The following text was written a few years ago, but much of it never got published. So, I thought that this might be a good opportunity to make it available, since what it says is still true today.
Since a phylogenetic tree is interpreted in terms of the monophyletic groups that it hypothesizes, it is important to quantitatively assess the robustness of all of these groups (i.e. the degree of support for each branch in the tree) — is the support for a particular group any better than would be expected from a random data set? This issue of clade robustness is the same as assessing branch support on the tree, since each branch represents a clade. Many different techniques have been developed, including:
 analytical procedures, such as interiorbranch tests (Nei et al. 1985; Sneath 1986), likelihoodratio tests (Felsenstein 1988; Huelsenbeck et al. 1996b), and clade significance (Lee 2000);
 resampling procedures, such as the bootstrap (Felsenstein 1985), the jackknife (Lanyon 1985), topologydependent permutation (Faith 1991), and clade credibility or posterior probability (Larget and Simon 1999); and
 nonstatistical procedures, such as the decay index (Bremer 1988), clade stability (Davis 1993), and spectral signals (Hendy and Penny 1993).
Of these, far and away the most popular and widely used method has been the bootstrap technique (Holmes 2003; Soltis and Soltis 2003).
The bootstrap
This method was first introduced by Efron (1979) as an alternative method to jackknifing for producing standard errors on estimates of central location other than the mean (e.g. the median), but it has since been expanded to cover probabilistic confidence intervals as well (Efron and Tibshirani 1993; Davison and Hinkley 1997). It was introduced into phylogenetic studies by Penny et al. (1982) and then formalized by Felsenstein (1985), who suggested that it could be implemented by holding the taxa constant and resampling the characters randomly with replacement, the treebuilding analysis then being applied to each of the bootstrap resamples.
Bootstrapping is a monte carlo procedure that it generates "pseudo" data sets from the original data, and uses these new data sets for its inferences. That is, it tries to derive the population inferences (i.e the "true" answer) from repeated generation of new samples, each sample being constrained by the characteristics of the original data sample. It thus relies on an explicit analogy between the sample and the appropriate population: that sampling from the sample is the same as sampling from the population. Clearly, the strongest requirement for bootstrapping to work is that the sample be a reasonable representation of the population.
Bootstrap confidence intervals are only ever approximate, especially for complex data structures, as they are a fundamentally more ambitious measure of accuracy than is a simple standard error (SE). For example, the usual formula for calculating a confidence interval (CI) when the population frequency distribution is assumed to be normal is: CI = t * SE, where t is the Student tvalue associated with the particular sample size and confidence percentage required. However, the main use of bootstrapping is in situations where the population frequency distribution is either indeterminate or is difficult to obtain empirically, and so this simple formula cannot be applied. Getting from the standard error to a confidence interval is then not straightforward. As a result, there are actually several quite distinct procedures for performing bootstrapping (Carpenter and Bithell 2000), with varying degrees of expected success.
Types of bootstrap
The original technique is called the
percentile bootstrap. It is based on the principle of using the minimum number of ad hoc assumptions, and so it merely counts the percentage of bootstrap resamples that meet the specified criteria. F§or example, to estimate the standard error of a median, the median can be calculated for each bootstrap resample and then the standard deviation of the resulting frequency distribution will be the estimated standard error of the original median. The method is thus rather simplistic, and is often referred to as the
naïve bootstrap, because it assumes no knowledge of how to calculate population estimates. It is a widespread method, as it can be applied even when the other methods cannot. However, it is known to have certain problems associated with the estimates produced, particularly for confidence intervals, such as bias and skewness (especially when the parent frequency distribution is not symmetrical). These were pointed out right from the start (Efron 1979), and efforts have subsequently been made to deal with them. Nevertheless, this is the form of bootstrap introduced by Felsenstein (1985), and it is the one used by most phylogeny computer programs. It is therefore the one that will be discussed in more detail below.
These known problems with the naïve bootstrap can be overcome by using biascorrected (BC) bootstrap estimates — that is, the bias is estimated and removed from the calculation of the confidence interval. Possible dependence of the standard error on the parameter being estimated, which creates skewness, can be dealt with by using
biascorrected and accelerated (BCa) bootstrap estimates, so that the bias and skewness are both estimated and removed from the calculation of the confidence interval. The BCa method is the one usually recommended for use (Carpenter and Bithell 2000), because it corrects for both bias and skewness. This method is much slower to calculate than the simple percentile bootstrap, because it requires an extra parameter to be estimated for each of the bias and skewness corrections, and the latter correction is actually estimated by performing a separate jackknife analysis on each bootstrap resample (which means that the analysis can take 100 times as long as a naïve analysis). There have been several attempts to apply this form of correction methodology to bootstrapping in a phylogenetic context (Rodrigo 1993; Zharkikh and Li 1995; Efron et al. 1996; Shimodaira 2002), but while these can be successful at correcting bias and skewness (Sanderson and Wojciechowski 2000) these have not caught on, possibly because of the time factor involved.
Alternatively, we can decide not to be naïve when calculating confidence intervals, and to therefore calculate them in the traditional manner, using the standard error and the tdistribution. However, we then need to overcome any nonnormal distribution problems of these two estimates by estimating both of them using bootstrapping. That is,
bootstrappedt confidence intervals are derived by calculating both the standard error and the tvalue using bootstrapping, and then calculating the confidence interval as ±t * SE. To many people, this is the most natural way to calculate confidence intervals, since it matches the usual parametric procedure, and thus it is frequently recommended (Carpenter and Bithell 2000). Once again, this method is much slower to calculate than the percentile bootstrap, because the tvalue is actually estimated by performing a separate bootstrap analysis on each bootstrap resample (which means that the analysis can take 100 times as long as a naïve analysis). This methodology seems not to have yet been suggested in a phylogenetic context, and in any case the time factor may be restrictive.
It is also possible to calculate
testinversion confidence intervals. This idea is based on the reciprocal relationship of statistical tests and confidence intervals, where (for example) nonoverlapping 95% confidence intervals indicate statistically significant patterns at p<0.05 and vice versa. Thus, if we work out the situation where the pattern has a probability of p=0.05 of occurring by chance, then this defines the 95% confidence limit of the pattern. Clearly, this can be a complex process, especially for twosided tests (which double the required number of calculations), as it can only be done by iteratively modifying the pattern and recalculating the probability until the solution is reached. Once again, noone yet seems to have suggested this methodology in a phylogenetic context, which is not unexpected given the general problems in deciding how to test branches statistically.
The above methods all count as nonparametric bootstrap methods. More recently,
parametric bootstrapping methods have also been developed, which make the more restrictive assumption that a parametric model can be applied to the data (e.g. that the standard deviation of the parameter can be reliably estimated). In parametric bootstrapping we generate simulated datasets based on the assumed frequency distribution of the data, rather than by resampling from the data set itself. That is, instead of sampling from the sample, we sample from the assumed theoretical distribution to generate the set of bootstrap samples. We can then apply the percentile, BCa or bootstrapt methods, described above, in the usual way. Clearly this method assumes that we know the appropriate frequency distribution; and the method will only be appropriate if this assumption is true, but not otherwise. However, if the assumption is correct, then this can be the most powerful method (Huelsenbeck et al. 1996a; Newton 1996) because it is not dependent on the representativeness of the data sample. The method has been introduced into phylogenetics in several contexts (Goldman 1993; Adell and Dopazo 1994; Huelsenbeck et al. 1996a), but the appropriate frequency distribution for branch support is not obvious (i.e. a phylogeny is a complex structure and cannot be represented by a single number but rather requires a model of sequence evolution and a model tree) and so it is not used for this purpose.
Issues with the bootstrap
Thus, for several reasons, all of the best bootstrapping methods are not likely to be available when assessing the robustness of a phylogenetic tree, and we are left with the naïve percentile bootstrap, which can be expected a priori to provide biased and skewed estimates of confidence intervals (because the frequency distribution associated with tree branches will not be symmetrical). Sadly, these problems have been repeatedly confirmed for the assessment of branch support in phylogenetic treebuilding, both theoretically (Zharkikh and Li 1992a, 1992b; Felsenstein and Kishino 1993; Li and Zharkikh 1994; Sitnikova et al. 1995; Berry and Gascuel 1996; Efron et al. 1996; Huelsenbeck et al. 1996a; Newton 1996; Sanderson and Wojciechowski 2000; Suzuki et al. 2002; Alfaro et al. 2003; Erixon et al. 2003; Galtier 2004; Huelsenbeck and Rannala 2004) and empirically (Sanderson 1989; Hillis and Bull 1993; Buckley et al. 2001; Buckley and Cunningham 2002; Wilcox et al. 2002; Taylor and Piel 2004).

An example of the relationship between naïve bootstrap probabilities and the true probability of a false positive result, showing that percentile bootstrap indices >75% tend to be underestimates of the amount of support while they are overestimates below this level. The graph is based upon 1000 bootstrap resamples of 100 simulated characters for a clade of three taxa plus outgroup (based on data presented by Zharkikh and Li 1992a). The true probability represents the amount of character support for the clade in the simulated data, while the bootstrap probability is the proportion of resamples that included the clade. 
These studies have demonstrated that the probability of bootstrap resampling supporting the true tree may be either under or overestimated, depending on the particular situation. For example, bootstrap values >75% tend to be underestimates of the amount of support, while they may be overestimates below this level, as shown in the first graph (above). That is, when the branch support is strong (i.e. the clade is part of the true tree) there will be an underestimation and when the support is weak (i.e. the clade is not part of the true tree) there will be an overestimation. This situation has been reported time and time again, with various theoretical explanations (e.g. Felsenstein and Kishino 1993; Efron et al. 1996; Newton 1996), although there are dissenting voices (e.g. Taylor and Piel 2004) as would be expected for a complex situation. Unfortunately, practitioners seem to ignore this fact, and to assume incorrectly that bootstrap values are
always underestimates.
Just as importantly, the theoretical studies show that the pattern of over and underestimation depends on (i) the shape of the tree and the branch lengths, (ii) the number of taxa, (iii) the number of characters, (iv) the evolutionary model used, and (v) the number of bootstrap resamples. This was first reported by Zharkikh and Li (1992a), and has been reconfirmed since then. For example, with few characters the bootstrap index tends to overestimate the support for a clade and to underestimate it for more characters. This is particularly true if the number of phylogenetically informative characters is increased or the number of nonindependent characters is increased; and the index becomes progressively more conservative (i.e. lower values) as the number of taxa is increased.
Moreover, these patterns of under and overestimation are increased with an increasing number of bootstrap replications, as shown in the next graph — this called "being wrong, with confidence".

An example of the relationship between the true clade probability and the observed nonparametric bootstrap proportion for two simulated data sets with different numbers of characters (as shown). The lines are based on data presented by Zharkikh & Li, (1995) for 1000 bootstrap resamples of a clade of three taxa plus outgroup. 
The following graph pair of graphs show the effect of varying the evolutionary model used to generate the data, where underspecification of the analysis model leads to a general overestimate of the true probability (crossover at p=0.8, as shown in the first graph of the pair), while matching the generating and analysis models leads to a general underestimation (crossover at p=0.3, as shown in the second graph of the pair).

An example of the relationship between the true tree probability and the difference between the observed percentile bootstrap proportion and the true probability for two simulated data sets. The label in the bottom corner shows the substitution model used to simulate the data, then the model assumed in the bootstrap analysis (the sequence length is 100 nucleotides); JC69 = JukesCantor, GTRG = general time reversible + gammadistributed amongsite rate variation. The points are based on data presented by Huelsenbeck & Rannala (2004). 
These are serious issues, which seem to be often ignored by practitioners. We can't just assume that the "true" support value is larger than our observed bootstrap value. In particular, this means that bootstrap values are not directly comparable between trees, even for the same taxa, and thus there can be no "agreed" level of bootstrap support that can be considered to be "statistically significant". A bootstrap value of 90% on a branch on one tree may actually represent
less support than a bootstrap value of 85% on another tree, depending on the characteristics of the dataset concerned and the bootstrapping procedure used (although within a single tree the values should be comparable).
This complex situation means that we have to consider carefully how best to interpret bootstrap values in a phylogenetic context (Sanderson 1995). The bootstrap proportion (i.e. the proportion of resampled trees containing the branch/clade of interest) has variously been interpreted as (Berry and Gascuel 1996):
 a measure of reliability, telling us what would be expected to happen if we repeated our experiment;
 a measure of accuracy, telling us about the probability of our experimental result being true; and
 a measure of confidence, interpreted as a conditional probability similar to those in standard statistical hypothesis tests (i.e. measuring Type I errors or false positives).
The bootstrap was originally designed for purpose (1), and all of the problems identified above relate to trying to use it for purposes (2) and (3). The values derived from the naïve bootstrap need correcting for purposes (2) and (3), and the degree of correction depends on the particular data set being examined (Efron et al. 1996; Goloboff et al. 2003).
The issue of support values depending on the number of bootstrap replicates is also of interest. It is usually recommended that at least 1,000–2,000 bootstrap resamples are taken for estimating confidence intervals, and this generality has been applied to phylogenetic trees (Hedges 1992). However, it is important to recognize that these suggestions relate to the precision of the confidence estimates not to their accuracy. Accuracy refers to how close the estimates are to the true value (i.e. correctness) while precision refers to how variable are the estimates (i.e. repeatability). Accuracy depends on a complex set of characteristics many of which have nothing to do with bootstrap replication. Precision, on the other hand, is entirely to do with the number of bootstrap replicates and the expected accuracy of the estimates. As shown in the next graph, 100 replicates at a conventional level of accuracy produces estimates that are expected to be within ±4% of the "true" values while 2,000 replicates produces estimates ±1%. This needs to be borne in mind when deciding whether to call a particular value "significant support" or not.

The number of bootstrap replicates needed to achieve a specified amount of precision, given statistical testing at two different levels of probability. For example (as shown by the dotted line), 100 bootstrap replicates means that, if the bootstrap value is accurate at the 95% confidence level, then the estimated bootstrap percentage will be precise to ±4.3%. In order to get ±1% precision then nearly 2,000 bootstrap replicates are needed. 
There have also been attempts to overcome some of the practical limitations of bootstrapping for large data sets by adopting heuristic procedures, including resampling estimated likelihoods for maximumlikelihood analyses (Waddell et al. 2002) and reduced treesearch effort for the bootstrap replicates. However, approaches using reduced treesearch effort produce even more conservative estimates of branch support, and the magnitude of the effect increases with decreasing bootstrap values (DeBry and Olmstead 2000; Mort et al. 2000; Sanderson and Wojciechowski 2000).
References
Adell J.C., Dopazo J. 1994. Monte Carlo simulation in phylogenies: an application to test the constancy of evolutionary rates. J. Mol. Evol. 38, 305309.
Alfaro M.E., Zoller S., Lutzoni F. 2003. Bayes or bootstrap? A simulation study comparing the performance of bayesian markov chain monte carlo sampling and bootstrapping in assessing phylogenetic confidence. Mol. Biol. Evol. 20, 255266.
Berry V., Gascuel O. 1996. On the interpretation of bootstrap trees: appropriate threshold of clade selection and induced gain. Mol. Biol. Evol. 13, 9991011.
Bremer K. 1988. The limits of amino acid sequence data in angiosperm phylogenetic reconstruction. Evolution 42, 795803.
Buckley T.R., Cunningham C.W. 2002. The effects of nucleotide substitution model assumptions on estimates of nonparametric bootstrap support. Mol. Biol. Evol. 19, 394405.
Buckley T.R., Simon C., Chambers G.K. 2001. Exploring amongsite rate variation models in a maximum likelihood framework using empirical data: effects of model assumptions on estimates of topology, branch lengths and bootstrap support. Syst. Biol. 50, 6786.
Carpenter J., Bithell J. 2000. Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. Stat. Med. 19, 11411164.
Davis J.I. 1993. Character removal as a means for assessing the stability of clades. Cladistics 9, 201210.
Davison A.C., Hinkley D.V. 1997. Bootstrap Methods and Their Applications. Cambridge Uni. Press, Cambridge.
DeBry R.W., Olmstead R.G. 2000. A simulation study of reduced treesearch effort in bootstrap resampling analysis. Syst. Biol. 49, 171179.
Efron B. 1979. Bootstrapping methods: another look at the jackknife. Ann. Stat. 7, 126.
Efron B., Halloran E., Holmes S. 1996. Bootstrap confidence levels for phylogenetic trees. Proc. Nat. Acad. Sci. U.S.A. 93, 70857090.
Efron B., Tibshirani R.J. 1993. An Introduction to the Bootstrap. Chapman & Hall, London.
Erixon P., Svennblad B., Britton T., Oxelman B. 2003. Reliability of bayesian probabilities and bootstrap frequencies in phylogenetics. Syst. Biol. 52, 665673.
Faith D.P. 1991. Cladistic permutation tests for monophyly and nonmonophyly. Syst. Zool. 40, 366375.
Felsenstein J. 1985. Confidence limits on phylogenies: an approach using the bootstrap. Evolution 39, 783791.
Felsenstein J. 1988. Phylogenies from molecular sequences: inference and reliability. Annu. Rev. Genet. 22, 521565.
Felsenstein J., Kishino H. 1993. Is there something wrong with the bootstrap on phylogenies? A reply to Hillis and Bull. Syst. Biol. 42, 193200.
Galtier N. 2004. Sampling properties of the bootstrap support in molecular phylogeny: influence of nonindependence among sites. Syst. Biol. 53, 3846.
Goldman N. 1993. Statistical tests of models of DNA substitution. J. Mol. Evol. 36, 182198.
Goloboff P.A., Farris J.S., Källersjö M., Oxelman B., Ramırez M.J., Szumik C.A. 2003. Improvements to resampling measures of group support. Cladistics 19, 324332.
Hedges S.B. 1992. The number of replications needed for accurate estimation of the bootstrap P value in phylogenetic studies. Mol. Biol. Evol. 9, 366369.
Hendy M.D., Penny D. 1993. Spectral analysis of phylogenetic data. J. Classific. 10, 524.
Hillis D.M., Bull J.J. 1993. An empirical test of bootstrapping as a method for assessing confidence in phylogenetic analysis. Syst. Biol. 42, 182192.
Holmes S. 2003. Bootstrapping phylogenetic trees: theory and methods. Statist. Sci. 18, 241255.
Huelsenbeck J.P., Hillis D.M., Jones R. 1996a. Parametric bootstrapping in molecular phylogenetics: applications and performance. In: Ferraris, J.D., Palumbi, S.R. (Eds), Molecular
Huelsenbeck J.P., Hillis D.M., Nielsen R. 1996b. A likelihood ratio test of monophyly. Syst. Biol. 45, 546558.
Huelsenbeck J.P., Rannala B. 2004. Frequentist properties of bayesian posterior probabilities of phylogenetic trees under simple and complex substitution models. Syst. Biol. 53, 904913.
Lanyon S.M. 1985. Detecting internal inconsistencies in distance data. Syst. Zool. 34, 397403.
Larget B., Simon D.L. 1999. Markov chain monte carlo algorithms for the bayesian analysis of phylogenetic trees. Mol. Biol. Evol. 16, 750759.
Lee M.S.Y. 2000. Tree robustness and clade significance. Syst. Biol. 49, 829836.
Li W.H., Zharkikh A. 1994. What is the bootstrap technique? Syst. Biol. 43, 424430.
Mort M.E., Soltis P.S., Soltis D.E., Mabry M.L. 2000. Comparison of three methods for estimating internal support on phylogenetic trees. Syst. Biol. 49, 160171.
Nei M., Stevens J.C., Saitou M. 1985. Methods for computing the standard errors of branching points in an evolutionary tree and their application to molecular data from humans and apes. Mol. Biol. Evol. 2, 6685.
Newton M.A. 1996. Bootstrapping phylogenies: large deviations and dispersion effects. Biometrika 83, 315328.
Penny D., Foulds L.R., Hendy M.D. 1982. Testing the theory of evolution by comparing phylogenetic trees constructed from five different protein sequences. Nature 297, 197200.
Rodrigo A.G. 1993. Calibrating the bootstrap test of monophyly. Int. J. Parasitol. 23, 507514.
Sanderson M.J. 1989. Confidence limits on phylogenies: the bootstrap revisited. Cladistics 5, 113129.
Sanderson M.J. 1995. Objections to bootstrapping phylogenies: a critique. Syst. Biol. 44, 299320.
Sanderson M.J., Wojciechowski M.F. 2000. Improved bootstrap confidence limits in largescale phylogenies, with an example from NeoAstragalus (Leguminosae). Syst. Biol. 49, 671685.
Shimodaira H. 2002. An approximately unbiased test of phylogenetic tree selection. Syst. Biol. 51, 492508.
Sitnikova T., Rzhetsky A., Nei M. 1995. Interiorbranch and bootstrap tests of phylogenetic trees. Mol. Biol. Evol. 12, 319333.
Sneath P.H.A. 1986. Estimating uncertainty in evolutionary trees from Manhattandistance triads. Syst. Zool. 35, 470–488.
Soltis P.S., Soltis D.E. 2003. Applying the bootstrap in phylogeny reconstruction. Statist. Sci. 18, 256267.
Suzuki Y., Glazko G.V., Nei M. 2002. Overcredibility of molecular phylogenies obtained by bayesian phylogenetics. Proc. Nat. Acad. Sci. U.S.A. 99, 1613816143.
Taylor D.J., Piel W.H. 2004. An assessment of accuracy, error, and conflict with support values from genomescale phylogenetic data. Mol. Biol. Evol. 21, 15341537.
Waddell P.J., Kishino H. and Ota, R. 2002). Very fast algorithms for evaluating the stability of ML and Bayesian phylogenetic trees from se quence data. Genome Informatics 13, 8292.
Wilcox T.P., Zwickl D., Heath T.A., Hillis D.M. 2002. Phylogenetic relationships of the dwarf boas and a comparison of bayesian and bootstrap measures of phylogenetic support. Mol. Phylogenet. Evol. 25, 361371.
Zharkikh A., Li W.H. 1992a. Statistical properties of bootstrap estimation of phylogenetic variability from nucleotide sequences. I. Four taxa with a molecular clock. Mol. Biol. Evol. 9, 11191147.
Zharkikh A., Li W.H. 1992b. Statistical properties of bootstrap estimation of phylogenetic variability from nucleotide sequences. II. Four taxa without a molecular clock. J. Mol. Evol. 35, 356366.
Zharkikh A., Li W.H. 1995. Estimation of confidence in phylogeny: the completeandpartial bootstrap technique. Mol. Phylogenet. Evol. 4, 4463.