Factors Associated With STEAM Career Interest Among Female Students in Bangladesh: A Cross‐Sectional Study

ABSTRACT

Background and Aims

Girls' access to education in Bangladesh has improved markedly. Yet women remain underrepresented in Science, Technology, Engineering, Arts and Mathematics (STEAM). Quantitative evidence on the factors linked to female students' STEAM career interest is limited. This study examined how confidence, family support, access to digital resources and female role‐model exposure relate to STEAM career interest. It also examined whether these factors differ between urban and rural students.

Methods

A cross‐sectional online survey was conducted between July and September 2025. Participants were 424 female students recruited through purposive sampling (65.09% undergraduate, 27.83% higher secondary, 6.84% secondary; 73.58% urban). Each construct was measured with a single five‐point Likert item. Analyses included Spearman correlations, chi‐square tests, Welch t‐tests, Mann‐Whitney U tests, and ordinal and binary logistic regression. MANOVA and K‐means clustering were used as exploratory analyses.

Results

In the binary logistic regression with all four factors, confidence (OR = 1.97, 95% CI = 1.53–2.53) and family support (OR = 2.23, 95% CI = 1.74–2.87) were associated with higher odds of STEAM career interest. Role‐model exposure was not associated with interest (OR = 1.15, p = 0.212). Access showed no positive association (OR = 0.76, p = 0.035). Results were similar after adjustment for location and educational level, and after excluding probable duplicate submissions. Urban respondents reported stronger family encouragement than rural respondents (χ²(4) = 18.43, p = 0.001, Cramér's V = 0.21). Confidence and digital access did not differ significantly by location.

Conclusion

In this research, STEAM career interest was associated jointly with self‐confidence and family support. Digital access and role‐model exposure showed no independent positive association. Interventions that combine confidence building with family and teacher engagement, and that embed role models in sustained mentoring, may be more promising than single‐factor approaches. Longitudinal and representative studies are needed to confirm these associations.

1. Introduction

Women's equitable participation in Science, Technology, Engineering, Arts and Mathematics (STEAM) supports inclusive innovation and sustainable development [1, 2]. Globally, however, women remain underrepresented. Young women now enroll in higher education at higher rates than young men. Even so, they account for only about 35% of science graduates, and this share has not improved over the past decade [3, 4]. The gap is concentrated in technical fields. Women earn 57% of degrees in the natural sciences and mathematics, but only 28% in engineering and 30% in information and communication technologies (ICT) [3]. In 2022, they held under a quarter of science, engineering and ICT jobs [3]. Across G20 economies, women made up about 35% of STEAM graduates and 22% of the STEAM workforce in 2023 [5]. Similar patterns are documented across regions [6, 7, 8]. In this paper, STEAM denotes the study domain. STEAM is retained when reporting sources that measured STEAM only.

Several barriers are well documented. They include gender bias in curricula and classrooms, low mathematical self‐confidence among girls despite comparable achievement, a male‐dominated digital transformation, and few visible women in scientific leadership [3, 9]. In Bangladesh, girls' enrollment at the primary and secondary levels has improved markedly. This progress has not carried through to STEAM in higher education and careers. Female students face limited learning resources, weak institutional support and gender stereotypes that discourage STEAM choices [10, 11]. These constraints are often stronger in rural areas [12, 13].

Two gaps limit the evidence base in Bangladesh. The first gap is theoretical. Confidence, family and institutional support, resource access and role‐model exposure are often treated as separate barriers. Each is then addressed by its own intervention. Self‐efficacy research suggests, however, that confidence develops partly through social reinforcement [14, 15]. These factors may therefore be interrelated rather than independent. Whether each contributes independently to career interest has not been examined with Bangladeshi data.

The second gap is methodological. Bangladeshi studies rely mainly on enrollment statistics, qualitative case studies and evaluations of single interventions [16, 17, 18]. Such designs cannot estimate the relative contribution of correlated factors. Multivariable approaches are common in the international literature but remain rare in local studies [19, 20]. The urban and rural divide is also often invoked but seldom tested directly [21]. It remains unclear whether rural disadvantage relates to material access, social encouragement or confidence. As a result, policies that emphasize role‐model visibility, digital infrastructure or confidence building proceed without local evidence on which factors matter most.

This study addresses these gaps with survey data from 424 female students at the secondary, higher secondary and undergraduate levels. It asks three research questions (RQ). RQ1: How are confidence, family support, access to digital resources and role‐model exposure related to interest in STEAM careers? RQ2: Do these factors differ between urban and rural female students? RQ3: Which factors are independently associated with STEAM career interest when all are considered together?

The study makes three contributions (Figure 1). First, it provides a multivariable analysis of female students' STEAM career interest in Bangladesh. Second, it estimates psychological, social and material factors jointly rather than one at a time. Third, it compares urban and rural respondents directly and statistically. Section 2 reviews the literature. Section 3 describes the methods. Section 4 presents the results. Section 5 discusses them, and Section 6 concludes.

Figure 1.

Main contributions of the study.

2. Literature Review

Research on gender disparities in STEAM increasingly treats the relevant factors as interconnected. Girls' engagement appears to arise from psychological factors, family and school influences, and access to resources acting together.

2.1. Global Trends in Women's Participation in STEAM

Women make up only about 35% of STEAM students in higher education worldwide [6]. They are especially underrepresented in engineering and information technology. Contributing factors include limited access to digital tools, cultural expectations, gender bias, and a lack of visible female role models [22]. Many girls internalize societal biases early. This leads to lower confidence and interest in science and technology careers, even when academic performance is strong [9].

2.2. South Asian and Bangladeshi Contexts

In South Asia, girls' education has progressed markedly at the primary level. Gender gaps widen, however, at higher levels and within STEAM fields. Gender norms, limited institutional support and a lack of inclusive curricula reduce girls' participation in STEAM pathways in Bangladesh [11, 23]. Socio‐economic constraints, family expectations and limited career awareness also shape female students' enrollment and persistence in STEAM [19].

Geographic inequality compounds these challenges. Rural students often face weaker infrastructure, fewer academic opportunities and stronger adherence to traditional gender roles [12, 24]. Early marriage, household responsibilities and safety concerns affect rural girls' educational continuity in particular [12]. Differences in teacher quality, resources and institutional capacity also create unequal learning environments between urban and rural schools [13]. In higher education, female enrollment has grown. Participation in technical and engineering disciplines remains low because of persistent stereotypes, limited mentorship and a perceived conflict between STEAM careers and social expectations [17, 18].

2.3. The Digital Gender Divide

Digital literacy is now central to academic success and workforce readiness, particularly in STEAM [25]. Access to digital tools remains unequal. Girls in rural or low‐income communities often fall behind in digital skills [26, 27]. This pattern is known as the digital gender gap [28]. Introducing rural students to basic programming and robotics has been shown to raise their interest and confidence in STEAM [29]. Many female students nevertheless start with fewer resources, less digital exposure and less encouragement to pursue technical paths.

2.4. Institutional and Psychological Barriers

Psychological and institutional factors also shape girls' attitudes towards STEAM [30]. Low self‐confidence and the belief that STEAM subjects suit boys better discourage many girls from enrolling. Teachers may unintentionally reinforce gender stereotypes, and female students often lack mentors or supportive guidance [31]. A comparative analysis of school textbooks in Malaysia, Indonesia, Pakistan and Bangladesh found women underrepresented and shown in gender‐stereotyped occupations [32]. Such portrayals may reinforce the view of science and engineering as male professions.

3. Methodology

3.1. Research Design

The study used a quantitative, cross‐sectional survey design. This design examines relationships among variables at a single point in time and suits large‐scale data collection [33]. It cannot establish causal direction.

3.2. Population and Sampling

The target population was female students at three educational stages in Bangladesh: secondary, higher secondary and undergraduate. These stages were chosen because STEAM aspirations, confidence and support evolve across the transition from school to higher education. More than 11 million female students are enrolled at these levels [34]. The main contextual comparison was between urban and rural respondents, because infrastructure, family expectations and institutional opportunities are known to differ between these settings. Educational level was examined in a robustness model (Section 3.6).

Participants were recruited between July and September 2025 through educational institutions, student networks and online platforms, using non‐probability purposive sampling. Eligible respondents were female students currently enrolled at the secondary, higher secondary or undergraduate level in Bangladesh, with exposure to at least one STEAM subject. All gave informed consent and completed the online questionnaire. The final sample comprised 424 respondents. Of these, 276 (65.09%) were undergraduates, 118 (27.83%) were higher secondary students and 29 (6.84%) were secondary students. One respondent (0.24%) selected both the higher secondary and undergraduate levels.

Ages ranged from 12 to 44 years (M = 22.82, SD = 3.88), and 19 respondents (4.48%) were under 18. Higher secondary respondents averaged 21.0 years and secondary respondents 18.1 years. This suggests that some respondents reported their highest completed level rather than their current level. Educational level is therefore interpreted with caution.

Of all respondents, 312 (73.58%) lived in urban areas and 112 (26.42%) in rural areas. This imbalance partly reflects the online mode of data collection, which favors students with better internet access and institutional connections. The rural subsample was large enough for statistical comparison. The urban and rural comparison is nevertheless interpreted as analytical rather than nationally representative. Because undergraduates formed the majority and only 29 secondary students took part, the findings mainly reflect undergraduate and higher secondary students.

3.3. Sample Size

A minimum sample of 385 was estimated with Cochran's formula for large populations, using 95% confidence, a 5% margin of error and p = 0.5 [35]. The calculation used an estimated population of 11.2 million female students at the target levels [34]. Because sampling was purposive, this figure serves as a precision benchmark rather than a guarantee of representativeness. The final sample of 424 also provided more than 20 respondents in the smaller outcome category for each regression predictor.

3.4. Questionnaire and Data Collection

Data were collected from 31 July to 14 September 2025, with 418 of the 424 responses received in August. A self‐administered online questionnaire was developed from the literature on women's participation in STEAM. It contained four background questions: age, educational level, institution type and residential location. It then presented ten statements rated on a five‐point Likert scale from 1 (Strongly Disagree) to 5 (Strongly Agree). Four statements addressed support and resources: family encouragement, institutional support, affordability of study resources, and access to internet and online learning tools. The other six addressed equal treatment by teachers, confidence in science and mathematics, perceived social pressure, female role models, belief that girls can succeed in STEAM, and interest in a STEAM career. Two optional open‐ended questions asked for suggestions and personal challenges. These received 64 and 62 responses and are not analysed in this paper.

Before data collection, three experts in social science, education and quantitative methods reviewed the questionnaire for clarity and relevance. A pilot with 30 female students tested readability and flow, and minor wording changes followed. Cronbach's alpha for the ten Likert items was 0.80.

3.5. Variables

The outcome was interest in a STEAM career, measured by the item “You are interested in building a career in STEAM fields”. It was analysed as an ordered five‐level variable in ordinal logistic regression. It was also analysed as a binary indicator in binary logistic regression, where Agree or Strongly Agree was coded 1 (n = 296) and all other responses 0 (n = 128).

Four predictors were specified in advance from the literature reviewed in Section 2 [6, 9, 15]. Confidence was measured by “You are confident in science and math subjects.” Family support was measured by “Your family encourages you to study STEAM subjects.” Access was measured by “You have access to the internet and online learning tools.” Role‐model exposure was measured by “You have female role models in STEAM, and you are inspired to be them.” All four were entered together in the primary models (Model 1). Model 2 added residential location (urban = 1) and educational level (university = 1; secondary or higher secondary = 0) as a robustness check. Age and institution type were described but not modeled, because age is closely aligned with educational level.

3.6. Data Preparation

Data were cleaned in Python 3 using pandas and NumPy. Text entries were standardized, and Likert responses were coded from 1 (Strongly Disagree) to 5 (Strongly Agree). There were no missing values on the study variables. Each construct was represented by a single item. Single items are directly interpretable, but their reliability cannot be estimated separately. For analyses involving educational level, secondary and higher secondary students were combined into a school group (n = 147). The respondent who selected two levels was excluded from these analyses.

Submissions were screened for duplicates. Twenty responses were identical to an earlier response on all 14 closed‐ended fields. A further six belonged to a sequence submitted 45–60 s apart that closely reproduced earlier responses. Because these 26 responses could not be confirmed as duplicates, all 424 responses were retained in the primary analysis. The models were then re‐estimated without the 26 flagged responses as a sensitivity check.

3.7. Statistical Analysis

The analyses followed the stages shown in Table 1. They were run in Python 3 with statsmodels 0.14, SciPy and scikit‐learn. Spearman correlations described bivariate relationships (RQ1). Urban and rural differences were tested with chi‐square tests for response distributions and with Welch t‐tests (RQ2). Mann‐Whitney U tests served as non‐parametric checks. Effect sizes are reported as Cramér's V, Cohen's d and r = |Z | /√N, and statistical power for the group comparisons was calculated.

Table 1.

Analytical framework and alignment with the research questions.

Stage	Methods	Purpose	Question
Descriptive	n, %, mean, SD	Describe the sample and variables	Context
Bivariate	Spearman correlation	Relationships among variables	RQ1
Group comparison	Chi‐square test, Welch t‐test, Mann‐Whitney U test	Urban and rural differences	RQ2
Primary models	Ordinal and binary logistic regression (Model 1: confidence, family support, access, role‐model exposure)	Independent associations with career interest	RQ3
Robustness	Model 2: Model 1 plus location and educational level; re‐estimation without flagged submissions	Stability of estimates	RQ3
Exploratory	One‐way MANOVA; K‐means clustering	Joint urban and rural difference; response profiles	Supplementary

For RQ3, proportional‐odds ordinal logistic regression and binary logistic regression were estimated with the four predictors entered together (Model 1). Model 2 added location and educational level. Results are reported as odds ratios (OR) with 95% confidence intervals (CI). Model fit was assessed with the likelihood‐ratio test and McFadden's pseudo‐R². The proportional‐odds assumption was tested with the [36]. Calibration of the binary model was assessed with the Hosmer‐Lemeshow test and discrimination with the area under the ROC curve (AUC) [37]. Multicollinearity was assessed with variance inflation factors (VIF).

Two exploratory analyses followed. A one‐way MANOVA compared urban and rural respondents on the five study variables jointly. Pillai's trace was reported, with Bonferroni‐adjusted univariate follow‐up tests. K‐means clustering was used to describe response profiles. The five variables were standardized as z‐scores, and clustering used k‐means + + initialization with 50 random starts. Solutions with k = 2–8 were compared using the within‐cluster sum of squares and the average silhouette coefficient (Supporting Information S1: Figure S1). The silhouette was highest for k = 2 (0.28) and similar for k = 3 (0.26). The three‐cluster solution was retained because it separated interpretable profiles. Both values indicate weak cluster structure. Significance was set at α = 0.05, two‐sided.

4. Results

4.1. Sample Characteristics

Table 2 summarizes the sample. Respondents were aged 12–44 years (M = 22.82, SD = 3.88), and 19 (4.48%) were under 18. Most were undergraduates (65.09%), attended private institutions (62.26%) and lived in urban areas (73.58%). Overall, 296 respondents (69.81%) agreed or strongly agreed that they were interested in a STEAM career.

Table 2.

Sample characteristics (N = 424).

Characteristic	Category	n	%
Age (years)	M = 22.82, SD = 3.88, range 12 to 44
	18 or younger	45	10.61
	19–20	40	9.43
	21–22	127	29.95
	23–24	120	28.30
	25 or older	92	21.70
Educational level	Undergraduate	276	65.09
	Higher secondary	118	27.83
	Secondary	29	6.84
	Both higher secondary and undergraduate selected	1	0.24
Institution type	Private	264	62.26
	Public	160	37.74
Location	Urban	312	73.58
	Rural	112	26.42
STEAM career interest	Agree or Strongly Agree	296	69.81

Family encouragement differed by location (Table 3, Figure 2). Strong agreement was reported by 30.45% of urban respondents but only 12.50% of rural respondents. Rural respondents more often answered neutral (25.89% vs. 17.31%) or strongly disagree (9.82% vs. 4.17%). The association was significant and small to moderate, χ²(4) = 18.43, p = 0.001, Cramér's V = 0.21.

Table 3.

Family encouragement to study STEAM subjects, by location, n (%).

Location	Strongly disagree	Disagree	Neutral	Agree	Strongly agree	Total
Urban	13 (4.17)	28 (8.97)	54 (17.31)	122 (39.10)	95 (30.45)	312
Rural	11 (9.82)	12 (10.71)	29 (25.89)	46 (41.07)	14 (12.50)	112
Total	24 (5.66)	40 (9.43)	83 (19.58)	168 (39.62)	109 (25.71)	424

Note: Row percentages. χ²(4) = 18.43, p = 0.001, Cramér's V = 0.21. The smallest expected cell count was 6.34.

Figure 2.

Family encouragement to study STEAM subjects, by location (row percentages).

4.2. Correlations

Table 4 presents descriptive statistics and Spearman correlations. Career interest correlated most strongly with family support (ρ = 0.49) and confidence (ρ = 0.44). It correlated more weakly with role‐model exposure (ρ = 0.25) and access (ρ = 0.17). All correlations were significant at p < 0.001. The largest correlation among predictors was between confidence and family support (ρ = 0.42). All VIFs were below 1.4, indicating no problematic multicollinearity. Figure 3 illustrates the association between confidence and career interest.

Table 4.

Descriptive statistics and Spearman correlations (N = 424).

Variable	M	SD	1	2	3	4
1. Confidence	3.64	1.07
2. Family support	3.70	1.12	0.42
3. Access	3.87	1.04	0.30	0.36
4. Role‐model exposure	3.39	1.14	0.31	0.31	0.19
5. Career interest	3.86	1.05	0.44	0.49	0.17	0.25

Note: All items were scored from 1 (Strongly Disagree) to 5 (Strongly Agree). All correlations were significant at p < 0.001.

Figure 3.

Confidence in science and mathematics and interest in a STEAM career. The line shows a least‐squares fit with a bootstrapped 95% confidence band. Points are jittered to show overlapping responses.

The association between confidence and interest was positive among both urban respondents (ρ = 0.48, p < 0.001, n = 312) and rural respondents (ρ = 0.31, p = 0.001, n = 112; Figure 4). The association appeared weaker in the rural group. However, a confidence by location interaction added to the binary model was not significant (OR = 1.50, 95% CI = 0.89–2.51, p = 0.128). This difference should therefore not be over‐interpreted.

Figure 4.

Confidence and interest in a STEAM career among (a) urban and (b) rural respondents.

4.3. Urban and Rural Comparisons

Urban and rural respondents did not differ significantly in confidence (d = 0.16, p = 0.163), access (d = 0.19, p = 0.111) or role‐model exposure (d = 0.12, p = 0.270; Table 5). Urban respondents reported higher family support (d = 0.43, p < 0.001) and higher career interest (d = 0.26, p = 0.022). Mann‐Whitney U tests gave the same pattern. The distributions of confidence and access are shown in Supporting Figure S2. With 312 urban and 112 rural respondents, power was 0.78 to detect d = 0.30 but only 0.44 to detect d = 0.20. The minimum detectable effect at 80% power was d = 0.31. Small differences in confidence and access therefore cannot be ruled out.

Table 5.

Urban and rural comparisons of the study variables.

Variable	Urban M (SD)	Rural M (SD)	Welch t (df)	p	d	Mann‐Whitney U	p	r
Confidence	3.69 (1.06)	3.52 (1.10)	1.40 (190.0)	0.163	0.16	19051.5	0.139	0.07
Family support	3.83 (1.09)	3.36 (1.14)	3.79 (188.7)	< 0.001	0.43	21697.5	< 0.001	0.19
Access	3.92 (0.99)	3.72 (1.16)	1.60 (172.9)	0.111	0.19	18952.0	0.160	0.07
Role‐model exposure	3.42 (1.15)	3.29 (1.12)	1.11 (200.7)	0.270	0.12	18819.5	0.206	0.06
Career interest	3.94 (1.03)	3.66 (1.10)	2.31 (184.6)	0.022	0.26	19933.5	0.020	0.11

Note: Urban n = 312; rural n = 112. d = Cohen's d; r = |Z | /√N.

4.4. Regression Models

Tables 6 and 7 report the regression models. In Model 1, confidence and family support were each associated with higher career interest. In the ordinal model, the odds ratio was 1.89 (95% CI 1.54–2.32) for confidence and 2.31 (95% CI 1.87–2.85) for family support. In the binary model, the corresponding odds ratios were 1.97 (95% CI 1.53–2.53) and 2.23 (95% CI 1.74–2.87). Each one‐point increase in confidence or family support was therefore associated with roughly double the odds of STEAM career interest.

Table 6.

Ordinal logistic regression predicting STEAM career interest.

Predictor	Model 1 OR (95% CI)	p	Model 2 OR (95% CI)	p
Confidence	1.89 (1.54, 2.32)	< 0.001	1.90 (1.54, 2.34)	< 0.001
Family support	2.31 (1.87, 2.85)	< 0.001	2.30 (1.85, 2.85)	< 0.001
Access	0.86 (0.71, 1.05)	0.131	0.85 (0.69, 1.03)	0.092
Role‐model exposure	1.12 (0.94, 1.33)	0.216	1.12 (0.94, 1.34)	0.198
Location (urban vs. rural)			1.05 (0.68, 1.63)	0.823
Educational level (university vs. school)			1.29 (0.86, 1.93)	0.213
Cut points (1 | 2, 2 | 3, 3 | 4, 4 | 5)	0.15, 2.92, 4.16, 6.28		0.27, 3.05, 4.31, 6.42
N	424		423
LR χ2 (df)	162.09 (4)	< 0.001	164.27 (6)	< 0.001
McFadden pseudo‐R2	0.142		0.144
Brant test χ2 (df)	12.91 (12)	0.376	15.91 (18)	0.599

Note: Proportional‐odds logit model. Model 1 coefficients (β, SE): confidence 0.637 (0.105); family support 0.839 (0.107); access −0.150 (0.099); role‐model exposure 0.110 (0.089). Cut points are on the latent logit scale. A non‐significant Brant test indicates that the proportional‐odds assumption holds.

Table 7.

Binary logistic regression predicting STEAM career interest.

Predictor	Model 1 OR (95% CI)	p	Model 2 OR (95% CI)	p
Confidence	1.97 (1.53, 2.53)	< 0.001	1.97 (1.53, 2.54)	< 0.001
Family support	2.23 (1.74, 2.87)	< 0.001	2.26 (1.75, 2.92)	< 0.001
Access	0.76 (0.59, 0.98)	0.035	0.73 (0.56, 0.95)	0.019
Role‐model exposure	1.15 (0.92, 1.44)	0.212	1.15 (0.92, 1.44)	0.211
Location (urban vs. rural)			0.90 (0.51, 1.59)	0.715
Educational level (university vs. school)			1.47 (0.86, 2.52)	0.157
Constant	0.023 (0.006, 0.085)	< 0.001	0.021 (0.006, 0.080)	< 0.001
N (interested)	424 (296)		423 (295)
LR χ2 (df)	115.69 (4)	< 0.001	118.84 (6)	< 0.001
McFadden pseudo‐R 2	0.223		0.229
AUC	0.82		0.82
Hosmer‐Lemeshow χ2(8)	20.17	0.010	16.08	0.041
Largest VIF	1.35		1.39

Note: The outcome was coded 1 for Agree or Strongly Agree and 0 otherwise. Model 1 coefficients (β, SE): constant −3.760 (0.664); confidence 0.677 (0.128); family support 0.803 (0.128); access −0.276 (0.131); role‐model exposure 0.141 (0.113).

Role‐model exposure was not independently associated with career interest in either model (binary OR = 1.15, p = 0.212). Access was positively correlated with career interest on its own (ρ = 0.17). It showed no positive association once the other predictors were controlled. It was non‐significant in the ordinal model (OR = 0.86, p = 0.131) and weakly negative in the binary model (OR = 0.76, 95% CI 0.59 to 0.98, p = 0.035). This change of sign suggests that the bivariate association of access largely reflected its overlap with family support (ρ = 0.36) and confidence (ρ = 0.30). Because the negative estimate was small and inconsistent across the two outcome codings, it is not interpreted substantively.

The proportional‐odds assumption held (Brant χ²(12) = 12.91, p = 0.376). The binary model discriminated well (AUC = 0.82), and 78.5% of respondents were correctly classified. The Hosmer‐Lemeshow test, however, indicated some miscalibration (χ²(8) = 20.17, p = 0.010). This is common when predictors take few discrete values. Predicted probabilities should therefore not be read as precise individual estimates.

4.5. Robustness and Sensitivity

Adding location and educational level (Model 2; N = 423) left the estimates almost unchanged. In the binary model, confidence (OR = 1.97, 95% CI 1.53–2.54) and family support (OR = 2.26, 95% CI 1.75–2.92) remained significant. Access remained negatively associated (OR = 0.73, 95% CI 0.56–0.95, p = 0.019), and role‐model exposure remained non‐significant (OR = 1.15, p = 0.211). Neither location (OR = 0.90, 95% CI 0.51–1.59, p = 0.715) nor educational level (OR = 1.47, 95% CI 0.86–2.52, p = 0.157) was independently associated with career interest. Together they did not improve model fit, LR χ²(2) = 2.01, p = 0.367. Rural respondents reported lower career interest (Table 5), but this difference was not evident once family support and the other predictors were held constant. Re‐estimating the models without the 26 flagged submissions (N = 398) produced the same pattern. Confidence (binary OR = 1.94, 95% CI 1.50–2.52) and family support (OR = 2.35, 95% CI 1.81–3.05) remained significant. Access remained negative (OR = 0.74, p = 0.024), and role‐model exposure remained non‐significant (OR = 1.10, p = 0.441).

4.6. Exploratory Multivariate Comparison by Location

A one‐way MANOVA compared urban and rural respondents on the five variables jointly (Table 8). Box's M indicated unequal covariance matrices (χ²(15) = 32.52, p = 0.005), and group sizes were unequal. Pillai's trace is therefore reported. The multivariate effect of location was significant but small, V = 0.036, F(5, 418) = 3.09, p = 0.009, partial η² = 0.036. Univariate follow‐up tests with a Bonferroni‐adjusted α of 0.01 showed that the difference lay in family support, F(1, 422) = 14.98, p < 0.001, η² = 0.034. Career interest differed only at the unadjusted level (p = 0.018). Confidence, access and role‐model exposure did not differ. Because the outcomes are single Likert items, this analysis is interpreted as exploratory.

Table 8.

Exploratory one‐way MANOVA comparing urban and rural respondents.

Effect	Value	F	df	p	Partial η2
Pillai's trace	0.036	3.09	5, 418	0.009	0.036
Wilks' λ	0.964	3.09	5, 418	0.009	0.036
Univariate follow‐up tests
Family support		14.98	1, 422	< 0.001	0.034
Career interest		5.68	1, 422	0.018	0.013
Access		2.97	1, 422	0.086	0.007
Confidence		2.03	1, 422	0.155	0.005
Role‐model exposure		1.20	1, 422	0.275	0.003

Note: Bonferroni‐adjusted α for the follow‐up tests = 0.01. Box's M χ²(15) = 32.52, p = 0.005.

4.7. Exploratory Response Profiles

The cluster analysis identified three profiles (Table 9, Figure 5). A high‐readiness profile (n = 217, 51.2%) scored high on every variable. The other two profiles had the same moderate confidence (M = 2.98) but differed in pattern. Respondents with low family support and low access (n = 81, 19.1%) still reported moderate career interest (M = 3.60). Respondents with high access but lower family support and role‐model exposure (n = 126, 29.7%) reported the lowest interest (M = 2.97). Because the silhouette coefficient was modest (0.26), these profiles are treated as descriptive groupings rather than distinct student types.

Table 9.

Exploratory response profiles: mean scores (1–5).

Profile	n (%)	Confidence	Family support	Access	Role‐model exposure	Career interest	Urban (%)
1. High readiness	217 (51.2)	4.28	4.35	4.29	3.90	4.48	81.6
2. Low support and access	81 (19.1)	2.98	2.69	2.23	3.07	3.60	66.7
3. Access without interest	126 (29.7)	2.98	3.23	4.18	2.70	2.97	64.3

Note: K‐means clustering on standardized scores (k = 3; average silhouette = 0.26).

Figure 5.

Mean scores by exploratory response profile. Error bars show 95% confidence intervals.

5. Discussion

5.1. Principal Findings

This study examined which factors are associated with Bangladeshi female students' interest in STEAM careers, and whether they contribute independently. Confidence and family support were each associated with roughly double the odds of career interest after mutual adjustment. Role‐model exposure showed no independent association. Access to digital resources showed no positive association once the other factors were considered. Urban and rural respondents reported similar confidence and access but differed in family encouragement. The data are cross‐sectional and self‐reported, and each construct was measured with a single item. The interpretations below are therefore theoretically informed explanations of these associations. They are not demonstrated mechanisms.

5.2. Confidence and Family Support

The association between confidence and career interest is consistent with Social Cognitive Career Theory (SCCT). In SCCT, self‐efficacy is a central antecedent of career interest [38]. The finding also agrees with international evidence linking self‐efficacy to STEM aspirations [14, 15, 39]. SCCT often models contextual supports as distal influences that act largely through self‐efficacy [38]. In our data, however, family support remained associated with career interest after adjustment for confidence, with an odds ratio of similar size.

One interpretation is that contextual supports can also act proximally, at the point of choice. Later formulations of SCCT anticipate this, particularly where educational decisions are made jointly with families [40]. This interpretation is plausible in Bangladesh, but the present design cannot test it. The same pattern is compatible with reverse influence, in which interested students elicit more encouragement. It is also compatible with shared‐method variance, or with unmeasured factors such as household income. The modest explanatory power of the models (pseudo‐R² = 0.14 to 0.23) shows that much of the variation in career interest lies outside the measured variables.

5.3. Role‐Model Exposure

Role‐model exposure correlated positively with career interest (ρ = 0.25) but was not independently associated with it. Two explanations are possible. First, its influence may be indirect. Role‐model exposure correlated with both confidence and family support (ρ = 0.31 for each). Part of any effect may therefore operate through these factors and be absorbed when they enter the model. Second, Gender Schema Theory suggests that brief or distant exposure to counter‐stereotypic women may be treated as an exception [41]. It may then fail to challenge gender‐typed expectations. Sustained and relatable contact is more likely to change interest [42, 43].

The single item did not capture the type or frequency of exposure. The data therefore cannot separate mediation, a genuine absence of effect, and measurement limits. The result is consistent with the argument that women's visibility in STEM matters but is not sufficient without broader institutional change [44].

5.4. Access to Digital Resources

Access correlated weakly and positively with career interest on its own (ρ = 0.17). It showed no positive association after adjustment and was weakly negative in the binary model. The simplest reading is that access shares much of its association with family support and confidence. Households that provide devices and connectivity may also be those that encourage STEAM study.

Two further points limit interpretation. Because the survey was online, all respondents were connected, and 71.70% agreed or strongly agreed that they had adequate access. This restricts the range of the variable. The exploratory profiles point the same way. The profile with the lowest career interest reported access almost as high as the high‐readiness profile. Access therefore appears necessary but not sufficient in this sample. The findings do not show that access is unimportant for students who lack it.

5.5. Urban and Rural Differences

Rural respondents reported less strong family encouragement (d = 0.43) and somewhat lower career interest (d = 0.26) than urban respondents. Confidence and access did not differ significantly. Location was not associated with career interest once family support and the other predictors were controlled. This pattern is consistent with family encouragement accounting for part of the rural gap, although the design cannot establish this.

Social Role Theory offers one interpretation. Material conditions such as connectivity can change quickly through policy. Normative expectations about women's roles tend to change more slowly [45]. The pattern is also consistent with Bangladeshi evidence that family and community norms shape girls' educational paths, especially in rural areas [11, 12].

Three cautions apply. First, non‐significant differences do not establish equivalence. The rural subsample (n = 112) could detect only differences of about d = 0.31 or larger. Second, the online purposive sample probably under‐represents rural students with poor connectivity. Third, the family support item measured encouragement, not social norms directly.

5.6. Contribution

The main contribution of the study is the joint estimation of confidence, family support, access and role‐model exposure in a Bangladeshi sample. Confidence and family support carried independent associations with career interest. Access and role‐model exposure did not. Interventions that target a single factor may therefore be insufficient. The exploratory profiles add a descriptive nuance. About three in ten respondents combined good access with lower family support, role‐model exposure and interest.

5.7. Implications

Efforts to increase girls' participation in STEAM may be more effective if confidence building is combined with the engagement of families and teachers. Schools could build confidence through inquiry‐based laboratory work, coding practice and formative feedback. Programs could involve parents directly, for example through information sessions on STEAM careers. Teachers could be trained in gender‐responsive practice. Role‐model initiatives may be more useful when embedded in sustained mentoring than when delivered as one‐off events. Digital access programs may need family and mentoring components, since access alone was not associated with higher interest in this sample. Rural respondents reported less strong family encouragement. Rural programs may therefore benefit from particular attention to family and community engagement. These suggestions come from associations in a non‐probability sample. They should be evaluated in intervention studies before being adopted at scale.

6. Conclusions

This study examined factors associated with STEAM career interest among 424 female students in Bangladesh. Confidence and family support were independently associated with career interest. Role‐model exposure and digital access showed no positive association after mutual adjustment. Urban and rural respondents reported similar confidence and access but differed in family encouragement.

As an interpretation rather than a demonstrated mechanism, we propose a distinction between capability and permission. Confidence and resources may be understood as capability, meaning what a student feels able to do. Family support may be understood as permission, meaning whether pursuing STEAM is regarded as appropriate for her. Where support is weak, gains in capability may not translate into aspiration. This interpretation yields a testable hypothesis. The relative weight of family support should be greater where educational choices are more strongly shaped by family and social expectations.

In practice, the findings suggest four priorities. Confidence‐building pedagogy should be combined with family and teacher engagement. Role models should be embedded in sustained mentoring. Access programs should be paired with these components. Rural settings deserve particular attention to family engagement.

Several limitations apply. The cross‐sectional design cannot establish causal direction, and reverse influence is possible. Self‐reported measures are subject to social desirability and common‐method bias. Each construct was measured with a single item, so construct reliability could not be estimated. The reported Cronbach's alpha refers to the ten Likert items together. Purposive online sampling produced an urban (73.58%) and undergraduate (65.09%) majority, and only 29 secondary students took part. All respondents were connected to the internet. The findings therefore mainly reflect urban, university‐level and connected respondents, and the urban and rural comparison is analytical rather than representative. Some respondents' ages suggest that they reported their highest completed level of education. The binary model showed some miscalibration, and the MANOVA and cluster analyses were exploratory.

Future research should use longitudinal designs and validated multi‐item scales. Probability samples stratified by location and educational stage are needed. Formal mediation analysis could clarify the role of role models. Comparisons across settings that differ in normative constraint could test the capability and permission distinction. Mixed‐method designs could draw on students' own accounts.

Increasing girls' participation in STEAM may require attention not only to their skills and access. It may also require attention to the family and social environments in which their educational choices are made.

Author Contributions

Farhana Yasmin: conceptualization, methodology, investigation, writing – original draft, formal analysis, data curation, resources, visualization. Mosammat Shamima Nasrain: conceptualization, investigation, writing – original draft, methodology, validation, project administration, supervision, resources. Pretty Bowm: conceptualization, data curation, resources, software, writing – original draft. Jemima Jahan Meem: visualization, validation, methodology, conceptualization, investigation. Israt Zahan: conceptualization, resources, supervision, project administration, writing – review and editing, writing – original draft, validation. Md. Alamgir Hossain: writing – original draft, conceptualization, validation, writing – review andediting, supervision, data curation, resources, software, formal analysis, project administration.

Funding

The authors have nothing to report.

Ethics Statement

Participation was voluntary, and respondents could withdraw from the survey at any time. The questionnaire did not require identifying information. All data were stored securely and used only for academic research. Ethical approval was obtained from the Skill Morph Ethics Committee (Reference: SkillMorph/ES/2025/02(05)) before data collection.

Conflicts of Interest

The authors declare no conflicts of interest.

Transparency Statement

The authors affirm that this manuscript is an honest, accurate and transparent account of the study being reported. No important aspects have been omitted, and any discrepancies from the study as planned have been explained.

Supporting information

Figure S1: Within‐cluster sum of squares (a) and average silhouette coefficient (b) for k = 2 to 8.

Figure S2: Confidence in science and mathematics (a) and access to internet and online learning tools (b), by location.

Data Availability Statement

The dataset and analysis code supporting this study are available from the corresponding author on reasonable request.