StATS: Continuous monitoring of the number needed to harm (created 2006-09-02).

The continuing review of clinical trials has to address "good news" issues. Does one arm of the study show substantially better efficacy? Does one arm of the study have a significantly better safety profile? There are rigorous and well accepted approaches for determining partway through a clinical trial whether one arm has a greater proportion of cured patients or a smaller proportion of harmed patients.

Continuing review also has to address "bad news" issues. Is the study falling behind schedule on its planned enrollment rates? Are patients dropping out of the study at an alarming rate? Are certain adverse drug reactions occurring at an unexpected rate?

The analysis of  "bad news" issues is more poorly developed. Often decisions about these issues are based on subjective opinions and ad hoc rules. Statistical process control charts and Bayesian statistical methods offer an approach to treat on-going review of rates not tied directly to an efficacy or safety comparison.

Consider a hypothetical research study that started in January 1997 with the intention to recruit 12 patients per year (one per month) over a ten year period, for a total sample size of 120 patients. By the end of June 2004, (roughly 7 1/2 years), the study has enrolled 42 patients (Table 1).

 02/26/1997 04/04/1997 07/07/1997
 07/25/1997 02/05/1998 02/15/1998
 03/06/1998 07/03/1998 08/03/1998
 02/08/1999 03/19/1999 04/20/1999
 05/29/1999 06/21/1999 07/27/1999
 09/06/1999 01/10/2000 01/11/2000
 02/28/2000 03/03/2000 04/13/2000
 05/30/2000 11/21/2000 12/18/2000
 02/06/2001 04/30/2001 08/03/2001
 01/20/2001 12/03/2001 12/07/2001
 09/27/2002 10/01/2002 02/02/2003
 03/03/2003 10/31/2003 11/04/2003
 11/11/2003 01/05/2004 02/02/2004
 04/15/2004 05/23/2004 06/02/2004
Table 1. List of accrual dates.

Clearly this clinical trial has problems. The actual accrual rate is a meager 5.6 patients per year, and now it is probably too late to fix things. In order to finish on time, the researchers would have to recruit at a rate more than 30 patients per year over the remainder of the study. This is more than 5 times faster than the current accrual rate and 2.5 times faster than the original planned accrual rate.

Wouldn't it be nicer if the researcher had noticed the problem two years into the study rather than 7 1/2 years out? The researcher would still have to hustle, but 14 patients per year would allow the study to still finish on time and it represents only a modest increase over the planned rate.

The traditional approach to examining rates is to set a time interval (weeks, months, or years, for example) and count the number of events per that time interval. For example, you could compute the monthly rates

Jan97 0
Feb97 1
Mar97 0
Apr97 1
May97 0
Jun97 0
Jul97 2

Or the yearly rates

1997 4
1998 5
1999 7
2000 8

or something in between like the quarterly rates

97Q1 1
97Q2 1
97Q3 2
97Q4 0
98Q1 3

A narrow time interval allows you to respond very rapidly, but the individual values (mostly zeros and ones) are so limited that the information value of this approach may be limited. The yearly approach has more information for any single time interval, but if you have to wait a full year or more to spot any important changes. A quarterly interval offers the best (worst?) of both worlds.

I am proposing a different approach that looks not at the events per time interval but rather the time interval per event. Start by calculating the date gap: the amount of time between successive events. The trial started on on January 1 and recruited the first two patients on February 26 and April 4. The gap between the start of the study and the first patient is 46 days and the gap between the first and second patients is 37 days. A plot of the date gaps for the entire study appears below:

Note that the vertical axis shows a mixture of time units. This was done intentionally to emphasize one of the biggest advantages of the date gap approach. Date gaps are self scaling and automatically configure themselves appropriately. If events occur frequently, the data will fit in the portion of the vertical axis where units are measured in days or weeks. If the events occur rarely, the data will fit in the portion of the graph where units are measured in months, quarters, or even years.

The date gap has a second important advantage. Each time a patient is recruited into the trial, another point appears on the chart. When you count the number of events per time interval, you have to wait until the end of that time interval before you can plot an additional data point. By recasting accrual rates in terms of the number of days between successive patients, we have liberated this problem from arbitrary calendar boundaries.

An important unsolved question is how to set control limits properly for a chart with highly skewed data. The control limits is traditionally divided into zones, but for this data set, some of the zones are in the negative territory. This suggests that perhaps the control limits should be set using an asymmetric rule, that the traditional use of zones in a control chart should be modified for skewed data, or possibly both. Another intriguing possibility is to transform the data prior to computing control limits.

Another intriguing prospect is to apply a CUSUM chart to this technique. An accrual rate of one per month implies an average date gap of 30 days. If you plot the cumulative sum of the deviations of each individual date gap from the target of 30 days you get the following chart.

The data has shown a clear and consistent problem from the very first date gap. After 10 patients, the study is more than a year behind schedule. While there are a few places where the accrual seems to be making up for lost time, more often than not, the study is falling further and further behind. The decision rules for CUSUM charts are not very well defined and we are investigating a Bayesian approach.

 02/26/1997  04/04/1997  07/07/1997
 07/25/1997D 02/05/1998  02/15/1998
 03/06/1998  07/03/1998  08/03/1998
 02/08/1999  03/19/1999  04/20/1999D
 05/29/1999  06/21/1999  07/27/1999
 09/06/1999  01/10/2000  01/11/2000
 02/28/2000D 03/03/2000D 04/13/2000
 05/30/2000  11/21/2000  12/18/2000
 02/06/2001D 04/30/2001  08/03/2001D
 01/20/2001  12/03/2001  12/07/2001
 09/27/2002  10/01/2002D 02/02/2003
 03/03/2003  10/31/2003  11/04/2003
 11/11/2003  01/05/2004  02/02/2004D
 04/15/2004  05/23/2004D 06/02/2004
Table 2. List of accrual dates with dropouts designated by D.

For this problem, note that you had to recruit 4 patients before one dropped out, another 8 patients before the second dropped out, another 6 patients before the next one dropped out, and so forth.

The gaps 4, 8, 7, 1, 5, 2, 5, 7, 2 can be plotted in sequence to look for trends. If the drop out rate is accelerating, then the number of patients seen between each dropout is going to be smaller.

The gray line in this graph is drawn at the mean. The mean of the gaps can be shown to be equivalent to the number needed to harm (under the presumption, of course, that a drop-out represents a harmful event).

This page was written by Steve Simon while working at Children's Mercy Hospital. Although I do not hold the copyright for this material, I am reproducing it here as a service, as it is no longer available on the Children's Mercy Hospital website. Need more information? I have a page with general help resources. You can also browse for pages similar to this one at Category: Adverse events in clinical trials.