 km^{2} and the total number; distance, distance of fragment to main forest block (i.e. block of > 100 km^{2}); –, no data available Region | Area (km^{2}) | No. fragments | Distance (km) | Age (year) | No. species |
---|

Range | Median | < 1 km^{2}/total (%) | Range | Median (*n*) | Range | Median (*n*) | Range | Median |
---|

Globe | 0·01–100 | 1·0 | 72/148 (49) | 0·1–53 | 2 (33) | 1·5–72 | 30 (69) | 1–7 | 3 |

Africa | 0·01–51 | 0·2* | 28/43 (65) | 0·2–7 | 2·1 (22) | 19–43 | 43 (23) | 1–7 | 3 |

Asia | 0·01–68 | 0·9 | 14/28 (50) | 12^{1} | 12 (1) | 9–45 | 34 (5) | 1–6 | 2 |

Madagascar | 0·01–75 | 0·4 | 10/19 (53) | – | – | 6–40 | 30 (18) | 1–7 | 3 |

S. America | 0·01–100 | 2·0* | 20/58 (34) | 0·1–53 | 0·2 (10) | 1·5–72 | 15 (23) | 1–7 | 3 |

Except for studies in Madagascar, where nocturnal primate species comprise a large proportion of the total (45%), nocturnal species were often not mentioned in the sources. The omission will not markedly change counts of species richness in the Americas, where less than 10% of the total complement of primate species is nocturnal. However, in Asia they comprise 15% of the primate fauna, and in Africa 20%. Thus, the Asian and African data must be considered a sample of the diurnal species. Because censuses usually miss species (Grayson & Livingston 1993), we used long-term studies as our sources whenever possible, and therefore the counts of the diurnal species should be largely complete.

Our analyses were limited to subsets of the total sample, because not all categories of data were available for every site or fragment.

#### analysis

The review considered only fragments in which primates were recorded. With more than one fragment per site, there was the possibility of pseudoreplication/spatial autocorrelation through site effects when sites were combined. We reduced the effect by running analyses that omitted the site per continent with the most fragments with the relevant data (area, isolation, age). With area, we also removed all sites with five or more fragments, but a sample size of only nine fragments remained. The slope was positive but not significant. Additionally, the site-by-site analysis (below) obviated pseudoreplication.

Many forms of species–area analysis have been applied in fragmentation and insularization studies (Tjørve 2003). Ordinary least-squares linear regression of double-log or semi-log data are common. However, ordinary least-squares regression is inappropriate for small sample size, because the data are effectively counts and therefore not a continuous distribution (McCullagh & Nelder 1989). Also, in the case of our data, the residuals (errors) did not approximate a normal distribution, whether logged or not (e.g. globe number species by log_{10} area, Shapiro–Wilk *W*= 0·95, *n*= 136, *P*= 0·0007). Poisson regression would be suitable in such cases (McCullagh & Nelder 1989). However, because of the very small number of primate species per fragment (never more than seven), we only used correlation tests on counts. These included non-parametric Spearman correlation tests because of the small sample size for Madagascar (*n* < 20).

We also ran a site-by-site analysis for degree and direction of slope of richness to area. Spearman correlation tests were used for each site. Besides the importance of examination of the effect at the local as well as regional scale, this site-by-site test prevented spatial autocorrelation.

With the global- and continental-level analyses, we ran ordinary least-squares regressions of arcsine square-root proportional richness (the proportion of the full complement of species in the main forest that was in each fragment) against log_{10} fragment area and isolation. These regressions provided more information than correlation tests, and the use of proportions accounted for the fact that the number of species available to be in a fragment could differ across sites (cf. Telleria, Baquero & Santos 2003). Use of proportions was valid because their values were more nearly continuously distributed than counts, and the distribution of errors in the proportions was not statistically different from a normal distribution (e.g. globe percentage species by log_{10} area, Shapiro–Wilk *W*= 0·97, *n*= 63, *P*= 0·41).

Because arguments exist for the use of other than double-log plots, and because of suggestions that the species–area relationship is not necessarily a straight line (Lomolino 2000), we also examined semi-log plots and polynomial functions. We obtained no consistent results as to whether or not polynomials provided a closer fit than a linear regression, and neither they nor semi-log plots changed any relationship from being significant to non-significant or vice versa. We therefore do not provide any of the results here.

We removed outliers in statistical analyses to allow investigation of what the majority of the sampled population was doing. In the same way that it is bad practice to claim a significant relationship that is determined by only a minority of the data, so it is bad practice to suggest no relationship when a relationship is obscured by a minority of the data. Hence outlier detection facilities in statistical programs (below).

When accounting for the interaction of influences (area, continent, isolation, age) on proportional richness, we used analysis of covariance and multiple regression. In regression analyses, the *r*^{2} value that we report is always the adjusted *r*^{2}. We used JMP 5·0·1·2 (SAS Institute Inc. 2002) and Siegel (1956) for statistical analysis. We identified outliers with JMP's Mahalanobis outlier analysis. *P*-values are two-tailed, despite predictions of the direction of effect, but we give significance values up to 0·1.