A search of the literature was performed in mid-January 2014 in the Pubmed, Medline, PsychInfo, Cinahl, Web of Science and Embase electronic databases. The search term employed was “Parkinson’s disease” in combination with “caregiver” or “carer” and with “burden” or “distress” or “stress” or “strain”. The search was limited to English and Scandinavian languages. There were no limitations in terms of publication year. A manual search was also made of the reference lists of the articles included in the review. A total of 581 articles were identified, of which 318 were duplicates.

All articles were assessed for inclusion based on the following criteria: 1) original empirical study, 2) published in peer-reviewed journals, 3) with an identifiable and separately analysed group of informal caregivers of persons diagnosed with PD, 4) explored characteristics that are assumed to influence caregiver burden and 5) burden was measured by means of a standardized caregiver instrument. The sample size of eligible studies was ≥50 caregivers, which has been suggested as a reasonable number of cases to ensure statistical power in correlation and regression analysis [20] . Exclusion criteria were intervention studies, medical trials and testing of instruments. Seventy titles were identified as relevant for inclusion. After reading the abstracts, 23 of the 70 articles were excluded in line with the inclusion and exclusion criteria, while another 16 articles were excluded after reading the full text, leaving a total of 31 studies for inclusion (Table 1). The flow chart illustrating study selection and the reasons for exclusion is presented in Figure 1.

For data extraction, two authors (IL, UTL) independently identified the characteristics of the studies such as authors, publication year, country of origin, aim, participants, recruitment methods, study design, main definitions, outcome measures, statistical analysis and key findings. The first author (IL) identified and categorized all de-

*When multivariate procedures are used, these are the only reported except when bivariate analyses show differences between groups. HRQoL: Health-Related Quality of Life, ICD: Impulse Control Disorder, NPI: Neuropsychiatric Inventory, NPI-CD: Neuropsychiatric Inventory-Caregiver Distress Scale, PD: Parkinson’s disease, PDD: Parkinson’s Disease Dementia, PD-MCI: Patients with Parkinson’s disease-mild cognitive impairment, PD-NC: Patients with Parkinson’s disease-no cognitive impairment, PDQ-39: 39-item Parkinson´s Disease Questionnaire, Qol: Quality of Life, SF-36: Short Form (36) Health Survey.

mographic and clinical characteristics examined in the studies regarding their possible influence on caregiver burden. The second author (UTL) made some spot tests to cross-check the validity of the latter extraction. In addition, the various instruments used to assess patient characteristics were collected.

Data from the included studies were narratively synthesized. A narrative approach was used in preference to a meta-analysis due to the heterogeneity of the studies with a huge variety of instruments applied to assess clinical aspects of PD and caregiver burden outcomes [21] .

Two of the authors (ES/IL) independently assessed the methodological quality of the studies using a modified version of the Norwegian Knowledge Centre for the Health Services (NOKC) check-list for cross-sectional studies [22] . The check-list is a tool for assessing the degree to which the methodology of the studies reduces the risk of systemic bias. It contains seven questions assessing criteria such as 1) population, 2) sampling methods, 3) comparison of respondents and non-respondents, 4) response rate, 5) data collection procedures, 6) reliability and validity and 7) statistical methods. For the purpose of this review, the check-list was supplemented by an additional criterion; ethical considerations (Criterion 8). The response alternatives for all questions were Yes (Y), No (N) and No information (NI) (Table 2). Differences in judgment between the two reviewers were resolved

Note: Y = yes; N = no; NI = no information. ^*Quality assessment questions 1. Was the population from which the sample was drawn clearly defined? 2. Were sampling methodsadequate? 3. Was it explained whether (and how) the participants who agreed to participate differed from those who refused? 4. Was the response rate adequate? 5. Were procedures for data collection standardized? 6. Were measures shown to be reliable and valid? 7. Were the statistical methods appropriate? 8. Wereethicalissuesconsidered? ^**Studies meeting > 50% of the criteria rated as acceptable in terms of quality, studies meeting ≤ 50% of the criteria rated as low quality [22] .

by re-reading the studies and subsequent discussions. When in doubt, a statistician was consulted for a final decision. Studies that met > 50% of the criteria were rated as acceptable in quality, whereas those that met ≤50% of the criteria were rated as being of low methodological quality [23] . Finally, 26 out of the 31 studies were deemed to be of acceptable quality. No studies were excluded due to low quality. Methodological shortcomings mainly concerned criteria 3, 4 and 6. Twenty one of the studies failed to explain whether and how the participants who agreed to participate differed from those who refused (criterion 3). Fifteen studies did not achieve an acceptable response rate of 65% [24] , or failed to provide information about the response rate (criterion 4). Fourteen studies did not comment on the validity or reliability of the outcome measure (criterion 6). Ethical issues were considered in 26 of the 31 selected studies (criterion 8).

Thirty one studies published between 1991 and 2014 in English-language journals were included in the review. The studies represent European, North- and South American, as well as Asian countries. Key features of the included studies are described in Table 1. All but one longitudinal study [25] have a cross sectional design. Fifteen were designed as single studies and 16 were sub-studies that emanated from an epidemiological study, a clinical trial or an intervention study. In the sub-studies, data collection was conducted in the original study or new information was collected from the respondents who participated in the original study, or in some cases by recruiting additional participants. Some studies partly share a database with one or more of the other included studies [25] -[28] and [29] [30] . However, they all have separate statistical analyses. In most studies, the caregiver was recruited together with the person she/he cares for. The caregiver sample size ranged from 50 to 2476, with the majority of studies having less than 100 participants, of whom spouses/partners constituted the main caregiver group. Fifteen studies provide a definition or description of caregiver burden, distress, strain or stress, referring to the concepts’ multifaceted nature of external, environmental and personal factors [9] , subjective and objective burden [31] [32] , emotional distress [33] , enduring change of well-being [12] , psychosocial load as a result of caring [29] [30] , impact on social life and health [34] and difficulties in fulfilling the caregiving role [25] [26] . Five studies [35] -[39] refer to the much cited definition of caregiver burden presented by Zarit et al. [40] , which defines it as “the extent to which caregivers perceived their emotional or physical health, social life, and financial status as suffering as a result of caring for their relative” (p. 261). Only one study [41] has an explicitly theoretical approach, based on a stress-appraisal model. Twenty two of the 31 included studies provide multivariate statistical procedures analysing the relationship among three or more variables [24] (multiple regression, analysis of covariance, multivariate analysis of variance, factor analysis, logistic regression and path analysis). Nine studies only provide bivariate statistical tests, analysing the relation between two variables [24] (Person’s product moment correlation, Spearman’s rho, Mann-Whitney U-test, analysis of variance and t-test). Nineteen studies report sub-scores of the caregiver burden instruments, whereas 12 only report the sum scores of the burden instruments. Eleven studies use some stratification of the sample in the statistical analysis.

Eleven different generic instruments for assessing caregiver burden, distress, strain or stress were identified. The instruments cover a wide range of domains with a variety of items and sub-scales (Table 3). Eleven studies used a caregiver burden instrument based on the ones developed by Zarit et al. [42] [43] . The Caregiver Burden Inventory [44] and The Relative Stress Scale [45] were each used in three studies. Four articles [25] -[28] refer to the caregiver section of the Family Caregiving Inventory [46] , although the number of scales and items used vary across studies. One study [47] used the original 10 item version of the Neuropsychiatric Inventory (NPI) [48] , whereas four studies [33] [49] -[51] employed the 12 item NPI Caregiver Distress Scale (NPI-CD) [52] . The original Caregiver Strain Index (CSI) [53] was used in the study by Tanji et al. [54] , whereas the modified version (MCSI) [55] with an amended response set, was employed by Kelly et al. [12] and Peters et al. [56] . The Care Management Stress Scale [57] , the Caregiver Burden Scale [58] , The Montgomery, Gonyea and Hooyman’s scale [59] and the Multidimensional Caregiver Strain Index [60] were used in only one study each.

Psychometric testing of internal consistency was carried out by means of Cronbach’s alpha for all but one instrument. Except for a few sub-scales the Cronbach’s alpha level was beyond the acceptable border of 0.70, de-

monstrating the internal consistency of the instruments [24] . In addition, construct and/or criteria validation was reported for six instruments. The validity of the instruments for use in a PD population was further confirmed for seven of the instruments with a Cronbach’s alpha > 0.75 for nearly all scales.

The included studies use various terminologies, including caregiver or carer, burden, distress, strain or stress. In this article we generally use the terms caregiver and burden. When referring to results from one particular article we use that article’s own terminology.

A range of different instruments was used to assess patient characteristics possibly associated with caregiver burden. These are listed consecutively when mentioned in the text. A summary of the synthesizing of demographics and the association of PD characteristics with caregiver burden is presented in Table 4.

Despite the great variety of instruments used in the analysis, similar associations were found between caregiver burden, demographic variables and certain patient characteristics across the reviewed studies. There is clear statistical evidence of the relationship between the severity of PD and increased caregiver burden. In particular, higher PD stage and patients’ problems performing ADL contribute to caregivers’ burden. Duration of PD, motor symptoms and complications were not identified as predictors of caregiver burden. Associations between caregiver burden and non-motor symptoms of PD, such as depression and psychosis, cognitive impairment and behavioural problems, are well documented in the reviewed studies. Moreover, several studies report patient depression and behavioural problems as predictors of caregiver burden. In addition, comorbidity of depression, apathy, psychosis, dementia and impulse control disorders in patients correlate with significantly greater burden for caregivers, whereas the individual contribution of patient dementia on caregiver burden was inconclusive. Furthermore, when comparing the impact of patient motor and non-motor symptoms on caregiver burden, several studies found that patients’ mental symptoms have a stronger impact on caregiver burden than motor symptoms or functional impairment. This finding differs from the review by Lau et al. [10] , who revealed that patient motor symptoms and dependency in ADL were most strongly correlated with caregiver distress, while the weakest association was the patient’s cognition. The different inclusion criteria applied in the review by Lau et al. [10] and the present review, which resulted in an overlap of only 6 studies, might explain the divergent result. The correlation between burden, assessed by various caregiver burden instruments, and patient HRQoL measured by disease specific as well as generic instruments, is clear. The only included study that employed a QoL instrument that goes beyond a health status assessment did not identify a significant relationship between QoL and caregiver burden. As the authors suggest, this indicates the relevance of studying subjective QoL and self-assessed health status as separate concepts [66] .

Despite the fact that many of the authors comment on cultural and ethnic family structures, as well as caring traditions that might influence caregiver burden in their sample [32] [34] [35] [38] [39] [51] [61] [63] , there is no obvious evidence of variation in caregiver burden between studies representing European, North- and South American, as well as Asian cultures. The only international comparative study [54] observed an equal amount of total strain in Japanese and US caregivers despite the differences in demographics and social support, although the two groups differed significantly on several single strain items and the factors that contribute to caregiver strain. Contrary to the expectations due to different caring traditions, the included studies reveal little gender difference in caregiver burden. In contrast, in their meta-analysis of 164 studies of different caregiver samples, Pinquart and Sörensen [96] reported a small but statistically and practically significant gender difference in caregiver burden, with women reporting the highest burden. The influence of patient or caregiver age on reported caregiver burden was found in two studies comparing various caregiver age groups. However, when comparing the younger with the older group the two studies were contradictory in terms of which group was the most burdened. There is also conflicting evidence in the studies about whether caregiver spouses are more burdened than non-spouses, as spouses and non-spouses tend to differ with respect to burden domains, how they are affected and which factors contribute to caregiver burden. In addition, other aspects such as the emotional relationship might explain the inconsistent results. In a literature review carried out by Carbonneau et al. [97] , the quality of the relationship was found to impact on caregiver stress. This is supported by a few of the reviewed studies, in which it is stated that caregivers’ assessment of the quality of their relationship with the care recipient [25] [41] and mutuality in the relationship [25] [26] had a protective effect on caregiver burden.

A number of instruments are available for measuring caregiver burden. Durme et al. [14] identified 55 scales (mainly generic) for assessing the negative impact of caregiving, 42 of which evaluate burden, strain or stress as the main dimension. There could be various reasons for choosing one instrument over another. In this review, the ZBI was the instrument used in most studies (11 studies). Durme et al. [14] rated the ZBI as the preferred caregiver burden instrument and good reliability has been reported in a PD setting [37] . According to Deeken et al. [15] and Den Oudsten et al. [98] , the choice of outcome instrument in research should be based on a clear understanding of the phenomenon under study. In the present review, 15 of the 31 studies presented a definition or conceptualization of caregiver burden, distress, strain or stress, while one study had an explicitly theoretical approach. The diversity of burden instruments, lack of a specific definition of burden and the absence of validity testing in a PD setting make it challenging to standardise tools for assessing caregiver burden in PD. Lack of conceptual consistency and standardization of burden measures leads to problems comparing the burden outcomes across instruments. This was illustrated in the study by Leroi et al. [49] , who included two caregiver burden instruments. Significant differences were found between caregivers of PDD patients compared to those caring for non-PDD patients when assessed by the ZBI, but not by the NPI-CD. To achieve a more unequivocal picture reflecting the burden of PD caregivers, a smaller number of consistent instruments should be used when planning future studies. Recommendations should be made about the most appropriate caregiver burden instruments for a PD setting, as has been done for other factors associated with PD, including depression [99] , fatigue [100] , psychoses [101] and HRQoL [102] .

With regard to the choice of caregiver burden instruments, a wide range of instruments were chosen in the selected studies to measure similar patient characteristics. For instance, five and seven different scales were used to measure patient functional ability and patient depression respectively. Practical or professional reasons within the research team might explain these choices. However, the variation in assessment tools impedes the opportunity to determine whether equal results indicate the same outcome and whether different results mirror real disparities. A lesser number of scales would make it possible to statistically integrate the results of the associations between caregiver burden and patient characteristics in a meta-analysis, which was not done in this review.

Comparison between studies also requires a clearly defined sampling unit in each study. Six of the 31 selected studies only included spouse or partner caregivers. However, in most studies the analyses are limited to the total group of caregivers. The absence of considerations about and stratification of subgroups in the analysis phase of caregiver studies has previously been criticized for limiting the understanding of caregiver burden [103] . This view is supported by the studies in the present review that provide separate analysis for different caregiver groups or control for variables that may influence the results. For instant, patient sex [25] [64] [65] , comorbidity [36] [37] [49] -[51] and the patient-caregiver relationship [34] [65] were differently correlated with caregiver burden in distinct groups of caregivers. In addition, multivariate analysis provided further information about how caregiver age [26] , presence of dementia in the patient [64] , as well as the patient’s depression and UPDRS mentation and motor scores [63] might make a different contribution to burden across caregiver groups.

The various caregiver scales, number of subscales and the different scope of items used in the reported studies demonstrate the multidimensional nature of caregiver burden. However, several of the included studies treat burden as a one-dimensional variable and only present a sum score showing the overall burden without any information about sub-scores. For many years lack of such information has been criticized for not permitting a distinction to be made between various dimensions of burden [44] , for masking variation in the correlates between different patient characteristics and the dimensions of strain [60] , and for overlooking the relative contribution of specific domains of burden to the overall burden scores [104] . The few included studies that provide information about subscales partly address this criticism. An example is the comparative study of Japanese and American caregivers in which the CSI sum scores were the same, whereas single item scores differed significantly between the two caregiver groups [54] . Furthermore, the study by Kim et al. [32] demonstrates that several patient characteristics were differently associated with caregivers’ subjective and objective burden scores. In addition, patient functional disability [28] [29] , caregiver age [26] [32] and the spouses’ relationship with the patient [29] have different associations within various subscales. Another aspect of the multifaceted character of caregiver burden is the assumption that it is influenced by many different factors. The majority of the selected studies considered this problem as the researchers performed a multivariate analysis. However, in contrast to the review by Greenwell et al. [18] who only included studies that identified predictive factors, the present study contains a correlation analysis to include studies that did not meet the criteria for a multiple regression analysis [105] .

This systematic review has several limitations. A meta-analysis was not possible due to the diversity of the instruments used in the various studies. In addition, only quantitative studies using a caregiver burden instrument as an outcome measure of caregiver stress and burden were included. Accordingly, studies employing other methods and instruments that could provide further knowledge about the subject were excluded. Finally, to fully understand the complexity of caring experiences and to plan effective interventions, caregiver-related variables as well as those that may be associated with positive outcomes for the caregiver are required. In addition, the risk of bias needs to be discussed with regard to how the systematic review was carried out. Two of the authors independently assessed the methodological quality. Differences in judgment were resolved by discussion between the authors, who agreed on the final version. However, it is important to acknowledge that this review represents the authors’ understanding of the studies analysed. When conducting a review, researchers are in danger of making choices regarding the selection of studies and data extraction that expose the review to the risk of bias. However, we have strived to reduce the risk of bias by means of a comprehensive search in several electronic databases, formulating explicit inclusion and exclusion criteria for these lection of studies and by cooperation between authors in the data extraction.