Rate Checks from Counts and Waiting Times

One Rate, Several Units of Description

A rate can be described as an expected number of events per unit time, as the reciprocal of a mean waiting time, or as a count divided by an observed duration. These descriptions are easy to confuse because they use different quantities and sometimes different time units. A careful rate check translates each description into the same units before comparing them.

Let \((\Omega,\mathcal{F},\mathbb{P})\) be a probability space, and let \((N_t)_{t\ge0}\) be a Poisson process in continuous time with state space \(S=\{0,1,2,\ldots\}\). Assume first that its rate \(\lambda\) is positive. For an interval of length \(T>0\), \(N_T\) counts arrivals in \((0,T]\); the process has \(N_0=0\), independent stationary increments, and \(N_T\) has a Poisson distribution with parameter \(\lambda T\). The successive interarrival times are \(\tau_k\), and the arrival times are \(S_k=\tau_1+\cdots+\tau_k\).

The key units are reciprocal. If time is measured in hours, a rate has units of events per hour and a mean gap has units of hours per event. Multiplying them gives one event, represented numerically by \(1\). Thus a rate of \(3\) events per hour corresponds to a mean gap of \(1/3\) hour, or \(20\) minutes. Conversely, a mean gap of \(20\) minutes implies \(60/20=3\) events per hour. The conversion is not an additional statistical model; it is a consistency relation between two ways of expressing the same positive rate.

Why the Two Model Quantities Match

For completeness, the count and waiting-time formulas can be checked directly. Write \(\theta=\lambda T\). From the Poisson probability mass function,

For positive \(\lambda\), the first interarrival time has density \(f_{\tau_1}(t)=\lambda e^{-\lambda t}\) for \(t\ge0\). Integrating by parts gives

The boundary term is zero because \(t e^{-\lambda t}\) tends to zero as \(t\) tends to infinity, and its value at zero is also zero. Dividing the count mean by \(T\) therefore gives \(\mathbb{E}[N_T]/T=\lambda\), while taking the reciprocal of the mean gap gives \(1/\mathbb{E}[\tau_1]=\lambda\). These are model-level identities: they do not say that every observed count divided by duration equals \(\lambda\), or that every observed gap equals its mean.

Proof. Each summary is already expressed as a number of events per unit time: \(c/T\) is the claimed mean count per duration, \(1/w\) is the reciprocal mean-gap rate, and \(g/L\) and \(a/H\) are observed counts divided by their corresponding total durations. If all four summaries describe one rate \(r\), each must equal \(r\), so they must be equal to one another. Conversely, if they are equal, let \(r\) be their common value. Then the claimed count summary, mean-gap summary, and both observed summaries all give the same events-per-time value \(r\), which is precisely mutual arithmetic compatibility. The positivity assumptions ensure that all divisions and the reciprocal are defined. \(\square\)

For the interarrival data, if the \(g\) complete gaps have total length \(L\), their arithmetic mean is \(L/g\). Its reciprocal is \(g/L\), the observed rate from those gaps. This identity is exact for the recorded gaps. It does not establish that the gaps are exponential or that their average equals the population mean.

Work the Rate Check in One Time Unit

The conversion should be done before comparing numbers. A value in events per minute cannot be compared directly with a value in events per hour, even if both are numerically written as rates. For positive rates, changing the time unit multiplies the rate and divides the mean gap by reciprocal factors.

Proof. In one hour, the numerical rate \(\lambda_{\mathrm{hour}}\) gives the expected number of events. That same hour consists of \(60\) minutes, so the number per minute is \(\lambda_{\mathrm{hour}}/60\). The reciprocal of this per-minute rate is \(60/\lambda_{\mathrm{hour}}\) minutes, while the reciprocal of the per-hour rate is \(1/\lambda_{\mathrm{hour}}\) hours. Because \(60\) minutes equal one hour, these are the same duration. The assumptions \(\lambda_{\mathrm{hour}}>0\) and \(60>0\) make both reciprocals and conversions valid. \(\square\)

What an Observed Rate Does—and Does Not—Say

An observed count divided by duration is calculated from one realized window. Under a Poisson model, the expected count in a window of length \(T\) is \(\lambda T\), but the realized count \(N_T\) is random. It need not equal \(\lambda T\). Likewise, an average of a finite set of observed gaps need not equal the population mean \(1/\lambda\). A rate check can identify inconsistent numerical interpretations, but it is not a statistical test, a model fit, or proof that the model is suitable.

Count conventions matter too. Here \(N_T\) counts arrivals in \((0,T]\), so an arrival at time \(0\) is not included. For a positive-rate Poisson process, an arrival at a fixed endpoint has probability zero, but the convention still makes the definition precise. When calculating an observed count rate, use the count and the duration that belong to the same observation window. When using gaps, use complete gaps and the sum of those same gaps; a partially observed initial or final waiting time is not a complete interarrival interval.

Common Mistakes and the Zero-Rate Case

Common mistake: invert without converting units. If a mean waiting time is \(20\) minutes, the calculation \(1/20=0.05\) events per hour is wrong: it treats \(20\) minutes as though it were \(20\) hours. The correction is to convert first: \(20\) minutes is \(20/60=1/3\) hour, and \(1/(1/3)=3\) events per hour. Checking in minutes gives \(1/20\) events per minute, and \((1/20)(60)=3\) events per hour, the same answer.

A second mistake is to treat an observed count as if it were automatically the expected count. For example, \(9\) observed arrivals in \(3\) hours gives a realized rate of \(9/3=3\) per hour. It does not, without further information, establish that the process has rate \(3\), nor does it mean the model’s expected count must equal \(9\) in every \(3\)-hour window. Keep empirical summaries and model expectations distinct.

Finally, the reciprocal-mean rule requires a positive rate. If \(\lambda=0\), the process has no arrivals, so \(\mathbb{E}[N_T]=0\) for every finite \(T\). The expression \(1/\lambda\) is undefined; it must not be reported as a finite mean gap. Thus a zero count rate is not handled by applying the positive-rate reciprocal formula.

In practice, write every candidate rate in one unit, list which values are model means and which are observed summaries, and compare only quantities intended to describe the same rate. This simple discipline catches many errors of scale and interpretation before any probability calculation is attempted.