Smartphones undermine social connectedness more in men than women

Earlier experiments on phones and well-being each studied one situation. They were also usually too small to ask who is affected most. We had run eight experiments in our own lab, four published (n = 809) and four unpublished (n = 969). I combined the raw data from all of them. Including the unpublished studies guards against publication bias, which favors small studies that found large effects.

The theory behind the work is the displacement, interference and complementarity framework from our review paper. It predicts that phones lower well-being when they interrupt pleasant social time, and can help when they give us information we need right now. We registered two predictions. People with a phone would report lower well-being (H1) and lower social connection (H2) than people without one, in every situation except waiting alone, where there was nobody to interrupt.

Measures and design

Social connectedness combined two items on a 7-point scale, adapted from the Social Connectedness Scale: “I feel close to people” minus “I feel distant from people.” The two items correlated between −.41 and −.59 across experiments.

Affect valence was how often people felt good, positive and pleasant minus bad, negative and unpleasant (reliability α = .85 to .90). The parenting experiments used one overall item instead. Both measures were standardized within each situation before pooling.

Design. People were randomly assigned to phone or no phone (50.1% had a phone). Whole tables and groups shared a condition. The purpose of each study was disguised. We counted everyone as assigned, which is the conservative choice. A sensitivity analysis showed the full sample could detect an effect as small as f² = .007 with 95% power.

Analysis plan. We first asked with ordinary regressions whether the effect differed by situation, then pooled the situations that did not differ. Because six situations are few clusters, we avoided a complex model at that step. We then ran multilevel models with people nested in situations, adding phone condition, then gender and age, then their interaction. Our first registration used ordinary regression throughout. We updated it, before looking at the pooled results, to multilevel models and dropped covariates that were not measured in every study (ethnicity was measured in 1,065 people, phone dependence in 713).

No. In the first regression no situation changed the phone effect on connectedness, so I pooled all six. The one interaction that stands out is phone by gender.

People with a phone were less connected than people without one (phone −.13, no phone .13 on the standardized scale; b = −.259, t(1749) = −4.33, p < .001, d = −.207). Adding gender and age barely changed it (b = −.241, d = −.181). Men reported less connection than women overall (b = −.134, p = .010, d = −.126), and age did not matter (p = .368).

Model 3 added the interaction (b = −.313, SE = .103, t(1684) = −3.03, p = .003, d = −.148). The simple effects show where it comes from: the phone effect was larger for men (b = −.461, p < .001, d = −.473) than for women (b = −.148, p = .041, d = −.152). Men and women who had no phone did not differ (b = −.018, p = .799, d = −.019).

For mood, one situation stood out: getting directions (Condition × Getting directions, b = .251, p = .024, d = .204). Once I removed that situation, no other interaction was reliable. As registered, I analyzed getting directions separately from the other five.

People with a phone reported more negative affect (b = −.105, t(1436) = −2.009, p = .045, d = −.106). With gender and age in the model it held (b = −.118, p = .027, d = −.119). Men reported more negative affect than women (b = −.185, p = .002, d = −.171). Age did not matter (p = .461). The interaction with gender was not reliable (b = −.159, p = .169, d = −.074).

Phones cost connection even in unpublished studies. Pooling the nonsignificant studies did not erase the effect. It is small (d from −.15 to −.21 in the models), but small effects can add up. A person who is a little less connected at every meal and wait could notice the difference over a day.

The waiting-alone result was not what we predicted. We expected no phone cost when nobody else was there. The effect was not reliably different, and a phone did not make waiting alone feel better either. Our one nonsocial situation cannot settle this.

Mood depends on what the phone is for. Phones lowered affect valence when they competed with people. They did not when the phone served a current goal. The published getting-directions study found two paths that point in opposite directions: finding the building faster felt good, and feeling less connected felt bad. A null main effect is consistent with those two cancelling.

Why men? We did not test the reasons. The paper lists four candidates. Men may use phones more in social settings, though earlier research suggests women use phones more. Men may be more distractible, though earlier research finds no gender gap in attention. Women may follow stricter norms about phones in company. Or women may use phones in more social ways, such as messaging. Each could be tested by measuring how people use their phones during a conversation.

Limits. The samples were mostly college-aged people from British Columbia, and most situations were pleasant social ones. Taking a phone away may add confounds. We analyzed data we had already seen, so our registered plans could have been shaped by that. The gender result was not registered and needs a new, confirmatory test. We kept everyone as randomly assigned, which makes the effect for people who actually followed instructions a little larger than the numbers here.

Every coefficient and test comes from the printed paper. The paper’s data, code and materials are public on OSF, including the supplemental tables and the preregistrations. I have not re-run the models on the raw data.

For the by-situation forest plots, I derived each 95% interval from the group size, with the standard error equal to the square root of 4 over N plus d squared over 2N. The same approach reproduces the paper’s own forest plot. The gender averages carry intervals of 1.96 standard errors from the printed standard errors. The paper’s Table 6 is partly garbled in my copy, so I used the values the text prints. The rain plots in the paper’s Figure 2 need the individual data, so I show the averages instead.