We analyzed two solar flares of 1992 September 5 and 6, using the high time resolution (data from BATSE/CGRO, and 100-3000 MHz radio (100 ms) dynamic spectra from PHOENIX. The broadband radio data reveal a separatrix frequency (at 620 and 750 MHz in the two flares) between normal- and reverse-drifting radio bursts, indicating a compact acceleration source where electron beams are injected in both the upward and downward direction. We find a mean injection rate of 1.2 bursts s-1 in one flare and more than 0.7 bursts s-1 in the other. From 12 broad-band, reverse-drifting radio bursts we find in five cases an unambiguous one-to-one correlation between the reverse-drifting radio bursts and hard X-ray (HXR) pulses of similar duration (400 +/- 220 ms). The high significance (15 +/- 6 sigma) of the HXR pulses and the small scatter (+/- 150 ms) in the relative timing strongly supports a close causal connection. The cross-correlation between HXR and radio pulses shows that the HXR pulses are coincident (within the instrumental time resolution) with the reverse-drifting bursts at the injection frequency (880 +/- 50 MHz), and lead the radio bursts by 270 +/- 150 ms at the highest observable frequency (1240 +/- 100 MHz). The average drift time of the downward propagating radio bursts is measured to 150 ms, corresponding to a drift rate of 2350 MHz s-1. We examined various effects to model the observed timing of radio and HXR pulses (propagation delays, radio wave growth and damping, group velocity delays, radio wave scattering, radio wave ducting, light path differences, etc.). Assuming an exciter velocity of v/c = 0.2 +/- 0.1 for the reverse-drifting radio bursts, we infer an altitude difference of H = 8000 +/- 3000 km between the injection site and the HXR source. The most likely explanation for the retarded radio emission seems to be a combination of the following two effects: (1) HXR-emitting (> 25 keV) and radio-emitting electrons have different energies (the exciter velocity of the reverse-drifting radio bursts is associated with less than or similar to 5 keV electrons), and (2) a low (marginal) growth rate for plasma emission at the second harmonics. Delay effects caused by group velocity, collisional damping, wave scattering, and wave ducting are found to be minor (< 30 ms each).