The method problem in the Bell congestion studies

Before the 1997 ETI study took the Bell filings apart one at a time, it assessed them as a group. The trouble was not arithmetic but method: a hand-picked sample, a cost study aimed at one end of the call, and a missing revenue line.

A 2.4 percent view of the network

Lee L. Selwyn and Joseph W. Laszlo opened their general assessment with a counting exercise. Between them, US West, Pacific Telesis, Bell Atlantic and NYNEX operated roughly 5,200 central office switches, part of a national base of 23,686. Their studies examined some 127 of those switches, about 2.4 percent of the four regions. The switches were not drawn at random. Each was selected because it happened to serve one or more online providers, which is to say each was chosen for the very condition the filings set out to prove was widespread. Even inside that sample, only a handful showed traffic anyone called a problem.

Pacific Bell ran 790 switches in California and looked at 29. One was singled out as a place where data calls made up a large share of traffic, and it sat in Silicon Valley, in what was probably the country's densest concentration of computer industry workers. By November 1996 the company was walking the claim back, telling reporters the difficulty was confined to that one location in the evening and telling the California Public Utilities Commission on November 14 that service had not deteriorated, with better than 99 percent of calls completing on the first attempt. The evidentiary base for a national change in policy came down to one switch on a busy night.

Costs measured where the money is not

The second flaw was where the studies pointed their instruments. A dial-up session has two ends. Whatever costs can be traced to online providers appear at the terminating end, on the switch serving the provider's modem bank, which is where the Bells measured. The revenue arrives at the other end, from the subscriber who dialed the call and pays a monthly bill. Origination was left out almost entirely. A cost study trained on one end of a transaction while the money flows in at the other will find a shortfall every time, and the report reduced it to a four-cell table: revenue high at origination and low at termination, costs low at origination and moderate at termination.

Underneath sat the claim that calls to an online service were free. They were not free, they were prepaid, the way a letter is paid for by its sender. The caller pays either a per-call charge or a flat monthly rate for unlimited local calling, and those state-regulated rates sat above the incremental cost of the usage. The report cited the FCC's own proxy costs: 0.2 to 0.4 cents per minute for end office switching and 0.15 cents for tandem, against local usage priced at 2 to 3 cents. Business lines, what providers bought, were not priced below cost either, and most paid extra for features such as hunting or direct inward dialing.

"The state's phone network is not in trouble because of the Internet."

John Britton, Pacific Bell spokesman, quoted in the San Jose Mercury News, November 8, 1996

The burden that was also a product

Then the report turned the accusation over. The additional residential line, the thing the Bells presented as an unfunded load, was the fastest growing item in their consumer business. Before 1990, between 2.5 and 3.0 percent of American households with telephone service bought a second line. That reached 9.2 percent by 1992 and 14.7 percent by 1995, with 1996 earnings reports suggesting 16 percent or higher. Inside NYNEX, additional residential lines grew 10.1 percent in 1994 and 9.6 percent in 1995, against about 3.4 percent for residential lines overall. ETI attributed roughly 6 million residential lines in 1995 principally to online access.

What made those lines lucrative was where they came from. Local carriers had for years pulled at least two pairs to a residence, sometimes four in expensive neighborhoods, so a second line often meant turning up copper already in the ground. Nor were the companies behaving like firms saddled with demand they did not want. Pacific Bell planned to hire 2,500 people in 1997 partly to keep up with line orders, several Bells advertised second lines as the way to get online, and one Pacific Telesis promotion paired a new line with free setup and five months of unlimited Pacific Bell Internet service.

Beaten by the Bells' own cost estimate

The comparison that ended the argument was arithmetic, and it used the industry's own number. Bellcore, the research arm the Bells jointly owned, put the cost of reinforcing the network at about $35 million per year per regional company, roughly $245 million nationally. ETI estimated revenue from additional residential lines installed for online use at $1.4 billion in 1995 alone and more than $3.5 billion cumulatively from 1990 through 1995. That single line item exceeded the industry's own repair bill by a factor of six, before counting what providers paid for business lines and features, or any measured local usage.

Why was that fatal rather than merely awkward? A new access charge needs a specific predicate: a class of traffic that imposes costs the carrier has no means of recovering. Enter the second line on the revenue side and the predicate evaporates. What remains is an ordinary argument about whether existing local rates are set correctly, a state ratemaking question rather than grounds for the FCC to invent a per-minute charge on internet providers. Scale pushed the same way. Bell Atlantic ascribed about $2.5 million of switch work to internet demand, and put $30 million as its total annual cost of serving providers, against $409 million spent on central office switching in 1995 alone.

"from idle capacity in an existing plant"

Raymond F. Smith, chairman and CEO of Bell Atlantic, on second-line growth, Merrill Lynch conference, March 19, 1996

Where the Bells were not wrong

None of this makes the engineering complaint imaginary, and the report never said it was. Data calls genuinely behaved unlike voice calls. They held a connection far longer than a typical conversation, and they peaked in the late evening rather than the mid-morning business hour, shifting the busy hour that capacity planning is built around. Pacific Telesis was describing a real pattern in its own traffic data. The report conceded the mechanism too: where a provider's lines terminate on a line concentration module, each call occupies a line port for its full duration, and a module that has not been rebalanced can block other subscribers.

The dispute was how far that observation reached. Providers could be, and often were, connected through trunk ports that bypass the concentrator entirely. Bell Atlantic itself reported that half the providers in its territory used PRI ISDN, which lands on a digital trunk controller and cannot block other users at the terminating switch. The studies pooled both arrangements into one aggregate burden. The report also separated two things that sound alike: a standard busy signal, which means one customer bought too few lines, and a fast busy or reorder tone, which means the switch or its trunks are exhausted. Busy signals at small providers were evidence of the first.

What a convincing study would have shown

Read as methodology rather than advocacy, the chapter is a list of things the filings could have done and did not. A representative sample, or a stratified one with stated weights, would support a claim about the network as a whole; 127 hand-picked offices support a claim only about those offices. A baseline would have shown what each switch was going to cost anyway as demand grew, since switches are modular and take new line and trunk cards on a routine schedule. Both ends of the call would have appeared in the ledger, serving arrangements would have been reported separately, and the method would have been published so others could reproduce it.

The report proposed a test the Bells could have run and did not: compare the share of local minutes attributable to online calling against the share of access lines. If online use were genuinely disproportionate, minutes would outrun lines. ETI's figures ran the other way. About 6 million dedicated lines plus roughly 4 million users dialing in over a voice line came to some 6.8 percent of the 147 million access lines in service at the end of 1995, while an assumed 15 hours per user per month yielded about 108 billion minutes, roughly 5.4 percent of the 2 trillion local minutes carried that year.