01.12.2013 | Research article | Ausgabe 1/2013 Open Access

# Quantitative summaries of treatment effect estimates obtained with network meta-analysis of survival curves to inform decision-making

- Zeitschrift:
- BMC Medical Research Methodology > Ausgabe 1/2013

## Electronic supplementary material

## Competing interests

## Authors’ contributions

## Background

## Methods

### Motivating example

#### Evidence base

### Network meta-analysis

## Results

### Treatment effects and functional estimates

_{ Ak }(t) is the HR of intervention k relative to A (i.e. DTIC), and d

_{0Ak }and d

_{1Ak }are the differences in scale and shape of treatment k relative to A as obtained with the NMA.

Outcome | DTIC | DTIC + IFN | DTIC + non-IFN | Non-DTIC |
---|---|---|---|---|

Median survival | 7.85 | 7.87 | 9.88 | 10.19 |

Expected survival (after all patients died) and 95% credible interval | 12.61 (11.31, 14.13) | 11.41 (8.44, 15.48) | 16.11 (11.21, 23.14) | 15.31 (9.17, 24.34) |

Mean survival at 22 months and 95% credible interval | 9.84 (9.13, 10.60) | 9.61 (7.66, 11.72) | 11.15 (8.88, 13.33) | 11.23 (8.05, 13.99) |

### Graphical and numerical summaries of rank probabilities

^{st}, 2

^{nd}, 3

^{rd}, and 4

^{th}best treatments was assessed on the basis of each of the aforementioned effect measures.

_{ k,b }to be among b best treatments:

## Discussion

### Advantages and disadvantages of different effect measures in relation to treatment ranking

Measure | Probability that a treatment is associated with: | Explicitly reflects time effect | Reflects cumulative effect over time | Requires baseline risk | Advantage | Disadvantage |
---|---|---|---|---|---|---|

Median survival | The greatest survival time when 50% patients are alive | No | Yes | Yes | Commonly used and clinically relevant; Easily summarized as statistic; May limit need for extrapolation; | Ignores what happens after 50% of subjects have experienced the event; |

Expected survival | The greatest expected survival | No | Yes | Yes | Directly relevant for cost-effectiveness; Easily summarized as statistic; | Sensitive to tail of distribution (may involve extrapolation); Does not illustrate time-varying results or time of greatest treatment effect; May not be as clinically relevant; |

Mean survival at time t | Greatest mean survival (area under the curve) up until time t | No | Yes | Yes | Limits need for extrapolation if time t corresponds to follow-up time of trial with shortest duration; Easily summarized as statistic | May be difficult to interpret; Requires subjective selection of time t; Ignores tails of distribution and does not illustrate time-varying results; |

Hazard (ratio) over time | The smallest hazard (ratio versus reference treatment) over time | Yes | No | Yes for hazard, | Directly relates to model and may help emphasize changes in treatment effect over time; | Does not capture cumulative effect of treatment over time; May lead to over interpretation near tail of distribution; Cannot be summarized as statistic (requires graphical illustration); May be more difficult to understand; |

No for hazard ratio | ||||||

Survival proportion over time (Cumulative hazard over time) | The greatest survival (proportion) over time | Yes | Yes | Yes | Highly intuitive and clinically relevant; Can be easily compared to data; | Cannot be summarized as statistic (requires graphical illustration); |

Mean survival over time | Greatest mean survival (area under the curve) over time | Yes | Yes | Yes | Reflects a cumulative summary of survival proportions up until that time point, thereby de-emphasizing tail of distribution; | Cannot be summarized as statistic (requires graphical illustration); May be more difficult to understand; |

### Time independent measures: median survival, expected survival, or mean survival at a specific time point?

### Time-varying measures: hazard, survival, or mean survival over time?

### Probability of being best treatment, rankograms, or SUCRA?

^{nd}best, 3

^{rd}best, etc. However, rankograms become more difficult to interpret for time-varying treatment effects as compared to one-dimensional effect measures. In the context of time-varying treatment effects, graphing SUCRA may provide a more concise summary measure than presenting all rank probabilities.