Cause of death (COD) information is essential to guide and inform health policy and priority debates [1]. Ideally, COD data would be based on accurate medical certification and registration of all deaths [2]. However, vital registration systems still function poorly in many countries, particularly in resource-poor settings where mortality rates are higher and accurate cause of death information is most crucial [3]. Verbal autopsy (VA) is now becoming recognized as the only feasible alternative to comprehensive medical certification of deaths in such settings. The World Health Organization has now called for wider use of VA to improve understanding of the causes of mortality and the nature of mortality change in national populations [4].

Although VAs have been incorporated into official data collection systems already in place in countries such as India [5], Brazil [6], Bangladesh [7], and Sri Lanka [8], as well as through the collection of VA samples during national censuses as in Mozambique [9], doubts have remained about the ability of VAs to provide accurate and timely information about the COD in populations. This can be attributed, in large part, to the initial reliance on physician certification of verbal autopsies (PCVA) in demographic and health surveillance research sites. PCVA is time-consuming and expensive, and it is difficult to maintain the quality of cause assignment on a large scale over long periods of time.

These problems, however, can be resolved by introducing automated VA diagnostic methods, which have been shown to out-perform PCVA in terms of their accuracy. They now offer the potential for inexpensive, rapid, and reliable COD assignments for deaths occurring outside of hospitals [10–13].

Current practice in the application of VA is to collect interview information using paper-based verbal autopsy instruments (VAIs), which have been largely derived from VA methods developed for research sites in the 1980s and 1990s [14, 15]. A barrier to their widespread adoption by national registration systems has been the amount of time and, hence, cost needed to conduct interviews and to maintain their quality. For widespread application, a short instrument is needed, but one that still enables automated diagnostic systems to make accurate predictions of causes of death. At the same time, electronic systems for data collection need to replace paper-based systems.

We address these needs in this paper and describe a shortened version of the VAI developed for the Population Health Metrics Research Consortium Gold Standard Verbal Autopsy Validation Study (PHMRC study) [16]. Using a formal empirically-based method to shorten the VAI, we identify the key survey questions and the optimal length of a shortened VAI.

Our general approach was first to establish a rank order of individual question items in the PHMRC VAI in terms of their importance in predicting COD. We did this using the Tariff 2.0 Method [17] to predict the COD for each VA in the PHMRC Gold Standard database and by comparing the predicted COD with the gold standard cause. Second, we reduced the size of the instrument by dropping items in reverse order of their importance. We assessed the predictive performance of the instrument at each stage of item reduction by calculating chance-corrected concordance (CCC) at the level of the individual and cause specific mortality fraction (CSMF) accuracy at the level of the population (see below). Finally, the optimum size of the shortened instrument was determined using a first derivative analysis of the decline in performance as the size of the VAI progressively decreased. We followed the same approach for adults, children, and neonates.

The reduced VAI analyzed with the Tariff 2.0 method ascertains causes of death in 34 mutually exclusive, collectively exhaustive categories for adults, 21 for children, and 6 for neonates, as the original PHMRC VAI. As would be expected, in terms of CCC and CSMF accuracy, the predictive performance of successively shortened versions of the questionnaire declined as question items were systematically removed (Fig. 1a). Performance metrics are shown in Table 2 for questionnaires both with and without text. Sensitivity and specificity for each COD with the long and shortened version of the questionnaire are presented in Additional files 4 and 5. For example, the shortened child module reduced to 69 questions had a CSMF accuracy of 78.3 % when all text items were included and 74.5 % when text items were excluded. CCC for the shortened child module was 52.5 % with text and 44.5 % without. These absolute differences of 3.5 % and 8.0 %, respectively, in diagnostic accuracy are reasonably substantial and continued to increase more or less monotonically as the number of question items was reduced, with the notable exception of CSMF accuracy for adult and neonate deaths without text items in the VAI.

			Chance-corrected concordance		CSMF accuracy
			Median (%)	95 % CI	Median (%)	95 % CI
Adult	Full PHMRC VAI	No Text	44.3	(44.1, 44.4)	74.5	(74.0, 74.8)
		Text	50.5	(50.3, 50.8)	77	(76.6, 77.5)
	Shortened PHMRC VAI	No Text	43.3	(43.1, 43.5)	74.6	(74.0, 75.2)
		Checklist	46.6	(46.5, 46.8)	76.2	(75.7, 76.7)
		Text	50	(49.8, 50.2)	76.6	(76.2, 77.2)
Child	Full PHMRC VAI	No Text	44.6	(44.2, 45.0)	74.3	(73.6, 75.2)
		Text	52.6	(52.2, 53.1)	78.3	(77.6, 78.7)
	Shortened PHMRC VAI	No Text	44.5	(44.1, 45.0)	74.5	(73.7, 75.2)
		Checklist	51.8	(51.4, 52.1)	78	(77.5, 78.7)
		Text	52.5	(52.2, 52.9)	78.3	(77.6, 78.9)
Neonate	Full PHMRC VAI	No Text	42.2	(41.9, 42.6)	81.4	(80.5, 82.4)
		Text	44.7	(44.3, 45.0)	82.8	(81.9, 83.5)
	Shortened PHMRC VAI	No Text	40.2	(40.0, 40.5)	81.1	(80.1, 81.8)
		Checklist	43.4	(43.0, 43.8)	82.1	(81.4, 83.0)
		Text	43.4	(43.1, 43.6)	82.2	(81.4, 83.1)

CSMF cause specific mortality fraction, CI confidence interval

Perhaps a better way to summarize the decline in predictive performance with a progressively shorter VAI is to examine the first derivative at different numbers of question items (Fig. 1b). Since the number of question items for the shortened instrument was chosen to maintain performance in terms of CCC and CSMF accuracy, both with and without text items, the first derivative of the curve of performance vs. the number of question items should not drop significantly below zero. Applying this criterion, we identified the optimum number of question items as 90, 60, and 50 for adults, children, and neonates, respectively. This translated into 91, 50, and 48 questions, respectively. It should be noted that significant drop-off in performance of the neonatal module was not seen until there were only about 25 to 30 question items remaining. This reflects the shorter list of target causes for neonates.

As described in the methods section, questions that would enable the shortened version to function as a stand-alone instrument in a survey of importance were added back into the survey. This process increased the number of questions for the short VA to 109, 69, and 67 questions for adults, children, and neonates, respectively.

It is difficult to see how substantial improvements in obtaining information about COD patterns in resource-poor populations can be achieved through expanding medical certification. Accurate certification depends on the physician being intimately familiar with the decedent’s clinical history and/or the details of the terminal illness. In resource-poor countries, only a small proportion of decedents have had access to a physician; this situation is unlikely to change in the medium term. If a physician has not been directly involved in the management of a patient, the COD on the person’s medical certificate has little more value than the COD obtained from an unstructured VA (a VA that does not follow a structured and validated format). The routine use of VA with known performance characteristics for all deaths that have not been attended by a physician is the only cost-effective means of obtaining information about cause of death patterns and about how these patterns are changing.

In the context of the PHMRC validation study, in which VAs and diagnosis methods were compared with gold standard hospital deaths, automated computer diagnosis of VAs was shown to be more accurate than PCVA [22]. Reliance on PCVA is inefficient and leads to unnecessary competition for resources with clinical services, often leading to long delays in diagnosing VAs. For example, the VA data collected in India’s Sample Registration System starting in 2002 have still not been reported or released because of delays to physician coding [23]. Outstanding barriers to widespread use are the length of time required to administer the full-length VAI (50–70 minutes) and the resources required for data entry of paper-based records of interview.

The Tariff Method was developed and validated using hospital gold standard deaths [10]. The Method is additive in the sense that it sums tariff scores by cause and by symptom to arrive at the most probable cause for an individual death. The use of hospital gold standard deaths as reference ensured that a full range of symptoms was available for the development of the tariff scores. The process of item-reduction was designed to remove from the questionnaire those symptoms which were not contributing significantly to the summed tariff scores. Tariff 2.0 sets cut-off points to establish that sufficient symptoms are available to assign a COD for an individual death [17]. Such cases are assigned to an indeterminate category under particular circumstances. If the symptom list in the item-reduced instrument were too short this would manifest itself in an increase in the size of the Indeterminate category.

In this paper, we have established the theoretical basis for the validity of a shortened VAI administered by means of hand-held electronic tablets. Although the shortening of an instrument may lead to a decrease in performance and some loss of specificity, at least for rare diseases, we have demonstrated by formal statistical methods applied to validation datasets, where the true COD is known, that it is possible to reduce the length of a VA questionnaire by 40 % or more without a significant drop in performance. The performance characteristics of the shortened VAI are now established. We have also shown that many of the advantages of an open-ended narrative in improving performance can be retained by use of an on-screen checklist. The application of this checklist will require special training of fieldworkers. It is worth highlighting that symptoms mentioned in the open narrative may have a different Tariff score for a given cause of death than their counterparts in the structured questionnaire. This may reflect the fact that more salient information may be easier to be mentioned spontaneously than being recalled after prompted in a questionnaire. The checklist of keywords from the open narrative is, therefore, an important innovation of this shortened questionnaire. It reduces the burden on the interviewer by registering keywords instead of registering the answer verbatim, and also captures these key items that have a substantial contribution to determine the cause of death.

We consider the greatest utility of the shortened VAI will be for the collection of COD data in civil registration systems and for the calculation of CSMFs in populations. To put our results in perspective: a CSMF accuracy of 76.2 % in adults, using the tablet with check list for data entry, compares with a reported CSMF accuracy of 82 % for medical certification of adult deaths in Mexican teaching hospitals [24]; the former equates to an adjusted accuracy of 32.8 % and the latter to an adjusted accuracy of 50 %.

We have set the number of questions items in the shortened VAI at 90 items for adults, 60 for children, and 50 for neonates. We justified the thresholds at the analytical level for question items for each questionnaire by an analysis of first derivatives, identifying the precise point at which they begin to deviate significantly from zero (signifying no apparent slope in the performance curve). In the case of the neonatal module, the suggested threshold was around 30 items. We took a conservative approach, however, and increased this to 50 items. A limited range of symptoms applies to conditions that cause death in neonates and we wished to avoid eliminating items that might add important contextual information for automated diagnosis.

CSMF accuracy declines less rapidly than does CCC with progressive reduction in the number of items. This can be attributed to the finding that random allocation of deaths to different causes would result in CSMF accuracy of at least 63 % [25]. We have, however, chosen to show absolute values for CSMF accuracy, as it is this measure that will influence the use of CSMFs derived from VAs as the basis for public health policy development.

We would argue that the different forms of the VAI and the different methods of data collection and analysis need to be tailored for use in particular circumstances. Here, we have described an approach which should be invaluable for the collection of vital statistics.

In this study, we developed a shortened VAI using a systematic approach and assessed its performance when administered by means of hand-held electronic tablets and analyzed using the Tariff 2.0 automated method. We demonstrated that, where the true cause of death is known, it is possible to reduce the length of a VA questionnaire by 40 % or more without a significant drop in performance. We have also shown that many of the advantages of the open-ended narrative in improving performance can be retained by use of an on-screen checklist. The reduced VAI developed has great utility to estimate COD data in civil registration and to calculate CSMFs in populations reducing the burden of time and resources required for data collection and analysis.

VA questionnaires have been constructed to elicit as much useful diagnostic information as possible for as many causes as possible for which symptomatic information is likely to be meaningful. Very little attention has been paid to the length of the interview or to the possible effects of interviewer/interviewee fatigue. If VA methods are to be useful beyond research settings to provide the essential intelligence on population cause of death patterns that governments and donors need, then it is critically important that they be rapidly integrated into national civil registration and vital statistics systems and routinely applied to all out-of-hospital deaths that are reported. New automated diagnostic methods and data collection platforms using tablets or mobile phones are now available for widespread use in civil registration systems. That utility will be even more enhanced if shorter VA questionnaires, such as the one reported here, are applied to cut interview time in half without any loss of diagnostic accuracy.