Discriminating Between Premigration and Postmigration HIV Acquisition Using Surveillance Data

Abstract

Background: 

Migrant populations are overrepresented among persons diagnosed with HIV in the European Union and the European Economic Area. Understanding the timing of HIV acquisition (premigration or postmigration) is crucial for developing public health interventions and for producing reliable estimates of HIV incidence and the number of people living with undiagnosed HIV infection. We summarize a recently proposed method for determining the timing of HIV acquisition and apply it to both real and simulated data.

Methods: 

The considered method combines estimates from a mixed model, applied to data from a large seroconverters' cohort, with biomarker measurements and individual characteristics to derive probabilities of premigration HIV acquisition within a Bayesian framework. The method is applied to a subset of data from the European Surveillance System (TESSy) and simulated data.

Findings: 

Simulation study results showed good performance with the probabilities of correctly classifying a premigration case or a postmigration case being 87.4% and 80.4%, respectively. Applying the method to TESSy data, we estimated the proportions of migrants who acquired HIV in the destination country were 31.9%, 37.1%, 45.3%, and 45.2% for those originating from Africa, Europe, Asia, and other regions, respectively.

Conclusions: 

Although the considered method was initially developed for cases with multiple biomarkers' measurements, its performance, when applied to data where only one CD4 count per individual is available, remains satisfactory. Application of the method to TESSy data, estimated that a substantial proportion of HIV acquisition among migrants occurs in destination countries, having important implications for public health policy and programs.

INTRODUCTION

Migrant populations in the European Union (EU) and the European Economic Area (EEA) have been disproportionally affected by HIV. Although approximately 12% of people living in the EU were born in a different country from the one where they were a resident,¹ the corresponding percentage among those diagnosed with HIV in 2019 was 44%.²

A substantial proportion of HIV-positive migrants originate from countries with generalized epidemics. Almost 4 of 10 migrants in Europe diagnosed with HIV in 2019 were born in sub-Saharan Africa,² where HIV prevalence is substantially higher. This fact has contributed to the widespread belief that migrants are infected in their countries of origin, before arrival in Europe, although this viewpoint has been challenged.³

As the evidence accumulates, it becomes more and more clear that migrants are at an increased risk of many diseases postmigration, including sexually transmitted infections.^4–7 Although migrants are typically a young and healthy population,⁸ limited access to HIV and sexually transmitted infections' prevention and testing, social inequalities, and structural factors may lead to individual behaviors, which increase their vulnerability and may explain the higher HIV prevalence in comparison with native populations.^4,5,9

Knowing whether HIV acquisition takes place before or after migration has important public health policy implications because it impacts the design and delivery of HIV prevention and testing services.¹⁰ In addition, having estimates of the proportion of migrants who acquired HIV before entering the country where they were diagnosed is essential for model-based methods that estimate HIV incidence and the number of people living with HIV¹¹ and thus for the prospect of achieving epidemic control through UNAIDS 90-90-90 targets.

Unfortunately, discriminating between premigration and postmigration acquisition is difficult without specific HIV test results, such as a positive test before migration or conversely a negative test after migration or a positive recency test after migration. When only a single, positive HIV test is available, we can attempt to relate the probable time of infection to the time of migration. Systematic analysis to elucidate whether HIV acquisition took place in the country of origin or in the destination country has only been used by few European countries.^3,12,13

The methods that have been used to estimate the time of infection, at the individual level, are mainly based on biomarkers. If the average evolution of a biomarker during untreated HIV infection is well characterized, levels of that biomarker at or after an individual's diagnosis can be used to ‟back calculate” the elapsed time since infection. This principal idea has been applied to specific settings, often in combination with other auxiliary information (eg, life events, behavioral data, and other clinical information).^3,12–20 These methods often require extended information not routinely available in surveillance data.

In this article, we adapt one of the most recent and flexible methods to estimate the time of infection¹⁹ and assess its performance using simulated data designed to resemble data sets collected routinely by European countries. Furthermore, we apply it to HIV surveillance data from the European Surveillance System (TESSy) of the European Centre for Disease Prevention and Control (ECDC) to generate estimates of the proportion of migrants infected premigration and postmigration in the EU/EEA.

METHODS

Data

We used data from TESSy,²¹ pooled in 2018, which included case reports of HIV diagnoses occurring between 1979 and the end of 2017 from 30 countries in the EU/EEA. Data from TESSy were provided by countries listed in the Acknowledgments and released by ECDC. We restricted the data to cases where age and sex were recorded; route of HIV transmission was through sex between men, heterosexual contact, or injecting drug use; and region of origin was not the same as the reporting country or unknown (N = 145,118). We excluded cases without any CD4 cell count measurement (32.4%) or where the year of arrival in the reporting country (used as a proxy for migration date) was not available (33.8%). Finally, we removed cases where the date of HIV diagnosis was reported to be before the date of migration (4.2%), assuming that these could be classified as premigration infections. From the resulting subset of 43,005 cases, we randomly sampled 10% to illustrate the method in a midsized sample; thus, the working data for the current application included 4301 individuals.

For the simulation study, we generated a set of 5000 cases with known HIV seroconversion dates using the same methods as in the study of Pantazis et al.¹⁹ Simulated data approximated the subset of TESSy data used in the main study about certain characteristics: date of birth, sex, route of HIV transmission, region of origin, date of diagnosis, and migration date. Given that the focus of the current work is on the performance of the methods when applied to data from surveillance systems, we removed from the initially simulated data all HIV-RNA measurements, retained the first available CD4 cell count measurement, and excluded cases where AIDS occurred before the first available CD4 measurement or the route of transmission was other than the 3 categories considered in the TESSy data application.

Statistical Analysis

As for most biomarker-based methods, we first model the HIV natural history using data from a cohort of individuals with a known date of HIV seroconversion and applying a mixed model approach. Estimates from both fixed effects and variance components are then taken into account to formally derive not only just a point estimate of an individual's time of infection but also the distribution of this time.

More specifically, the method assumes that a bivariate linear mixed model, with known parameters, correctly characterizes the evolution of CD4 cell count and HIV-RNA viral load (appropriately transformed and denoted by superscripts $c$and $r$, respectively) over time since HIV seroconversion ($t$) while individuals are ART naive and AIDS free. Given the observed measurements, ${\text{y}}_{\text{i}}^{\text{T}}=\left({\text{y}}_{\text{i}}^{\text{c}},{\text{y}}_{\text{i}}^{\text{r}}\right)$, at times ${\text{t}}_{\text{i}}^{\text{T}}=\left({\text{t}}_{\text{i}}^{\text{c}}{\text{,t}}_{\text{i}}^{\text{r}}\right)$for the i-th individual, the distribution function can be written as $\text{f}\left({\text{y}}_{\text{i}}^{\text{c}}{\text{,y}}_{\text{i}}^{\text{r}}\left|{\text{t}}_{\text{i}}^{\text{c}}{\text{,t}}_{\text{i}}^{\text{r}}\right.\right)$. Denoting the observed times of markers' measurements relative to diagnosis by d, it follows that the time of the j-th measurement for the i-th individual since seroconversion can be expressed as ${\text{t}}_{\text{ij}}^{\text{c}}={\text{d}}_{\text{ij}}^{\text{c}}+{\text{w}}_{\text{i}}$ and ${\text{t}}_{\text{ij}}^{\text{r}}={\text{d}}_{\text{ij}}^{\text{r}}+{\text{w}}_{\text{i}}$ with the unknown quantity ${\text{w}}_{\text{i}}$ denoting the time gap between HIV seroconversion and diagnosis date for the i-th individual. Thus, the distribution of the biomarkers, conditional on ${\text{w}}_{\text{i}}$, can be derived by replacing ${\text{t}}_{\text{ij}}^{\text{c}}$ and ${\text{t}}_{\text{ij}}^{\text{r}}$ with ${\text{d}}_{\text{ij}}^{\text{c}}+{\text{w}}_{\text{i}}$ and ${\text{d}}_{\text{ij}}^{\text{r}}+{\text{w}}_{\text{i}}$, respectively. Given the observed measurements, ${\text{y}}_{\text{i}}^{\text{T}}=\left({\text{y}}_{\text{i}}^{\text{c}}{\text{,y}}_{\text{i}}^{\text{r}}\right)$, the posterior distribution of the unknown ${\text{w}}_{\text{i}}$ conditional on ${\text{y}}_{\text{i}}$ can be derived through the Bayes theorem as $\text{f}\left({\text{w}}_{\text{i}}\left|{\text{y}}_i\right.\right)=\frac{\text{f}\left({\text{y}}_{\text{i}}\left|{\text{w}}_{\text{i}}\right.\right)\text{f}\left({\text{w}}_{\text{i}}\right)}{{\displaystyle {\int }_0^{u_i}\text{f}\left({\text{y}}_{\text{i}}\left|{\text{w}}_{\text{i}}\right.\right)\text{f}\left({\text{w}}_{\text{i}}\right)d{\text{w}}_{\text{i}}}},0<{\text{w}}_{\text{i}}<{\text{u}}_{\text{i}}$ where ${\text{u}}_{\text{i}}$ is the upper limit for the possible values of the gap between HIV seroconversion and diagnosis ${\text{w}}_{\text{i}}$ (assuming that HIV seroconversion must have occurred after the age of 10, after 1/1/1980, and after any documented HIV negative test). A uniform prior distribution for ${\text{w}}_{\text{i}}$, over the interval $(0,{\text{u}}_{\text{i}})$, is assumed, but this distribution can be updated based on additional information.

The model $\text{f}\left({\text{y}}_{\text{i}}^{\text{c}}{\text{,y}}_{\text{i}}^{\text{r}}\left|{\text{t}}_{\text{i}}^{\text{c}}{\text{,t}}_{\text{i}}^{\text{r}}\right.\right)$used to characterize the natural history evolution of CD4 cell count and HIV-RNA viral load was a bivariate linear mixed model. The model was fitted to the CASCADE seroconverters'²² natural history data, in which seroconversion dates are well estimated. The following variables were used as covariates in the model: sex, age at HIV seroconversion, region of origin (Africa, Europe, Asia, or others), route of HIV transmission, and calendar year of HIV seroconversion. Category “others” in the region of origin variable mainly comprised Latin America and the Caribbean. Continuous covariates (year of seroconversion and age at seroconversion) entered the model through splines. CD4 decline was assumed linear after a square root transformation, and ${\text{log}}_{10}$-transformed viral load was modeled through a fractional polynomial of time.

Given that the presence or absence of an AIDS-defining illness at or after HIV diagnosis carries additional information regarding the time gap between HIV infection and diagnosis, the formula for the posterior distribution of ${\text{w}}_{\text{i}}$ can be accordingly updated. For example, a person diagnosed without AIDS would be more likely to have acquired HIV recently compared with a similar person diagnosed while having already progressed to AIDS. Thus, for an AIDS-free subject and not on therapy until some time ${\text{d}}_{\text{i}}$, since HIV diagnosis, the posterior distribution of ${\text{w}}_{\text{i}}$ becomes $\text{f}\left({\text{w}}_{\text{i}}\left|{\text{y}}_{\text{i}}{\text{,T}}_{\text{i}}{\text{>d}}_{\text{i}}{+\text{w}}_{\text{i}}\right.\right)=\frac{\text{f}\left({\text{y}}_{\text{i}}\left|{\text{w}}_{\text{i}}\right.\right)\text{S}\left({\text{d}}_{\text{i}}{+\text{w}}_{\text{i}}\left|{\text{X}}_{\text{i}}^{\text{s}}\right.\right)\text{f}\left({\text{w}}_{\text{i}}\right)}{{\displaystyle {\int }_{\text{0}}^{{\text{u}}_{\text{i}}}\text{f}\left({\text{y}}_{\text{i}}\left|{\text{w}}_{\text{i}}\right.\right)\text{S}\left({\text{d}}_{\text{i}}{+\text{w}}_{\text{i}}\left|{\text{X}}_{\text{i}}^{\text{s}}\right.\right)\text{f}\left({\text{w}}_{\text{i}}\right)d{\text{w}}_{\text{i}}}}$ where ${\text{T}}_{\text{i}}$ is a latent variable representing the time from HIV diagnosis to AIDS onset and $\text{S}\left(\text{t}\left|{\text{X}}_{\text{i}}^{\text{S}}\right.\right)$ denotes the corresponding survival function conditional on subject-specific covariates.

This AIDS-free survival function $\text{S}(\text{t}|{\text{X}}_{\text{i}}^{\text{s}})$ was estimated also using CASCADE data,²² truncated after 1/1/1996 (after that point, effective antiretroviral therapy which substantially reduced the probability of developing clinical AIDS became widely available) using a Weibull proportional hazards model with age at seroconversion and sex included as covariates.

Subject-specific estimates of the unknown gap between HIV seroconversion and diagnosis ${\text{w}}_{\text{i}}$, can be derived through the posterior mean, median, or mode, whereas the posterior probabilities of HIV acquisition postmigration for an individual, $\text{i}$, can be expressed as ${\text{π}}_i=P\left(w_i{\text{< m}}_{\text{i}}\right)={\displaystyle \underset{0}{\overset{{\text{m}}_{\text{i}}}{\int }}\text{f}\left(w_i\left|{\text{y}}_i\right.\right)}d{w}_i$ where ${\text{m}}_{\text{i}}$ denotes the (known) time gap between migration and HIV diagnosis.

The method can be applied to individuals with multiple CD4 cell count and viral load measurements (provided they are taken before ART initiation or AIDS onset). However, it is also applicable when there is only one such measurement (CD4 or viral load) or even using information on AIDS status at diagnosis in case pre-ART and pre-AIDS CD4 cell count and viral load measurements are not available.

The estimation procedure was built in R (version 4.0.2)²³ using the base functions optim and integrate along with the additional library mvtnorm (version 1.1). For more details on the method, see reference 19, and for a previous application, see reference 3.

Role of the Funding Source

This work was supported by the European Centre for Disease Prevention and Control (Framework Contract ECDC/2018/014). Other than the coauthors from the ECDC who participated as individual experts, the study sponsor did not influence the study design, analysis, data interpretation, writing of the report, or the decision to submit the article for publication. The ECDC provided data from TESSy. A part of these data were used in analyses presented in this work.

RESULTS

Application to the Subset of TESSY Data

The subset of the TESSy data used for the current application included 4301 individuals. Their characteristics are summarized in Table 1. Most of the migrants (70.9%) in the sample originated from Africa (97% of whom were from sub-Saharan African countries), whereas smaller percentages originated from Europe (14.5%), Asia (5.4%), or other world regions (9.2%). European migrants in the sample originated equally from western (50.3%) and other European countries (49.7%).

TABLE 1. - Characteristics of the Study Sample (TESSy Subset) by the Region of Origin and Overall

	Region of Origin (Grouped)					P
	Africa (n = 3,049, 70.9%)	Asia (n = 232, 5.4%)	Europe (n = 626, 14.5%)	Others (n = 394, 9.2%)	Overall (n = 4,301, 100%)
	N (%)	N (%)	N (%)	N (%)	N (%)
Sex						<0.001
Male	1188 (39.0)	159 (68.5)	505 (80.7)	306 (77.7)	2158 (50.2)
Female	1861 (61.0)	73 (31.5)	121 (19.3)	88 (22.3)	2143 (49.8)
Route of transmission						<0.001
Heterosexual contact	2885 (94.6)	129 (55.6)	205 (32.7)	156 (39.6)	3375 (78.5)
Sex between men	149 (4.9)	100 (43.1)	359 (57.3)	233 (59.1)	841 (19.5)
Injecting drug use	15 (0.5)	3 (1.3)	62 (9.9)	5 (1.3)	85 (2.0)
Region of origin (detailed)						<0.001
Sub-Saharan Africa	2957 (97.0)				2957 (68.8)
Western Europe			315 (50.3)		315 (7.3)
Southern Asia		195 (84.1)			195 (4.5)
Latin America				189 (48.0)	189 (4.4)
Central Europe			152 (24.3)		152 (3.5)
Caribbean				134 (34.0)	134 (3.1)
Northern Africa–Middle East	92 (3.0)				92 (2.1)
Europe (unspecified)			86 (13.7)		86 (2.0)
Eastern Europe			73 (11.7)		73 (1.7)
Northern America				48 (12.2)	48 (1.1)
Eastern Asia–Pacific		37 (15.9)			37 (0.9)
Australia–New Zealand				23 (5.8)	23 (0.5)
Year of diagnosis						<0.001
Pre-2008	1491 (48.9)	62 (26.7)	147 (23.5)	98 (24.9)	1798 (41.8)
2008–2012	776 (25.5)	62 (26.7)	167 (26.7)	108 (27.4)	1113 (25.9)
2013–2017	782 (25.6)	108 (46.6)	312 (49.8)	188 (47.7)	1390 (32.3)
Information used for prediction						<0.001
Only 1 CD4 measurement	2887 (94.7)	216 (93.1)	609 (97.3)	378 (95.9)	4090 (95.1)
Only the presence of AIDS	158 (5.2)	10 (4.3)	10 (1.6)	10 (2.5)	188 (4.4)
More than 1 CD4 or HIV-RNA measurements	4 (0.1)	6 (2.6)	7 (1.1)	6 (1.5)	23 (0.5)

	Region of Origin (Grouped)					P
	Africa (n = 3,049, 70.9%)	Asia (n = 232, 5.4%)	Europe (n = 626, 14.5%)	Other (n = 394, 9.2%)	Overall (n = 4,301, 100%)
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)
Age at diagnosis, yr	35 (29–42)	34.5 (29–41)	32 (27–40)	34.5 (28–42)	34 (29–41)	<0.001
CD4 count at diagnosis, cells/μL	270 (126–433)	273.5 (97–456.5)	409 (240–593)	410 (245–583)	300 (140–480)	<0.001
Year of diagnosis	2008 (2004–2013)	2012 (2007–2015)	2012 (2008–2015)	2012 (2008–2015)	2009 (2005–2014)	<0.001
Year of migration	2002 (2000–2007)	2005 (2000–2010)	2007 (2001–2012)	2005 (2000–2012)	2003 (2000–2009)	<0.001
Time from migration to diagnosis, yr	3.0 (0.8–7.2)	4.7 (0.9–10.4)	2.3 (0.6–7.8)	3.5 (0.6–9.9)	3.0 (0.7–7.6)	0.001

Figures in italic font denote observed ("true") proportions of premigration HIV infections.

Heterosexual contact was the predominant route of transmission (78.5%), followed by sex between men (19.5%) and injecting drug use (2.0%). The median time between arrival to the reporting country and diagnosis ranged from 2.3 to 4.7 years, for migrants originating from Europe and Asia, respectively. The median age at diagnosis was 34 years with European migrants being slightly younger (median 32 years). The median CD4 count at diagnosis was around 270 cells/μL for African or Asian migrants and around 410 cells/μL for migrants from Europe or other countries.

Most of the migrants (95.1%) in the sample had 1 CD4 cell count measurement available, taken soon after HIV diagnosis (93% within 3 months from diagnosis). A small proportion (4.4%) was diagnosed with clinical AIDS close to diagnosis and did not have any pre-ART/pre-AIDS CD4 cell count or HIV-RNA measurements. Finally, a very small proportion (0.5%) had more than 1 pre-ART/pre-AIDS CD4 cell count measurement or 1 CD4 cell count and 1 HIV-RNA viral measurement.

The results from the application of the method are summarized in Table 2. Estimated probabilities of premigration HIV acquisition were highest (0.64) for African migrants and 0.54 to 0.60 for migrants originating from other regions. Using the probability of 0.5 as a threshold, the percentages of migrants who were classified as having acquired HIV in the country of origin were 68.1%, 62.9%, 54.7%, and 54.8%, for those originating from Africa, Europe, Asia, and other regions, respectively. The median values for the estimated time between HIV acquisition and diagnosis ranged between 4.97 and 6.23 years, for migrants from other regions and Africa, respectively. Differences between the region of origin in all aforementioned estimates were highly statistically significant (P < 0.001).

TABLE 2. - Results From the Application of the Method to the TESSy Subset Data

Region of Origin (Grouped)	Postmigration HIV Infection N (%)	Premigration HIV Infection N (%)	Probability of Premigration HIV Infection Mean (SD)	Time Between HIV Infection and Diagnosis, yr Median (IQR)
Africa	973 (31.9)	2076 (68.1)	0.64 (0.33)	6.23 (5.45–7.97)
Europe	232 (37.1)	394 (62.9)	0.60 (0.37)	5.30 (4.86–6.55)
Asia	105 (45.3)	127 (54.7)	0.54 (0.38)	5.81 (5.08–7.81)
Others	178 (45.2)	216 (54.8)	0.54 (0.38)	4.97 (4.50–6.06)
Overall	1488 (34.6)	2813 (65.4)	0.62 (0.35)	5.98 (5.19–7.63)

Simulation Study

The simulated data set included 5000 cases contributing data on first CD4 count measurement after diagnosis and before the initiation of ART or progression to clinical AIDS. Characteristics of the simulated cases are summarized in Table 3. The distribution of demographic and clinical characteristics of the simulated cases closely approximated those in the subset of TESSy data with the exception of CD4 counts at diagnosis which were slightly higher in the simulated data (median 358 vs. 300 cells/μL, respectively).

TABLE 3. - Characteristics of the Simulated Cases

	Region of Origin					P
	Africa	Asia	Europe	Others	Overall
	N (%)	N (%)	N (%)	N (%)	N (%)
Total	3660 (100.0)	296 (100.0)	652 (100.0)	392 (100.0)	5000 (100.0)
Sex						<0.001
Male	1442 (39.4)	193 (65.2)	516 (79.1)	305 (77.8)	2456 (49.1)
Female	2218 (60.6)	103 (34.8)	136 (20.9)	87 (22.2)	2544 (50.9)
Route of transmission						<0.001
Heterosexual contact	3477 (95.0)	169 (57.1)	214 (32.8)	152 (38.8)	4012 (80.2)
Sex between men	166 (4.5)	122 (41.2)	384 (58.9)	237 (60.5)	909 (18.2)
Injecting drug use	17 (0.5)	5 (1.7)	54 (8.3)	3 (0.8)	79 (1.6)

	Region of Origin					P
	Africa	Asia	Europe	Others	Overall
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)
Age at diagnosis, yr	33 (28–40)	32 (27–40)	31 (26–37)	32 (26–38)	33 (27–40)	<0.001
CD4 count at diagnosis, cells/μL	354 (197–608)	277 (127–498)	393 (234–646)	365 (203–592)	358 (198–604)	<0.001
Year of diagnosis	2008 (2004–2012)	2011 (2007–2014)	2010 (2006–2014)	2010 (2007–2014)	2008 (2004–2012)	0.005
Year of migration	2003 (1998–2008)	2005 (2000–2010)	2006 (2001–2011)	2005 (2001–2010)	2004 (1999–2009)	<0.001
Time from migration to diagnosis, yr	2.9 (0.8–7.1)	3.9 (1.5–8.2)	2.3 (0.7–5.6)	2.8 (0.9–6.5)	2.8 (0.9–6.9)	<0.001

The results from the application of the method to the simulated data set are summarized in Table 4. The overall observed proportion of premigration HIV acquisition cases in the simulated data was 69.4%. The mean of the corresponding estimated probabilities was 0.64. Overall, 66.7% of cases were classified as premigration HIV infections. The corresponding pairs of observed (“true”) and estimated proportions of premigration HIV infections by the region of origin were 70.2%–67.3%, 70.2%–68.7%, 62.2%–57.8%, and 66.6%–63.8% for migrants from Africa, Europe, Asia, and other regions, respectively. Overall, the method correctly classified 85.2% of cases with a sensitivity (ie, probability of correctly classifying a premigration case) of 87.4% and a specificity (ie, probability of correctly classifying a postmigration case) of 80.4%, with the positive and negative predictive values (ie, probabilities of being a premigration or postmigration case, given the corresponding classification by the method, respectively) being 91.0% and 73.7%, respectively.

TABLE 4. - Results From the Application of the Method to the Simulated Data

Region of Origin (Grouped)		Postmigration HIV Infection N (%)	Premigration HIV Infection N (%)	Probability of Premigration HIV Infection Mean (SD)	Time Between HIV Infection and Diagnosis, yr Median (IQR)
Africa	Estimate	1196 (32.7)	2464 (67.3)	0.65 (0.32)	6.06 (5.33–7.43)
	True	1091 (29.8)	2569 (70.2)		6.46 (3.64–10.00)
Europe	Estimate	204 (31.3)	448 (68.7)	0.66 (0.32)	5.45 (4.89–6.64)
	True	194 (29.8)	458 (70.2)		5.12 (2.85–8.27)
Asia	Estimate	125 (42.2)	171 (57.8)	0.59 (0.34)	5.98 (5.13–7.94)
	True	112 (37.8)	184 (62.2)		6.46 (3.56–9.91)
Others	Estimate	142 (36.2)	250 (63.8)	0.61 (0.33)	5.23 (4.59–6.36)
	True	131 (33.4)	261 (66.6)		5.15 (3.09–8.37)
Overall	Estimate	1667 (33.3)	3333 (66.7)	0.64 (0.32)	5.95 (5.17–7.30)
	True	1528 (30.6)	3472 (69.4)		6.19 (3.48–9.64)
Correctly classified	Sensitivity	Specificity	PPV	NPV	Lin coefficient of concordance
85.22%	87.36%	80.37%	91.00%	73.67%	0.409

Using the estimated probability of premigration infection as a continuous measure for discriminating between premigration and postmigration infections in an ROC analysis resulted in an area under the curve equal to 0.924 (Fig. 1).

FIGURE 1.: Receiver operating characteristic curve for the performance of the estimated probability of premigration infection.

Estimated time gaps between infection and diagnosis were close to the ‟true” ones with the difference in the overall median values being 0.24 years and no more than 0.5 years across different regions of origin. However, the agreement between estimated and observed time gaps from infection to diagnosis, for each case, was only moderate (the Lin coefficient of concordance = 0.409).

DISCUSSION

In this work, we adapted, evaluated in a simulation study, and applied to surveillance data a recently proposed statistical method¹⁹ to estimate premigration or postmigration HIV acquisition among migrants to the EU/EEA. The simulation results showed a very satisfactory performance, whereas the application to data from European surveillance systems showed that a large proportion of HIV acquisitions occur postmigration.

This method was previously applied to a rich data set which included demographic and clinical variables, repeat biomarkers (CD4 cell counts and HIV-RNA measurements), and behavioral data within the aMASE study.³ Our current application of the method uses a subset of routinely collected data from TESSy²¹ in which the relevant information is much more restricted (eg, usually only one CD4 cell count, no HIV-RNA measurements, and no behavioral data). The accompanying simulation study was also tailored to mimic the data availability in typical surveillance systems.

The results from the simulation study showed that restricting the relevant data to variables collected in TESSy had an impact on the accuracy of the estimated time gaps between HIV acquisition and diagnosis. The concordance coefficient for simulated data mimicking the aMASE study¹⁹ was 0.69 and dropped to 0.40 in the current study. However, the percentage of cases which were correctly classified, based on the timing of HIV acquisition relative to date of migration, was 85.2% in the current simulation study, comparable with the aMASE study (84%).³ Similar findings hold for sensitivity and specificity of the method. These results are encouraging because they indicate that the method has good classification performance even in the absence of multiple biomarker measurements and behavioral data.

Applying the method to a subset of TESSy data, we estimated in a substantial proportion of migrants, HIV was acquired postmigration. The corresponding percentages were 31.9%, 37.1%, 45.3%, and 45.2% for those originating from Africa, Europe, Asia, and other regions, respectively. These estimates were lower than those found in the aMASE study (45% for African migrants and 69%–71% for migrants from other regions³). The difference between the 2 sets of results can be probably explained by differences in the selection of the 2 samples: The aMASE study sample had to fulfill specific criteria (eg, diagnosed in the preceding 5 years, living in the destination country for at least 6 months, and only 9 European countries contributed data), whereas the subset of TESSy data includes diagnoses from a much wider time period over which the likelihood of premigration/postmigration HIV acquisition may have altered substantially. HIV cohorts, although broadly representative of the diagnosed population, were shown to underrepresent specific patient groups,²⁴ which may also affect the estimated proportion of premigration HIV acquisition. On the other hand, the high proportion of missing data on CD4 cell count and time of migration in TESSy resulted in a possibly selected subset which reflects the current data availability because these variables were generally more complete for the recent years of data. Thus, they may not be a representative of the whole EU/EEA migrant population diagnosed with HIV.

The main strength of the examined method is that it has a solid theoretical justification based on the full structure of a mixed model for biomarker(s) evolution during the natural history of HIV. The method takes into account the evolution of the average biomarker levels over time and the evolution of their variance and covariance. In addition, the statistical and computational framework of the method allows the proper treatment of covariates which change over time (eg, age and calendar time), whose values are known at the time of diagnosis but unknown at the time of infection. The incorporation of additional external information is also feasible through informative prior distributions. Finally, the data used to fit the initial mixed model are derived from one of the largest collaborative seroconverter studies,²² which comprised 19,788 individuals contributing 125,195 CD4 cell count measurements, 106,160 HIV-RNA VL measurements, and a wide range of information on demographic and clinical characteristics.¹⁹

The main limitation of the examined method is that it assumes that, given a set of crucial covariates, the individuals we want to classify are similar to the seroconverters whose data were used to fit the initial mixed model. This is an assumption used in all biomarker-based methods, but a previous study²⁵ showed that estimates of HIV progression derived from seroconverters are likely to hold more generally for the HIV-positive population. Another drawback of the examined method is that it is more complex and computationally intensive compared with other methods which are based only on the evolution of the mean CD4 cell count and disregard the variance–covariance structure of the seroconverters' mixed model.^13,15 In this regard, a tool to support the application of the proposed method by public health researchers working in HIV epidemiology would be useful.

Finally, it should be mentioned that our method does not take into account details of the migration process (eg, duration and route) or the probability of acquiring HIV during a short visit to the country of origin after migration. Behavioral data could be also used³ to improve the estimates of our method, but unfortunately, such data or data on the migration process are not routinely collected within surveillance systems. In general, the quality of data as collected in surveillance systems always remains a challenge, and in our model, especially, the poor completeness of CD4 count and time of arrival may lead to biased conclusions.

Besides the biomarker-based methods, another family of methods exists to assess premigration or postmigration HIV acquisition, based on molecular data analysis. For example, Paraskevis et al²⁶ used phylogenetic analysis to directly classify HIV infections as local or imported based on the idea of local transmission networks. Puller et al²⁷ presented a method to estimate the time gap between HIV infection and diagnosis based on the idea that diversity in the pol gene increases with time since infection. Finally, the concept of the molecular clock²⁸ could be applied as a method to elucidate the place of infection by estimating the time of infection and its relation to the time of migration. Unfortunately, molecular analysis methods require data and procedures that at this time are not routinely collected/performed within typical surveillance systems, although such data would be very important when studying and monitoring the HIV epidemic, and could be considered in surveillance in the future. It should be noted that the examined biomarker method could incorporate information from molecular methods because of its Bayesian nature.

To conclude, the ability to discriminate between premigration and postmigration HIV acquisition is important for both monitoring the epidemic and designing testing and prevention strategies. The biomarker method we applied showed a very satisfactory ability to correctly classify diagnoses according to the place of HIV acquisition based on the information currently collected in the European surveillance system. The method is flexible and able to incorporate additional information (eg, from HIV-RNA viral load measurements or molecular data); thus, improvements in the quality and quantity of routinely collected data could further improve its performance. Most importantly, it is essential for surveillance systems to maintain and improve the collection of data on CD4 cell counts at diagnosis and migrants' time of arrival. Finally, our study also confirms that postmigration HIV acquisition is likely quite high and it merits routine monitoring in the future.