We describe an independent analysis of LIGO data for black hole coalescence events. Gravitational wave strain waveforms are extracted directly from the data using a filtering method that exploits the observed or expected time-dependent frequency content. Statistical analysis of residual noise, after filtering out spectral peaks (and considering finite bandwidth), shows no evidence of non-Gaussian behaviour. There is also no evidence of anomalous causal correlation between noise signals at the Hanford and Livingston sites. The extracted waveforms are consistent with black hole coalescence template waveforms provided by LIGO. Simulated events, with known signals injected into real noise, are used to determine uncertainties due to residual noise and demonstrate that our results are unbiased. Conceptual and numerical differences between our RMS signal-to-noise ratios (SNRs) and the published matched-filter detection SNRs are discussed.